Make rustfmt happy

This reverts commit 7a4b6c065482f988b01298642f4c18775503f92f.

.dockerignore

@@ -2,4 +2,3 @@ target
 Dockerfile
 .dockerignore
 .gitignore
-**/.git

.github/workflows/publish-apt-brew-pkg.yml

@@ -35,7 +35,7 @@ jobs:
     - name: Build deb package
       run: cargo deb -p meilisearch -o target/debian/meilisearch.deb
     - name: Upload debian pkg to release
-      uses: svenstaro/upload-release-action@2.5.0
+      uses: svenstaro/upload-release-action@2.6.1
       with:
         repo_token: ${{ secrets.MEILI_BOT_GH_PAT }}
         file: target/debian/meilisearch.deb

.github/workflows/publish-binaries.yml

@@ -54,7 +54,7 @@ jobs:
     # No need to upload binaries for dry run (cron)
     - name: Upload binaries to release
       if: github.event_name == 'release'
-      uses: svenstaro/upload-release-action@2.5.0
+      uses: svenstaro/upload-release-action@2.6.1
       with:
         repo_token: ${{ secrets.MEILI_BOT_GH_PAT }}
         file: target/release/meilisearch
@@ -87,7 +87,7 @@ jobs:
     # No need to upload binaries for dry run (cron)
     - name: Upload binaries to release
       if: github.event_name == 'release'
-      uses: svenstaro/upload-release-action@2.5.0
+      uses: svenstaro/upload-release-action@2.6.1
       with:
         repo_token: ${{ secrets.MEILI_BOT_GH_PAT }}
         file: target/release/${{ matrix.artifact_name }}
@@ -121,7 +121,7 @@ jobs:
       - name: Upload the binary to release
         # No need to upload binaries for dry run (cron)
         if: github.event_name == 'release'
-        uses: svenstaro/upload-release-action@2.5.0
+        uses: svenstaro/upload-release-action@2.6.1
         with:
           repo_token: ${{ secrets.MEILI_BOT_GH_PAT }}
           file: target/${{ matrix.target }}/release/meilisearch
@@ -183,7 +183,7 @@ jobs:
       - name: Upload the binary to release
         # No need to upload binaries for dry run (cron)
         if: github.event_name == 'release'
-        uses: svenstaro/upload-release-action@2.5.0
+        uses: svenstaro/upload-release-action@2.6.1
         with:
           repo_token: ${{ secrets.MEILI_BOT_GH_PAT }}
           file: target/${{ matrix.target }}/release/meilisearch

.github/workflows/publish-docker-images.yml

@@ -58,13 +58,9 @@ jobs:
 
       - name: Set up QEMU
         uses: docker/setup-qemu-action@v2
-        with:
-          platforms: linux/amd64,linux/arm64
 
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v2
-        with:
-          platforms: linux/amd64,linux/arm64
 
       - name: Login to Docker Hub
         uses: docker/login-action@v2
@@ -92,13 +88,10 @@ jobs:
           push: true
           platforms: linux/amd64,linux/arm64
           tags: ${{ steps.meta.outputs.tags }}
-          builder: ${{ steps.buildx.outputs.name }}
           build-args: |
             COMMIT_SHA=${{ github.sha }}
             COMMIT_DATE=${{ steps.build-metadata.outputs.date }}
             GIT_TAG=${{ github.ref_name }}
-          cache-from: type=gha
-          cache-to: type=gha,mode=max
 
       # /!\ Don't touch this without checking with Cloud team
       - name: Send CI information to Cloud team

.github/workflows/sdks-tests.yml

@@ -3,6 +3,11 @@ name: SDKs tests
 
 on:
   workflow_dispatch:
+    inputs:
+      docker_image:
+        description: 'The Meilisearch Docker image used'
+        required: false
+        default: nightly
   schedule:
     - cron: "0 6 * * MON" # Every Monday at 6:00AM
 
@@ -11,13 +16,28 @@ env:
   MEILI_NO_ANALYTICS: 'true'
 
 jobs:
+  define-docker-image:
+    runs-on: ubuntu-latest
+    outputs:
+      docker-image: ${{ steps.define-image.outputs.docker-image }}
+    steps:
+      - uses: actions/checkout@v3
+      - name: Define the Docker image we need to use
+        id: define-image
+        run: |
+          event=${{ github.event.action }}
+          echo "docker-image=nightly" >> $GITHUB_OUTPUT
+          if [[ $event == 'workflow_dispatch' ]]; then
+            echo "docker-image=${{ github.event.inputs.docker_image }}" >> $GITHUB_OUTPUT
+          fi
 
   meilisearch-js-tests:
+    needs: define-docker-image
     name: JS SDK tests
     runs-on: ubuntu-latest
     services:
       meilisearch:
-        image: getmeili/meilisearch:nightly
+        image: getmeili/meilisearch:${{ github.event.inputs.docker_image }}
         env:
           MEILI_MASTER_KEY: ${{ env.MEILI_MASTER_KEY }}
           MEILI_NO_ANALYTICS: ${{ env.MEILI_NO_ANALYTICS }}
@@ -47,11 +67,12 @@ jobs:
         run: yarn test:env:browser
 
   instant-meilisearch-tests:
+    needs: define-docker-image
     name: instant-meilisearch tests
     runs-on: ubuntu-latest
     services:
       meilisearch:
-        image: getmeili/meilisearch:nightly
+        image: getmeili/meilisearch:${{ github.event.inputs.docker_image }}
         env:
           MEILI_MASTER_KEY: ${{ env.MEILI_MASTER_KEY }}
           MEILI_NO_ANALYTICS: ${{ env.MEILI_NO_ANALYTICS }}
@@ -73,11 +94,12 @@ jobs:
         run: yarn build
 
   meilisearch-php-tests:
+    needs: define-docker-image
     name: PHP SDK tests
     runs-on: ubuntu-latest
     services:
       meilisearch:
-        image: getmeili/meilisearch:nightly
+        image: getmeili/meilisearch:${{ github.event.inputs.docker_image }}
         env:
           MEILI_MASTER_KEY: ${{ env.MEILI_MASTER_KEY }}
           MEILI_NO_ANALYTICS: ${{ env.MEILI_NO_ANALYTICS }}
@@ -103,11 +125,12 @@ jobs:
           composer remove --dev guzzlehttp/guzzle http-interop/http-factory-guzzle
 
   meilisearch-python-tests:
+    needs: define-docker-image
     name: Python SDK tests
     runs-on: ubuntu-latest
     services:
       meilisearch:
-        image: getmeili/meilisearch:nightly
+        image: getmeili/meilisearch:${{ github.event.inputs.docker_image }}
         env:
           MEILI_MASTER_KEY: ${{ env.MEILI_MASTER_KEY }}
           MEILI_NO_ANALYTICS: ${{ env.MEILI_NO_ANALYTICS }}
@@ -127,11 +150,12 @@ jobs:
         run: pipenv run pytest
 
   meilisearch-go-tests:
+    needs: define-docker-image
     name: Go SDK tests
     runs-on: ubuntu-latest
     services:
       meilisearch:
-        image: getmeili/meilisearch:nightly
+        image: getmeili/meilisearch:${{ github.event.inputs.docker_image }}
         env:
           MEILI_MASTER_KEY: ${{ env.MEILI_MASTER_KEY }}
           MEILI_NO_ANALYTICS: ${{ env.MEILI_NO_ANALYTICS }}
@@ -139,7 +163,7 @@ jobs:
           - '7700:7700'
     steps:
       - name: Set up Go
-        uses: actions/setup-go@v3
+        uses: actions/setup-go@v4
         with:
           go-version: stable
       - uses: actions/checkout@v3
@@ -156,11 +180,12 @@ jobs:
         run: go test -v ./...
 
   meilisearch-ruby-tests:
+    needs: define-docker-image
     name: Ruby SDK tests
     runs-on: ubuntu-latest
     services:
       meilisearch:
-        image: getmeili/meilisearch:nightly
+        image: getmeili/meilisearch:${{ github.event.inputs.docker_image }}
         env:
           MEILI_MASTER_KEY: ${{ env.MEILI_MASTER_KEY }}
           MEILI_NO_ANALYTICS: ${{ env.MEILI_NO_ANALYTICS }}
@@ -180,11 +205,12 @@ jobs:
         run: bundle exec rspec
 
   meilisearch-rust-tests:
+    needs: define-docker-image
     name: Rust SDK tests
     runs-on: ubuntu-latest
     services:
       meilisearch:
-        image: getmeili/meilisearch:nightly
+        image: getmeili/meilisearch:${{ github.event.inputs.docker_image }}
         env:
           MEILI_MASTER_KEY: ${{ env.MEILI_MASTER_KEY }}
           MEILI_NO_ANALYTICS: ${{ env.MEILI_NO_ANALYTICS }}

.github/workflows/test-suite.yml

@@ -43,7 +43,7 @@ jobs:
           toolchain: nightly
           override: true
       - name: Cache dependencies
-        uses: Swatinem/rust-cache@v2.2.1
+        uses: Swatinem/rust-cache@v2.4.0
       - name: Run cargo check without any default features
         uses: actions-rs/cargo@v1
         with:
@@ -65,7 +65,7 @@ jobs:
     steps:
       - uses: actions/checkout@v3
       - name: Cache dependencies
-        uses: Swatinem/rust-cache@v2.2.1
+        uses: Swatinem/rust-cache@v2.4.0
       - name: Run cargo check without any default features
         uses: actions-rs/cargo@v1
         with:
@@ -105,6 +105,29 @@ jobs:
           command: test
           args: --workspace --locked --release --all-features
 
+  test-disabled-tokenization:
+    name: Test disabled tokenization
+    runs-on: ubuntu-latest
+    container:
+      image: ubuntu:18.04
+    if: github.event_name == 'schedule'
+    steps:
+      - uses: actions/checkout@v3
+      - name: Install needed dependencies
+        run: |
+          apt-get update
+          apt-get install --assume-yes build-essential curl
+      - uses: actions-rs/toolchain@v1 
+        with: 
+          toolchain: stable 
+          override: true 
+      - name: Run cargo tree without default features and check lindera is not present
+        run: |
+          cargo tree -f '{p} {f}' -e normal --no-default-features | grep lindera -vqz
+      - name: Run cargo tree with default features and check lindera is pressent
+        run: |
+          cargo tree -f '{p} {f}' -e normal | grep lindera -qz
+                
   # We run tests in debug also, to make sure that the debug_assertions are hit
   test-debug:
     name: Run tests in debug
@@ -123,7 +146,7 @@ jobs:
           toolchain: stable
           override: true
       - name: Cache dependencies
-        uses: Swatinem/rust-cache@v2.2.1
+        uses: Swatinem/rust-cache@v2.4.0
       - name: Run tests in debug
         uses: actions-rs/cargo@v1
         with:
@@ -142,7 +165,7 @@ jobs:
           override: true
           components: clippy
       - name: Cache dependencies
-        uses: Swatinem/rust-cache@v2.2.1
+        uses: Swatinem/rust-cache@v2.4.0
       - name: Run cargo clippy
         uses: actions-rs/cargo@v1
         with:
@@ -161,7 +184,7 @@ jobs:
           override: true
           components: rustfmt
       - name: Cache dependencies
-        uses: Swatinem/rust-cache@v2.2.1
+        uses: Swatinem/rust-cache@v2.4.0
       - name: Run cargo fmt
         # Since we never ran the `build.rs` script in the benchmark directory we are missing one auto-generated import file.
         # Since we want to trigger (and fail) this action as fast as possible, instead of building the benchmark crate

Cargo.lock

@@ -1207,6 +1207,12 @@ dependencies = [
  "winapi",
 ]
 
+[[package]]
+name = "doc-comment"
+version = "0.3.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "fea41bba32d969b513997752735605054bc0dfa92b4c56bf1189f2e174be7a10"
+
 [[package]]
 name = "dump"
 version = "1.2.0"
@@ -1763,6 +1769,15 @@ dependencies = [
  "byteorder",
 ]
 
+[[package]]
+name = "hashbrown"
+version = "0.11.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ab5ef0d4909ef3724cc8cce6ccc8572c5c817592e9285f5464f8e86f8bd3726e"
+dependencies = [
+ "ahash 0.7.6",
+]
+
 [[package]]
 name = "hashbrown"
 version = "0.12.3"
@@ -1794,7 +1809,7 @@ checksum = "95505c38b4572b2d910cecb0281560f54b440a19336cbbcb27bf6ce6adc6f5a8"
 [[package]]
 name = "heed"
 version = "0.12.5"
-source = "git+https://github.com/meilisearch/heed?tag=v0.12.5#4158a6c484752afaaf9e2530a6ee0e7ab0f24ee8"
+source = "git+https://github.com/meilisearch/heed?tag=v0.12.6#8c5b94225fc949c02bb7b900cc50ffaf6b584b1e"
 dependencies = [
  "byteorder",
  "heed-traits",
@@ -1811,12 +1826,12 @@ dependencies = [
 [[package]]
 name = "heed-traits"
 version = "0.7.0"
-source = "git+https://github.com/meilisearch/heed?tag=v0.12.5#4158a6c484752afaaf9e2530a6ee0e7ab0f24ee8"
+source = "git+https://github.com/meilisearch/heed?tag=v0.12.6#8c5b94225fc949c02bb7b900cc50ffaf6b584b1e"
 
 [[package]]
 name = "heed-types"
 version = "0.7.2"
-source = "git+https://github.com/meilisearch/heed?tag=v0.12.5#4158a6c484752afaaf9e2530a6ee0e7ab0f24ee8"
+source = "git+https://github.com/meilisearch/heed?tag=v0.12.6#8c5b94225fc949c02bb7b900cc50ffaf6b584b1e"
 dependencies = [
  "bincode",
  "heed-traits",
@@ -1864,6 +1879,22 @@ dependencies = [
  "digest",
 ]
 
+[[package]]
+name = "hnsw"
+version = "0.11.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2b9740ebf8769ec4ad6762cc951ba18f39bba6dfbc2fbbe46285f7539af79752"
+dependencies = [
+ "ahash 0.7.6",
+ "hashbrown 0.11.2",
+ "libm",
+ "num-traits",
+ "rand_core",
+ "serde",
+ "smallvec",
+ "space",
+]
+
 [[package]]
 name = "http"
 version = "0.2.9"
@@ -1994,7 +2025,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "bd070e393353796e801d209ad339e89596eb4c8d430d18ede6a1cced8fafbd99"
 dependencies = [
  "autocfg",
- "hashbrown",
+ "hashbrown 0.12.3",
  "serde",
 ]
 
@@ -2086,7 +2117,7 @@ checksum = "37228e06c75842d1097432d94d02f37fe3ebfca9791c2e8fef6e9db17ed128c1"
 dependencies = [
  "cedarwood",
  "fxhash",
- "hashbrown",
+ "hashbrown 0.12.3",
  "lazy_static",
  "phf",
  "phf_codegen",
@@ -2596,6 +2627,7 @@ dependencies = [
  "num_cpus",
  "obkv",
  "once_cell",
+ "ordered-float",
  "parking_lot",
  "permissive-json-pointer",
  "pin-project-lite",
@@ -2715,6 +2747,7 @@ dependencies = [
  "bimap",
  "bincode",
  "bstr",
+ "bytemuck",
  "byteorder",
  "charabia",
  "concat-arrays",
@@ -2730,6 +2763,7 @@ dependencies = [
  "geoutils",
  "grenad",
  "heed",
+ "hnsw",
  "insta",
  "itertools",
  "json-depth-checker",
@@ -2744,6 +2778,7 @@ dependencies = [
  "once_cell",
  "ordered-float",
  "rand",
+ "rand_pcg",
  "rayon",
  "roaring",
  "rstar",
@@ -2753,6 +2788,7 @@ dependencies = [
  "smallstr",
  "smallvec",
  "smartstring",
+ "space",
  "tempfile",
  "thiserror",
  "time",
@@ -2941,9 +2977,9 @@ checksum = "624a8340c38c1b80fd549087862da4ba43e08858af025b236e509b6649fc13d5"
 
 [[package]]
 name = "ordered-float"
-version = "3.6.0"
+version = "3.7.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "13a384337e997e6860ffbaa83708b2ef329fd8c54cb67a5f64d421e0f943254f"
+checksum = "2fc2dbde8f8a79f2102cc474ceb0ad68e3b80b85289ea62389b60e66777e4213"
 dependencies = [
  "num-traits",
 ]
@@ -3327,6 +3363,16 @@ dependencies = [
  "getrandom",
 ]
 
+[[package]]
+name = "rand_pcg"
+version = "0.3.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "59cad018caf63deb318e5a4586d99a24424a364f40f1e5778c29aca23f4fc73e"
+dependencies = [
+ "rand_core",
+ "serde",
+]
+
 [[package]]
 name = "rayon"
 version = "1.7.0"
@@ -3764,6 +3810,9 @@ name = "smallvec"
 version = "1.10.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "a507befe795404456341dfab10cef66ead4c041f62b8b11bbb92bffe5d0953e0"
+dependencies = [
+ "serde",
+]
 
 [[package]]
 name = "smartstring"
@@ -3786,6 +3835,16 @@ dependencies = [
  "winapi",
 ]
 
+[[package]]
+name = "space"
+version = "0.17.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c5ab9701ae895386d13db622abf411989deff7109b13b46b6173bb4ce5c1d123"
+dependencies = [
+ "doc-comment",
+ "num-traits",
+]
+
 [[package]]
 name = "spin"
 version = "0.5.2"
@@ -4433,7 +4492,7 @@ version = "0.16.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "9c531a2dc4c462b833788be2c07eef4e621d0e9edbd55bf280cc164c1c1aa043"
 dependencies = [
- "hashbrown",
+ "hashbrown 0.12.3",
  "once_cell",
 ]
 

Dockerfile

@@ -1,4 +1,3 @@
-# syntax=docker/dockerfile:1.4
 # Compile
 FROM    rust:alpine3.16 AS compiler
 
@@ -12,7 +11,7 @@ ARG     GIT_TAG
 ENV     VERGEN_GIT_SHA=${COMMIT_SHA} VERGEN_GIT_COMMIT_TIMESTAMP=${COMMIT_DATE} VERGEN_GIT_SEMVER_LIGHTWEIGHT=${GIT_TAG}
 ENV     RUSTFLAGS="-C target-feature=-crt-static"
 
-COPY    --link . .
+COPY    . .
 RUN     set -eux; \
         apkArch="$(apk --print-arch)"; \
         if [ "$apkArch" = "aarch64" ]; then \
@@ -31,7 +30,7 @@ RUN     apk update --quiet \
 
 # add meilisearch to the `/bin` so you can run it from anywhere and it's easy
 # to find.
-COPY    --from=compiler --link /meilisearch/target/release/meilisearch /bin/meilisearch
+COPY    --from=compiler /meilisearch/target/release/meilisearch /bin/meilisearch
 # To stay compatible with the older version of the container (pre v0.27.0) we're
 # going to symlink the meilisearch binary in the path to `/meilisearch`
 RUN     ln -s /bin/meilisearch /meilisearch

assets/prometheus-basic-scraper.yml

@@ -0,0 +1,19 @@
+global:
+  scrape_interval:     15s # By default, scrape targets every 15 seconds.
+
+  # Attach these labels to any time series or alerts when communicating with
+  # external systems (federation, remote storage, Alertmanager).
+  external_labels:
+    monitor: 'codelab-monitor'
+
+# A scrape configuration containing exactly one endpoint to scrape:
+# Here it's Prometheus itself.
+scrape_configs:
+  # The job name is added as a label `job=<job_name>` to any timeseries scraped from this config.
+  - job_name: 'meilisearch'
+
+    # Override the global default and scrape targets from this job every 5 seconds.
+    scrape_interval: 5s
+
+    static_configs:
+      - targets: ['localhost:7700']

config.toml

@@ -1,128 +1,131 @@
 # This file shows the default configuration of Meilisearch.
 # All variables are defined here: https://www.meilisearch.com/docs/learn/configuration/instance_options#environment-variables
 
-db_path = "./data.ms"
 # Designates the location where database files will be created and retrieved.
 # https://www.meilisearch.com/docs/learn/configuration/instance_options#database-path
+db_path = "./data.ms"
 
-env = "development"
 # Configures the instance's environment. Value must be either `production` or `development`.
 # https://www.meilisearch.com/docs/learn/configuration/instance_options#environment
+env = "development"
 
-http_addr = "localhost:7700"
 # The address on which the HTTP server will listen.
+http_addr = "localhost:7700"
 
-# master_key = "YOUR_MASTER_KEY_VALUE"
 # Sets the instance's master key, automatically protecting all routes except GET /health.
 # https://www.meilisearch.com/docs/learn/configuration/instance_options#master-key
+# master_key = "YOUR_MASTER_KEY_VALUE"
 
-# no_analytics = true
 # Deactivates Meilisearch's built-in telemetry when provided.
 # Meilisearch automatically collects data from all instances that do not opt out using this flag.
 # All gathered data is used solely for the purpose of improving Meilisearch, and can be deleted at any time.
 # https://www.meilisearch.com/docs/learn/configuration/instance_options#disable-analytics
+# no_analytics = true
 
-http_payload_size_limit = "100 MB"
 # Sets the maximum size of accepted payloads.
 # https://www.meilisearch.com/docs/learn/configuration/instance_options#payload-limit-size
+http_payload_size_limit = "100 MB"
 
-log_level = "INFO"
 # Defines how much detail should be present in Meilisearch's logs.
 # Meilisearch currently supports six log levels, listed in order of increasing verbosity:  `OFF`, `ERROR`, `WARN`, `INFO`, `DEBUG`, `TRACE`
 # https://www.meilisearch.com/docs/learn/configuration/instance_options#log-level
+log_level = "INFO"
 
-# max_indexing_memory = "2 GiB"
 # Sets the maximum amount of RAM Meilisearch can use when indexing.
 # https://www.meilisearch.com/docs/learn/configuration/instance_options#max-indexing-memory
+# max_indexing_memory = "2 GiB"
 
-# max_indexing_threads = 4
 # Sets the maximum number of threads Meilisearch can use during indexing.
 # https://www.meilisearch.com/docs/learn/configuration/instance_options#max-indexing-threads
+# max_indexing_threads = 4
 
 #############
 ### DUMPS ###
 #############
 
-dump_dir = "dumps/"
 # Sets the directory where Meilisearch will create dump files.
 # https://www.meilisearch.com/docs/learn/configuration/instance_options#dump-directory
+dump_dir = "dumps/"
 
-# import_dump = "./path/to/my/file.dump"
 # Imports the dump file located at the specified path. Path must point to a .dump file.
 # https://www.meilisearch.com/docs/learn/configuration/instance_options#import-dump
+# import_dump = "./path/to/my/file.dump"
 
-ignore_missing_dump = false
 # Prevents Meilisearch from throwing an error when `import_dump` does not point to a valid dump file.
 # https://www.meilisearch.com/docs/learn/configuration/instance_options#ignore-missing-dump
+ignore_missing_dump = false
 
-ignore_dump_if_db_exists = false
 # Prevents a Meilisearch instance with an existing database from throwing an error when using `import_dump`.
 # https://www.meilisearch.com/docs/learn/configuration/instance_options#ignore-dump-if-db-exists
+ignore_dump_if_db_exists = false
 
 
 #################
 ### SNAPSHOTS ###
 #################
 
-schedule_snapshot = false
 # Enables scheduled snapshots when true, disable when false (the default).
 # If the value is given as an integer, then enables the scheduled snapshot with the passed value as the interval
 # between each snapshot, in seconds.
 # https://www.meilisearch.com/docs/learn/configuration/instance_options#schedule-snapshot-creation
+schedule_snapshot = false
 
-snapshot_dir = "snapshots/"
 # Sets the directory where Meilisearch will store snapshots.
 # https://www.meilisearch.com/docs/learn/configuration/instance_options#snapshot-destination
+snapshot_dir = "snapshots/"
 
-# import_snapshot = "./path/to/my/snapshot"
 # Launches Meilisearch after importing a previously-generated snapshot at the given filepath.
 # https://www.meilisearch.com/docs/learn/configuration/instance_options#import-snapshot
+# import_snapshot = "./path/to/my/snapshot"
 
-ignore_missing_snapshot = false
 # Prevents a Meilisearch instance from throwing an error when `import_snapshot` does not point to a valid snapshot file.
 # https://www.meilisearch.com/docs/learn/configuration/instance_options#ignore-missing-snapshot
+ignore_missing_snapshot = false
 
-ignore_snapshot_if_db_exists = false
 # Prevents a Meilisearch instance with an existing database from throwing an error when using `import_snapshot`.
 # https://www.meilisearch.com/docs/learn/configuration/instance_options#ignore-snapshot-if-db-exists
+ignore_snapshot_if_db_exists = false
 
 
 ###########
 ### SSL ###
 ###########
 
-# ssl_auth_path = "./path/to/root"
 # Enables client authentication in the specified path.
 # https://www.meilisearch.com/docs/learn/configuration/instance_options#ssl-authentication-path
+# ssl_auth_path = "./path/to/root"
 
-# ssl_cert_path = "./path/to/certfile"
 # Sets the server's SSL certificates.
 # https://www.meilisearch.com/docs/learn/configuration/instance_options#ssl-certificates-path
+# ssl_cert_path = "./path/to/certfile"
 
-# ssl_key_path = "./path/to/private-key"
 # Sets the server's SSL key files.
 # https://www.meilisearch.com/docs/learn/configuration/instance_options#ssl-key-path
+# ssl_key_path = "./path/to/private-key"
 
-# ssl_ocsp_path = "./path/to/ocsp-file"
 # Sets the server's OCSP file.
 # https://www.meilisearch.com/docs/learn/configuration/instance_options#ssl-ocsp-path
+# ssl_ocsp_path = "./path/to/ocsp-file"
 
-ssl_require_auth = false
 # Makes SSL authentication mandatory.
 # https://www.meilisearch.com/docs/learn/configuration/instance_options#ssl-require-auth
+ssl_require_auth = false
 
-ssl_resumption = false
 # Activates SSL session resumption.
 # https://www.meilisearch.com/docs/learn/configuration/instance_options#ssl-resumption
+ssl_resumption = false
 
-ssl_tickets = false
 # Activates SSL tickets.
 # https://www.meilisearch.com/docs/learn/configuration/instance_options#ssl-tickets
+ssl_tickets = false
 
 #############################
 ### Experimental features ###
 #############################
 
-experimental_enable_metrics = false
 # Experimental metrics feature. For more information, see: <https://github.com/meilisearch/meilisearch/discussions/3518>
 # Enables the Prometheus metrics on the `GET /metrics` endpoint.
+experimental_enable_metrics = false
+
+# Experimental RAM reduction during indexing, do not use in production, see: <https://github.com/meilisearch/product/discussions/652>
+experimental_reduce_indexing_memory_usage = false

index-scheduler/src/batch.rs

@@ -24,6 +24,7 @@ use std::io::BufWriter;
 
 use dump::IndexMetadata;
 use log::{debug, error, info};
+use meilisearch_types::error::Code;
 use meilisearch_types::heed::{RoTxn, RwTxn};
 use meilisearch_types::milli::documents::{obkv_to_object, DocumentsBatchReader};
 use meilisearch_types::milli::heed::CompactionOption;
@@ -1491,7 +1492,12 @@ fn delete_document_by_filter(filter: &serde_json::Value, index: Index) -> Result
     Ok(if let Some(filter) = filter {
         let mut wtxn = index.write_txn()?;
 
-        let candidates = filter.evaluate(&wtxn, &index)?;
+        let candidates = filter.evaluate(&wtxn, &index).map_err(|err| match err {
+            milli::Error::UserError(milli::UserError::InvalidFilter(_)) => {
+                Error::from(err).with_custom_error_code(Code::InvalidDocumentFilter)
+            }
+            e => e.into(),
+        })?;
         let mut delete_operation = DeleteDocuments::new(&mut wtxn, &index)?;
         delete_operation.delete_documents(&candidates);
         let deleted_documents =

index-scheduler/src/error.rs

@@ -46,6 +46,8 @@ impl From<DateField> for Code {
 #[allow(clippy::large_enum_variant)]
 #[derive(Error, Debug)]
 pub enum Error {
+    #[error("{1}")]
+    WithCustomErrorCode(Code, Box<Self>),
     #[error("Index `{0}` not found.")]
     IndexNotFound(String),
     #[error("Index `{0}` already exists.")]
@@ -144,6 +146,7 @@ impl Error {
     pub fn is_recoverable(&self) -> bool {
         match self {
             Error::IndexNotFound(_)
+            | Error::WithCustomErrorCode(_, _)
             | Error::IndexAlreadyExists(_)
             | Error::SwapDuplicateIndexFound(_)
             | Error::SwapDuplicateIndexesFound(_)
@@ -176,11 +179,16 @@ impl Error {
             Error::PlannedFailure => false,
         }
     }
+
+    pub fn with_custom_error_code(self, code: Code) -> Self {
+        Self::WithCustomErrorCode(code, Box::new(self))
+    }
 }
 
 impl ErrorCode for Error {
     fn error_code(&self) -> Code {
         match self {
+            Error::WithCustomErrorCode(code, _) => *code,
             Error::IndexNotFound(_) => Code::IndexNotFound,
             Error::IndexAlreadyExists(_) => Code::IndexAlreadyExists,
             Error::SwapDuplicateIndexesFound(_) => Code::InvalidSwapDuplicateIndexFound,

index-scheduler/src/index_mapper/index_map.rs

@@ -5,6 +5,7 @@ use std::collections::BTreeMap;
 use std::path::Path;
 use std::time::Duration;
 
+use meilisearch_types::heed::flags::Flags;
 use meilisearch_types::heed::{EnvClosingEvent, EnvOpenOptions};
 use meilisearch_types::milli::Index;
 use time::OffsetDateTime;
@@ -53,6 +54,7 @@ pub struct IndexMap {
 pub struct ClosingIndex {
     uuid: Uuid,
     closing_event: EnvClosingEvent,
+    enable_mdb_writemap: bool,
     map_size: usize,
     generation: usize,
 }
@@ -68,6 +70,7 @@ impl ClosingIndex {
     pub fn wait_timeout(self, timeout: Duration) -> Option<ReopenableIndex> {
         self.closing_event.wait_timeout(timeout).then_some(ReopenableIndex {
             uuid: self.uuid,
+            enable_mdb_writemap: self.enable_mdb_writemap,
             map_size: self.map_size,
             generation: self.generation,
         })
@@ -76,6 +79,7 @@ impl ClosingIndex {
 
 pub struct ReopenableIndex {
     uuid: Uuid,
+    enable_mdb_writemap: bool,
     map_size: usize,
     generation: usize,
 }
@@ -103,7 +107,7 @@ impl ReopenableIndex {
                 return Ok(());
             }
             map.unavailable.remove(&self.uuid);
-            map.create(&self.uuid, path, None, self.map_size)?;
+            map.create(&self.uuid, path, None, self.enable_mdb_writemap, self.map_size)?;
         }
         Ok(())
     }
@@ -170,16 +174,17 @@ impl IndexMap {
         uuid: &Uuid,
         path: &Path,
         date: Option<(OffsetDateTime, OffsetDateTime)>,
+        enable_mdb_writemap: bool,
         map_size: usize,
     ) -> Result<Index> {
         if !matches!(self.get_unavailable(uuid), Missing) {
             panic!("Attempt to open an index that was unavailable");
         }
-        let index = create_or_open_index(path, date, map_size)?;
+        let index = create_or_open_index(path, date, enable_mdb_writemap, map_size)?;
         match self.available.insert(*uuid, index.clone()) {
             InsertionOutcome::InsertedNew => (),
             InsertionOutcome::Evicted(evicted_uuid, evicted_index) => {
-                self.close(evicted_uuid, evicted_index, 0);
+                self.close(evicted_uuid, evicted_index, enable_mdb_writemap, 0);
             }
             InsertionOutcome::Replaced(_) => {
                 panic!("Attempt to open an index that was already opened")
@@ -212,17 +217,30 @@ impl IndexMap {
     /// | Closing         | Closing       |
     /// | Available       | Closing       |
     ///
-    pub fn close_for_resize(&mut self, uuid: &Uuid, map_size_growth: usize) {
+    pub fn close_for_resize(
+        &mut self,
+        uuid: &Uuid,
+        enable_mdb_writemap: bool,
+        map_size_growth: usize,
+    ) {
         let Some(index) = self.available.remove(uuid) else { return; };
-        self.close(*uuid, index, map_size_growth);
+        self.close(*uuid, index, enable_mdb_writemap, map_size_growth);
     }
 
-    fn close(&mut self, uuid: Uuid, index: Index, map_size_growth: usize) {
+    fn close(
+        &mut self,
+        uuid: Uuid,
+        index: Index,
+        enable_mdb_writemap: bool,
+        map_size_growth: usize,
+    ) {
         let map_size = index.map_size().unwrap_or(DEFAULT_MAP_SIZE) + map_size_growth;
         let closing_event = index.prepare_for_closing();
         let generation = self.next_generation();
-        self.unavailable
-            .insert(uuid, Some(ClosingIndex { uuid, closing_event, map_size, generation }));
+        self.unavailable.insert(
+            uuid,
+            Some(ClosingIndex { uuid, closing_event, enable_mdb_writemap, map_size, generation }),
+        );
     }
 
     /// Attempts to delete and index.
@@ -282,11 +300,15 @@ impl IndexMap {
 fn create_or_open_index(
     path: &Path,
     date: Option<(OffsetDateTime, OffsetDateTime)>,
+    enable_mdb_writemap: bool,
     map_size: usize,
 ) -> Result<Index> {
     let mut options = EnvOpenOptions::new();
     options.map_size(clamp_to_page_size(map_size));
     options.max_readers(1024);
+    if enable_mdb_writemap {
+        unsafe { options.flag(Flags::MdbWriteMap) };
+    }
 
     if let Some((created, updated)) = date {
         Ok(Index::new_with_creation_dates(options, path, created, updated)?)

index-scheduler/src/index_mapper/mod.rs

@@ -66,6 +66,8 @@ pub struct IndexMapper {
     index_base_map_size: usize,
     /// The quantity by which the map size of an index is incremented upon reopening, in bytes.
     index_growth_amount: usize,
+    /// Whether we open a meilisearch index with the MDB_WRITEMAP option or not.
+    enable_mdb_writemap: bool,
     pub indexer_config: Arc<IndexerConfig>,
 }
 
@@ -88,8 +90,17 @@ pub enum IndexStatus {
 pub struct IndexStats {
     /// Number of documents in the index.
     pub number_of_documents: u64,
-    /// Size of the index' DB, in bytes.
+    /// Size taken up by the index' DB, in bytes.
+    ///
+    /// This includes the size taken by both the used and free pages of the DB, and as the free pages
+    /// are not returned to the disk after a deletion, this number is typically larger than
+    /// `used_database_size` that only includes the size of the used pages.
     pub database_size: u64,
+    /// Size taken by the used pages of the index' DB, in bytes.
+    ///
+    /// As the DB backend does not return to the disk the pages that are not currently used by the DB,
+    /// this value is typically smaller than `database_size`.
+    pub used_database_size: u64,
     /// Association of every field name with the number of times it occurs in the documents.
     pub field_distribution: FieldDistribution,
     /// Creation date of the index.
@@ -105,10 +116,10 @@ impl IndexStats {
     ///
     /// - rtxn: a RO transaction for the index, obtained from `Index::read_txn()`.
     pub fn new(index: &Index, rtxn: &RoTxn) -> Result<Self> {
-        let database_size = index.on_disk_size()?;
         Ok(IndexStats {
             number_of_documents: index.number_of_documents(rtxn)?,
-            database_size,
+            database_size: index.on_disk_size()?,
+            used_database_size: index.used_size()?,
             field_distribution: index.field_distribution(rtxn)?,
             created_at: index.created_at(rtxn)?,
             updated_at: index.updated_at(rtxn)?,
@@ -123,15 +134,22 @@ impl IndexMapper {
         index_base_map_size: usize,
         index_growth_amount: usize,
         index_count: usize,
+        enable_mdb_writemap: bool,
         indexer_config: IndexerConfig,
     ) -> Result<Self> {
+        let mut wtxn = env.write_txn()?;
+        let index_mapping = env.create_database(&mut wtxn, Some(INDEX_MAPPING))?;
+        let index_stats = env.create_database(&mut wtxn, Some(INDEX_STATS))?;
+        wtxn.commit()?;
+
         Ok(Self {
             index_map: Arc::new(RwLock::new(IndexMap::new(index_count))),
-            index_mapping: env.create_database(Some(INDEX_MAPPING))?,
-            index_stats: env.create_database(Some(INDEX_STATS))?,
+            index_mapping,
+            index_stats,
             base_path,
             index_base_map_size,
             index_growth_amount,
+            enable_mdb_writemap,
             indexer_config: Arc::new(indexer_config),
         })
     }
@@ -162,6 +180,7 @@ impl IndexMapper {
                     &uuid,
                     &index_path,
                     date,
+                    self.enable_mdb_writemap,
                     self.index_base_map_size,
                 )?;
 
@@ -273,7 +292,11 @@ impl IndexMapper {
             .ok_or_else(|| Error::IndexNotFound(name.to_string()))?;
 
         // We remove the index from the in-memory index map.
-        self.index_map.write().unwrap().close_for_resize(&uuid, self.index_growth_amount);
+        self.index_map.write().unwrap().close_for_resize(
+            &uuid,
+            self.enable_mdb_writemap,
+            self.index_growth_amount,
+        );
 
         Ok(())
     }
@@ -338,6 +361,7 @@ impl IndexMapper {
                                 &uuid,
                                 &index_path,
                                 None,
+                                self.enable_mdb_writemap,
                                 self.index_base_map_size,
                             )?;
                         }

index-scheduler/src/lib.rs

@@ -31,7 +31,7 @@ mod uuid_codec;
 pub type Result<T> = std::result::Result<T, Error>;
 pub type TaskId = u32;
 
-use std::collections::HashMap;
+use std::collections::{BTreeMap, HashMap};
 use std::ops::{Bound, RangeBounds};
 use std::path::{Path, PathBuf};
 use std::sync::atomic::AtomicBool;
@@ -233,6 +233,8 @@ pub struct IndexSchedulerOptions {
     pub task_db_size: usize,
     /// The size, in bytes, with which a meilisearch index is opened the first time of each meilisearch index.
     pub index_base_map_size: usize,
+    /// Whether we open a meilisearch index with the MDB_WRITEMAP option or not.
+    pub enable_mdb_writemap: bool,
     /// The size, in bytes, by which the map size of an index is increased when it resized due to being full.
     pub index_growth_amount: usize,
     /// The number of indexes that can be concurrently opened in memory.
@@ -374,6 +376,11 @@ impl IndexScheduler {
         std::fs::create_dir_all(&options.indexes_path)?;
         std::fs::create_dir_all(&options.dumps_path)?;
 
+        if cfg!(windows) && options.enable_mdb_writemap {
+            // programmer error if this happens: in normal use passing the option on Windows is an error in main
+            panic!("Windows doesn't support the MDB_WRITEMAP LMDB option");
+        }
+
         let task_db_size = clamp_to_page_size(options.task_db_size);
         let budget = if options.indexer_config.skip_index_budget {
             IndexBudget {
@@ -396,25 +403,37 @@ impl IndexScheduler {
             .open(options.tasks_path)?;
         let file_store = FileStore::new(&options.update_file_path)?;
 
+        let mut wtxn = env.write_txn()?;
+        let all_tasks = env.create_database(&mut wtxn, Some(db_name::ALL_TASKS))?;
+        let status = env.create_database(&mut wtxn, Some(db_name::STATUS))?;
+        let kind = env.create_database(&mut wtxn, Some(db_name::KIND))?;
+        let index_tasks = env.create_database(&mut wtxn, Some(db_name::INDEX_TASKS))?;
+        let canceled_by = env.create_database(&mut wtxn, Some(db_name::CANCELED_BY))?;
+        let enqueued_at = env.create_database(&mut wtxn, Some(db_name::ENQUEUED_AT))?;
+        let started_at = env.create_database(&mut wtxn, Some(db_name::STARTED_AT))?;
+        let finished_at = env.create_database(&mut wtxn, Some(db_name::FINISHED_AT))?;
+        wtxn.commit()?;
+
         // allow unreachable_code to get rids of the warning in the case of a test build.
         let this = Self {
             must_stop_processing: MustStopProcessing::default(),
             processing_tasks: Arc::new(RwLock::new(ProcessingTasks::new())),
             file_store,
-            all_tasks: env.create_database(Some(db_name::ALL_TASKS))?,
-            status: env.create_database(Some(db_name::STATUS))?,
-            kind: env.create_database(Some(db_name::KIND))?,
-            index_tasks: env.create_database(Some(db_name::INDEX_TASKS))?,
-            canceled_by: env.create_database(Some(db_name::CANCELED_BY))?,
-            enqueued_at: env.create_database(Some(db_name::ENQUEUED_AT))?,
-            started_at: env.create_database(Some(db_name::STARTED_AT))?,
-            finished_at: env.create_database(Some(db_name::FINISHED_AT))?,
+            all_tasks,
+            status,
+            kind,
+            index_tasks,
+            canceled_by,
+            enqueued_at,
+            started_at,
+            finished_at,
             index_mapper: IndexMapper::new(
                 &env,
                 options.indexes_path,
                 budget.map_size,
                 options.index_growth_amount,
                 budget.index_count,
+                options.enable_mdb_writemap,
                 options.indexer_config,
             )?,
             env,
@@ -554,10 +573,16 @@ impl IndexScheduler {
         &self.index_mapper.indexer_config
     }
 
+    /// Return the real database size (i.e.: The size **with** the free pages)
     pub fn size(&self) -> Result<u64> {
         Ok(self.env.real_disk_size()?)
     }
 
+    /// Return the used database size (i.e.: The size **without** the free pages)
+    pub fn used_size(&self) -> Result<u64> {
+        Ok(self.env.non_free_pages_size()?)
+    }
+
     /// Return the index corresponding to the name.
     ///
     /// * If the index wasn't opened before, the index will be opened.
@@ -737,6 +762,38 @@ impl IndexScheduler {
         Ok(tasks)
     }
 
+    /// The returned structure contains:
+    /// 1. The name of the property being observed can be `statuses`, `types`, or `indexes`.
+    /// 2. The name of the specific data related to the property can be `enqueued` for the `statuses`, `settingsUpdate` for the `types`, or the name of the index for the `indexes`, for example.
+    /// 3. The number of times the properties appeared.
+    pub fn get_stats(&self) -> Result<BTreeMap<String, BTreeMap<String, u64>>> {
+        let rtxn = self.read_txn()?;
+
+        let mut res = BTreeMap::new();
+
+        res.insert(
+            "statuses".to_string(),
+            enum_iterator::all::<Status>()
+                .map(|s| Ok((s.to_string(), self.get_status(&rtxn, s)?.len())))
+                .collect::<Result<BTreeMap<String, u64>>>()?,
+        );
+        res.insert(
+            "types".to_string(),
+            enum_iterator::all::<Kind>()
+                .map(|s| Ok((s.to_string(), self.get_kind(&rtxn, s)?.len())))
+                .collect::<Result<BTreeMap<String, u64>>>()?,
+        );
+        res.insert(
+            "indexes".to_string(),
+            self.index_tasks
+                .iter(&rtxn)?
+                .map(|res| Ok(res.map(|(name, bitmap)| (name.to_string(), bitmap.len()))?))
+                .collect::<Result<BTreeMap<String, u64>>>()?,
+        );
+
+        Ok(res)
+    }
+
     /// Return true iff there is at least one task associated with this index
     /// that is processing.
     pub fn is_index_processing(&self, index: &str) -> Result<bool> {
@@ -1471,6 +1528,7 @@ mod tests {
                 dumps_path: tempdir.path().join("dumps"),
                 task_db_size: 1000 * 1000, // 1 MB, we don't use MiB on purpose.
                 index_base_map_size: 1000 * 1000, // 1 MB, we don't use MiB on purpose.
+                enable_mdb_writemap: false,
                 index_growth_amount: 1000 * 1000, // 1 MB
                 index_count: 5,
                 indexer_config,

index-scheduler/src/utils.rs

@@ -466,7 +466,7 @@ impl IndexScheduler {
                         }
                     }
                     Details::DocumentDeletionByFilter { deleted_documents, original_filter: _ } => {
-                        assert_eq!(kind.as_kind(), Kind::DocumentDeletionByFilter);
+                        assert_eq!(kind.as_kind(), Kind::DocumentDeletion);
                         let (index_uid, _) = if let KindWithContent::DocumentDeletionByFilter {
                             ref index_uid,
                             ref filter_expr,

meilisearch-auth/src/lib.rs

@@ -45,6 +45,11 @@ impl AuthController {
         self.store.size()
     }
 
+    /// Return the used size of the `AuthController` database in bytes.
+    pub fn used_size(&self) -> Result<u64> {
+        self.store.used_size()
+    }
+
     pub fn create_key(&self, create_key: CreateApiKey) -> Result<Key> {
         match self.store.get_api_key(create_key.uid)? {
             Some(_) => Err(AuthControllerError::ApiKeyAlreadyExists(create_key.uid.to_string())),

meilisearch-auth/src/store.rs

@@ -55,9 +55,11 @@ impl HeedAuthStore {
         let path = path.as_ref().join(AUTH_DB_PATH);
         create_dir_all(&path)?;
         let env = Arc::new(open_auth_store_env(path.as_ref())?);
-        let keys = env.create_database(Some(KEY_DB_NAME))?;
+        let mut wtxn = env.write_txn()?;
+        let keys = env.create_database(&mut wtxn, Some(KEY_DB_NAME))?;
         let action_keyid_index_expiration =
-            env.create_database(Some(KEY_ID_ACTION_INDEX_EXPIRATION_DB_NAME))?;
+            env.create_database(&mut wtxn, Some(KEY_ID_ACTION_INDEX_EXPIRATION_DB_NAME))?;
+        wtxn.commit()?;
         Ok(Self { env, keys, action_keyid_index_expiration, should_close_on_drop: true })
     }
 
@@ -73,6 +75,11 @@ impl HeedAuthStore {
         Ok(self.env.real_disk_size()?)
     }
 
+    /// Return the number of bytes actually used in the database
+    pub fn used_size(&self) -> Result<u64> {
+        Ok(self.env.non_free_pages_size()?)
+    }
+
     pub fn set_drop_on_close(&mut self, v: bool) {
         self.should_close_on_drop = v;
     }

meilisearch-types/src/deserr/mod.rs

@@ -150,6 +150,7 @@ make_missing_field_convenience_builder!(MissingApiKeyActions, missing_api_key_ac
 make_missing_field_convenience_builder!(MissingApiKeyExpiresAt, missing_api_key_expires_at);
 make_missing_field_convenience_builder!(MissingApiKeyIndexes, missing_api_key_indexes);
 make_missing_field_convenience_builder!(MissingSwapIndexes, missing_swap_indexes);
+make_missing_field_convenience_builder!(MissingDocumentFilter, missing_document_filter);
 
 // Integrate a sub-error into a [`DeserrError`] by taking its error message but using
 // the default error code (C) from `Self`

meilisearch-types/src/error.rs

@@ -214,12 +214,14 @@ InvalidApiKeyUid                      , InvalidRequest       , BAD_REQUEST ;
 InvalidContentType                    , InvalidRequest       , UNSUPPORTED_MEDIA_TYPE ;
 InvalidDocumentCsvDelimiter           , InvalidRequest       , BAD_REQUEST ;
 InvalidDocumentFields                 , InvalidRequest       , BAD_REQUEST ;
+MissingDocumentFilter                 , InvalidRequest       , BAD_REQUEST ;
 InvalidDocumentFilter                 , InvalidRequest       , BAD_REQUEST ;
 InvalidDocumentGeoField               , InvalidRequest       , BAD_REQUEST ;
+InvalidVectorDimensions               , InvalidRequest       , BAD_REQUEST ;
+InvalidVectorsType                    , InvalidRequest       , BAD_REQUEST ;
 InvalidDocumentId                     , InvalidRequest       , BAD_REQUEST ;
 InvalidDocumentLimit                  , InvalidRequest       , BAD_REQUEST ;
 InvalidDocumentOffset                 , InvalidRequest       , BAD_REQUEST ;
-InvalidDocumentDeleteFilter           , InvalidRequest       , BAD_REQUEST ;
 InvalidIndexLimit                     , InvalidRequest       , BAD_REQUEST ;
 InvalidIndexOffset                    , InvalidRequest       , BAD_REQUEST ;
 InvalidIndexPrimaryKey                , InvalidRequest       , BAD_REQUEST ;
@@ -332,6 +334,8 @@ impl ErrorCode for milli::Error {
                     UserError::InvalidSortableAttribute { .. } => Code::InvalidSearchSort,
                     UserError::CriterionError(_) => Code::InvalidSettingsRankingRules,
                     UserError::InvalidGeoField { .. } => Code::InvalidDocumentGeoField,
+                    UserError::InvalidVectorDimensions { .. } => Code::InvalidVectorDimensions,
+                    UserError::InvalidVectorsType { .. } => Code::InvalidVectorsType,
                     UserError::SortError(_) => Code::InvalidSearchSort,
                     UserError::InvalidMinTypoWordLenSetting(_, _) => {
                         Code::InvalidSettingsTypoTolerance

meilisearch-types/src/tasks.rs

@@ -395,7 +395,6 @@ impl std::error::Error for ParseTaskStatusError {}
 pub enum Kind {
     DocumentAdditionOrUpdate,
     DocumentDeletion,
-    DocumentDeletionByFilter,
     SettingsUpdate,
     IndexCreation,
     IndexDeletion,
@@ -412,7 +411,6 @@ impl Kind {
         match self {
             Kind::DocumentAdditionOrUpdate
             | Kind::DocumentDeletion
-            | Kind::DocumentDeletionByFilter
             | Kind::SettingsUpdate
             | Kind::IndexCreation
             | Kind::IndexDeletion
@@ -430,7 +428,6 @@ impl Display for Kind {
         match self {
             Kind::DocumentAdditionOrUpdate => write!(f, "documentAdditionOrUpdate"),
             Kind::DocumentDeletion => write!(f, "documentDeletion"),
-            Kind::DocumentDeletionByFilter => write!(f, "documentDeletionByFilter"),
             Kind::SettingsUpdate => write!(f, "settingsUpdate"),
             Kind::IndexCreation => write!(f, "indexCreation"),
             Kind::IndexDeletion => write!(f, "indexDeletion"),

meilisearch/Cargo.toml

@@ -48,6 +48,7 @@ mime = "0.3.17"
 num_cpus = "1.15.0"
 obkv = "0.2.0"
 once_cell = "1.17.1"
+ordered-float = "3.7.0"
 parking_lot = "0.12.1"
 permissive-json-pointer = { path = "../permissive-json-pointer" }
 pin-project-lite = "0.2.9"

meilisearch/src/analytics/mock_analytics.rs

@@ -5,7 +5,7 @@ use actix_web::HttpRequest;
 use meilisearch_types::InstanceUid;
 use serde_json::Value;
 
-use super::{find_user_id, Analytics, DocumentDeletionKind};
+use super::{find_user_id, Analytics, DocumentDeletionKind, DocumentFetchKind};
 use crate::routes::indexes::documents::UpdateDocumentsQuery;
 use crate::routes::tasks::TasksFilterQuery;
 use crate::Opt;
@@ -71,6 +71,8 @@ impl Analytics for MockAnalytics {
         _request: &HttpRequest,
     ) {
     }
+    fn get_fetch_documents(&self, _documents_query: &DocumentFetchKind, _request: &HttpRequest) {}
+    fn post_fetch_documents(&self, _documents_query: &DocumentFetchKind, _request: &HttpRequest) {}
     fn get_tasks(&self, _query: &TasksFilterQuery, _request: &HttpRequest) {}
     fn health_seen(&self, _request: &HttpRequest) {}
 }

meilisearch/src/analytics/mod.rs

@@ -67,6 +67,12 @@ pub enum DocumentDeletionKind {
     PerFilter,
 }
 
+#[derive(Copy, Clone, Debug, PartialEq, Eq)]
+pub enum DocumentFetchKind {
+    PerDocumentId,
+    Normal { with_filter: bool, limit: usize, offset: usize },
+}
+
 pub trait Analytics: Sync + Send {
     fn instance_uid(&self) -> Option<&InstanceUid>;
 
@@ -90,6 +96,12 @@ pub trait Analytics: Sync + Send {
         request: &HttpRequest,
     );
 
+    // this method should be called to aggregate a fetch documents request
+    fn get_fetch_documents(&self, documents_query: &DocumentFetchKind, request: &HttpRequest);
+
+    // this method should be called to aggregate a fetch documents request
+    fn post_fetch_documents(&self, documents_query: &DocumentFetchKind, request: &HttpRequest);
+
     // this method should be called to aggregate a add documents request
     fn delete_documents(&self, kind: DocumentDeletionKind, request: &HttpRequest);
 

meilisearch/src/analytics/segment_analytics.rs

@@ -23,7 +23,9 @@ use tokio::select;
 use tokio::sync::mpsc::{self, Receiver, Sender};
 use uuid::Uuid;
 
-use super::{config_user_id_path, DocumentDeletionKind, MEILISEARCH_CONFIG_PATH};
+use super::{
+    config_user_id_path, DocumentDeletionKind, DocumentFetchKind, MEILISEARCH_CONFIG_PATH,
+};
 use crate::analytics::Analytics;
 use crate::option::{default_http_addr, IndexerOpts, MaxMemory, MaxThreads, ScheduleSnapshot};
 use crate::routes::indexes::documents::UpdateDocumentsQuery;
@@ -72,6 +74,8 @@ pub enum AnalyticsMsg {
     AggregateAddDocuments(DocumentsAggregator),
     AggregateDeleteDocuments(DocumentsDeletionAggregator),
     AggregateUpdateDocuments(DocumentsAggregator),
+    AggregateGetFetchDocuments(DocumentsFetchAggregator),
+    AggregatePostFetchDocuments(DocumentsFetchAggregator),
     AggregateTasks(TasksAggregator),
     AggregateHealth(HealthAggregator),
 }
@@ -139,6 +143,8 @@ impl SegmentAnalytics {
             add_documents_aggregator: DocumentsAggregator::default(),
             delete_documents_aggregator: DocumentsDeletionAggregator::default(),
             update_documents_aggregator: DocumentsAggregator::default(),
+            get_fetch_documents_aggregator: DocumentsFetchAggregator::default(),
+            post_fetch_documents_aggregator: DocumentsFetchAggregator::default(),
             get_tasks_aggregator: TasksAggregator::default(),
             health_aggregator: HealthAggregator::default(),
         });
@@ -205,6 +211,16 @@ impl super::Analytics for SegmentAnalytics {
         let _ = self.sender.try_send(AnalyticsMsg::AggregateUpdateDocuments(aggregate));
     }
 
+    fn get_fetch_documents(&self, documents_query: &DocumentFetchKind, request: &HttpRequest) {
+        let aggregate = DocumentsFetchAggregator::from_query(documents_query, request);
+        let _ = self.sender.try_send(AnalyticsMsg::AggregateGetFetchDocuments(aggregate));
+    }
+
+    fn post_fetch_documents(&self, documents_query: &DocumentFetchKind, request: &HttpRequest) {
+        let aggregate = DocumentsFetchAggregator::from_query(documents_query, request);
+        let _ = self.sender.try_send(AnalyticsMsg::AggregatePostFetchDocuments(aggregate));
+    }
+
     fn get_tasks(&self, query: &TasksFilterQuery, request: &HttpRequest) {
         let aggregate = TasksAggregator::from_query(query, request);
         let _ = self.sender.try_send(AnalyticsMsg::AggregateTasks(aggregate));
@@ -225,6 +241,7 @@ impl super::Analytics for SegmentAnalytics {
 struct Infos {
     env: String,
     experimental_enable_metrics: bool,
+    experimental_reduce_indexing_memory_usage: bool,
     db_path: bool,
     import_dump: bool,
     dump_dir: bool,
@@ -258,6 +275,7 @@ impl From<Opt> for Infos {
         let Opt {
             db_path,
             experimental_enable_metrics,
+            experimental_reduce_indexing_memory_usage,
             http_addr,
             master_key: _,
             env,
@@ -300,6 +318,7 @@ impl From<Opt> for Infos {
         Self {
             env,
             experimental_enable_metrics,
+            experimental_reduce_indexing_memory_usage,
             db_path: db_path != PathBuf::from("./data.ms"),
             import_dump: import_dump.is_some(),
             dump_dir: dump_dir != PathBuf::from("dumps/"),
@@ -338,6 +357,8 @@ pub struct Segment {
     add_documents_aggregator: DocumentsAggregator,
     delete_documents_aggregator: DocumentsDeletionAggregator,
     update_documents_aggregator: DocumentsAggregator,
+    get_fetch_documents_aggregator: DocumentsFetchAggregator,
+    post_fetch_documents_aggregator: DocumentsFetchAggregator,
     get_tasks_aggregator: TasksAggregator,
     health_aggregator: HealthAggregator,
 }
@@ -400,6 +421,8 @@ impl Segment {
                         Some(AnalyticsMsg::AggregateAddDocuments(agreg)) => self.add_documents_aggregator.aggregate(agreg),
                         Some(AnalyticsMsg::AggregateDeleteDocuments(agreg)) => self.delete_documents_aggregator.aggregate(agreg),
                         Some(AnalyticsMsg::AggregateUpdateDocuments(agreg)) => self.update_documents_aggregator.aggregate(agreg),
+                        Some(AnalyticsMsg::AggregateGetFetchDocuments(agreg)) => self.get_fetch_documents_aggregator.aggregate(agreg),
+                        Some(AnalyticsMsg::AggregatePostFetchDocuments(agreg)) => self.post_fetch_documents_aggregator.aggregate(agreg),
                         Some(AnalyticsMsg::AggregateTasks(agreg)) => self.get_tasks_aggregator.aggregate(agreg),
                         Some(AnalyticsMsg::AggregateHealth(agreg)) => self.health_aggregator.aggregate(agreg),
                         None => (),
@@ -450,6 +473,10 @@ impl Segment {
             .into_event(&self.user, "Documents Deleted");
         let update_documents = std::mem::take(&mut self.update_documents_aggregator)
             .into_event(&self.user, "Documents Updated");
+        let get_fetch_documents = std::mem::take(&mut self.get_fetch_documents_aggregator)
+            .into_event(&self.user, "Documents Fetched GET");
+        let post_fetch_documents = std::mem::take(&mut self.post_fetch_documents_aggregator)
+            .into_event(&self.user, "Documents Fetched POST");
         let get_tasks =
             std::mem::take(&mut self.get_tasks_aggregator).into_event(&self.user, "Tasks Seen");
         let health =
@@ -473,6 +500,12 @@ impl Segment {
         if let Some(update_documents) = update_documents {
             let _ = self.batcher.push(update_documents).await;
         }
+        if let Some(get_fetch_documents) = get_fetch_documents {
+            let _ = self.batcher.push(get_fetch_documents).await;
+        }
+        if let Some(post_fetch_documents) = post_fetch_documents {
+            let _ = self.batcher.push(post_fetch_documents).await;
+        }
         if let Some(get_tasks) = get_tasks {
             let _ = self.batcher.push(get_tasks).await;
         }
@@ -1135,3 +1168,76 @@ impl HealthAggregator {
         })
     }
 }
+
+#[derive(Default, Serialize)]
+pub struct DocumentsFetchAggregator {
+    #[serde(skip)]
+    timestamp: Option<OffsetDateTime>,
+
+    // context
+    #[serde(rename = "user-agent")]
+    user_agents: HashSet<String>,
+
+    #[serde(rename = "requests.max_limit")]
+    total_received: usize,
+
+    // a call on ../documents/:doc_id
+    per_document_id: bool,
+    // if a filter was used
+    per_filter: bool,
+
+    // pagination
+    #[serde(rename = "pagination.max_limit")]
+    max_limit: usize,
+    #[serde(rename = "pagination.max_offset")]
+    max_offset: usize,
+}
+
+impl DocumentsFetchAggregator {
+    pub fn from_query(query: &DocumentFetchKind, request: &HttpRequest) -> Self {
+        let (limit, offset) = match query {
+            DocumentFetchKind::PerDocumentId => (1, 0),
+            DocumentFetchKind::Normal { limit, offset, .. } => (*limit, *offset),
+        };
+        Self {
+            timestamp: Some(OffsetDateTime::now_utc()),
+            user_agents: extract_user_agents(request).into_iter().collect(),
+            total_received: 1,
+            per_document_id: matches!(query, DocumentFetchKind::PerDocumentId),
+            per_filter: matches!(query, DocumentFetchKind::Normal { with_filter, .. } if *with_filter),
+            max_limit: limit,
+            max_offset: offset,
+        }
+    }
+
+    /// Aggregate one [DocumentsFetchAggregator] into another.
+    pub fn aggregate(&mut self, other: Self) {
+        if self.timestamp.is_none() {
+            self.timestamp = other.timestamp;
+        }
+        for user_agent in other.user_agents {
+            self.user_agents.insert(user_agent);
+        }
+
+        self.total_received = self.total_received.saturating_add(other.total_received);
+        self.per_document_id |= other.per_document_id;
+        self.per_filter |= other.per_filter;
+
+        self.max_limit = self.max_limit.max(other.max_limit);
+        self.max_offset = self.max_offset.max(other.max_offset);
+    }
+
+    pub fn into_event(self, user: &User, event_name: &str) -> Option<Track> {
+        // if we had no timestamp it means we never encountered any events and
+        // thus we don't need to send this event.
+        let timestamp = self.timestamp?;
+
+        Some(Track {
+            timestamp: Some(timestamp),
+            user: user.clone(),
+            event: event_name.to_string(),
+            properties: serde_json::to_value(self).ok()?,
+            ..Default::default()
+        })
+    }
+}

meilisearch/src/error.rs

@@ -1,5 +1,6 @@
 use actix_web as aweb;
 use aweb::error::{JsonPayloadError, QueryPayloadError};
+use byte_unit::Byte;
 use meilisearch_types::document_formats::{DocumentFormatError, PayloadType};
 use meilisearch_types::error::{Code, ErrorCode, ResponseError};
 use meilisearch_types::index_uid::{IndexUid, IndexUidFormatError};
@@ -26,8 +27,8 @@ pub enum MeilisearchHttpError {
     InvalidExpression(&'static [&'static str], Value),
     #[error("A {0} payload is missing.")]
     MissingPayload(PayloadType),
-    #[error("The provided payload reached the size limit.")]
-    PayloadTooLarge,
+    #[error("The provided payload reached the size limit. The maximum accepted payload size is {}.",  Byte::from_bytes(*.0 as u64).get_appropriate_unit(true))]
+    PayloadTooLarge(usize),
     #[error("Two indexes must be given for each swap. The list `[{}]` contains {} indexes.",
         .0.iter().map(|uid| format!("\"{uid}\"")).collect::<Vec<_>>().join(", "), .0.len()
     )]
@@ -60,9 +61,9 @@ impl ErrorCode for MeilisearchHttpError {
             MeilisearchHttpError::MissingPayload(_) => Code::MissingPayload,
             MeilisearchHttpError::InvalidContentType(_, _) => Code::InvalidContentType,
             MeilisearchHttpError::DocumentNotFound(_) => Code::DocumentNotFound,
-            MeilisearchHttpError::EmptyFilter => Code::InvalidDocumentDeleteFilter,
+            MeilisearchHttpError::EmptyFilter => Code::InvalidDocumentFilter,
             MeilisearchHttpError::InvalidExpression(_, _) => Code::InvalidSearchFilter,
-            MeilisearchHttpError::PayloadTooLarge => Code::PayloadTooLarge,
+            MeilisearchHttpError::PayloadTooLarge(_) => Code::PayloadTooLarge,
             MeilisearchHttpError::SwapIndexPayloadWrongLength(_) => Code::InvalidSwapIndexes,
             MeilisearchHttpError::IndexUid(e) => e.error_code(),
             MeilisearchHttpError::SerdeJson(_) => Code::Internal,

meilisearch/src/extractors/payload.rs

@@ -11,6 +11,7 @@ use crate::error::MeilisearchHttpError;
 pub struct Payload {
     payload: Decompress<dev::Payload>,
     limit: usize,
+    remaining: usize,
 }
 
 pub struct PayloadConfig {
@@ -43,6 +44,7 @@ impl FromRequest for Payload {
         ready(Ok(Payload {
             payload: Decompress::from_headers(payload.take(), req.headers()),
             limit,
+            remaining: limit,
         }))
     }
 }
@@ -54,12 +56,14 @@ impl Stream for Payload {
     fn poll_next(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Option<Self::Item>> {
         match Pin::new(&mut self.payload).poll_next(cx) {
             Poll::Ready(Some(result)) => match result {
-                Ok(bytes) => match self.limit.checked_sub(bytes.len()) {
+                Ok(bytes) => match self.remaining.checked_sub(bytes.len()) {
                     Some(new_limit) => {
-                        self.limit = new_limit;
+                        self.remaining = new_limit;
                         Poll::Ready(Some(Ok(bytes)))
                     }
-                    None => Poll::Ready(Some(Err(MeilisearchHttpError::PayloadTooLarge))),
+                    None => {
+                        Poll::Ready(Some(Err(MeilisearchHttpError::PayloadTooLarge(self.limit))))
+                    }
                 },
                 x => Poll::Ready(Some(x.map_err(MeilisearchHttpError::from))),
             },

meilisearch/src/lib.rs

@@ -232,6 +232,7 @@ fn open_or_create_database_unchecked(
             dumps_path: opt.dump_dir.clone(),
             task_db_size: opt.max_task_db_size.get_bytes() as usize,
             index_base_map_size: opt.max_index_size.get_bytes() as usize,
+            enable_mdb_writemap: opt.experimental_reduce_indexing_memory_usage,
             indexer_config: (&opt.indexer_options).try_into()?,
             autobatching_enabled: true,
             max_number_of_tasks: 1_000_000,

meilisearch/src/main.rs

@@ -29,6 +29,11 @@ fn setup(opt: &Opt) -> anyhow::Result<()> {
 async fn main() -> anyhow::Result<()> {
     let (opt, config_read_from) = Opt::try_build()?;
 
+    anyhow::ensure!(
+        !(cfg!(windows) && opt.experimental_reduce_indexing_memory_usage),
+        "The `experimental-reduce-indexing-memory-usage` flag is not supported on Windows"
+    );
+
     setup(&opt)?;
 
     match (opt.env.as_ref(), &opt.master_key) {

meilisearch/src/metrics.rs

@@ -4,20 +4,32 @@ use prometheus::{
     register_int_gauge_vec, HistogramVec, IntCounterVec, IntGauge, IntGaugeVec,
 };
 
-const HTTP_RESPONSE_TIME_CUSTOM_BUCKETS: &[f64; 14] = &[
-    0.0005, 0.0008, 0.00085, 0.0009, 0.00095, 0.001, 0.00105, 0.0011, 0.00115, 0.0012, 0.0015,
-    0.002, 0.003, 1.0,
-];
+/// Create evenly distributed buckets
+fn create_buckets() -> [f64; 29] {
+    (0..10)
+        .chain((10..100).step_by(10))
+        .chain((100..=1000).step_by(100))
+        .map(|i| i as f64 / 1000.)
+        .collect::<Vec<_>>()
+        .try_into()
+        .unwrap()
+}
 
 lazy_static! {
-    pub static ref HTTP_REQUESTS_TOTAL: IntCounterVec = register_int_counter_vec!(
-        opts!("http_requests_total", "HTTP requests total"),
+    pub static ref HTTP_RESPONSE_TIME_CUSTOM_BUCKETS: [f64; 29] = create_buckets();
+    pub static ref MEILISEARCH_HTTP_REQUESTS_TOTAL: IntCounterVec = register_int_counter_vec!(
+        opts!("meilisearch_http_requests_total", "Meilisearch HTTP requests total"),
         &["method", "path"]
     )
     .expect("Can't create a metric");
     pub static ref MEILISEARCH_DB_SIZE_BYTES: IntGauge =
-        register_int_gauge!(opts!("meilisearch_db_size_bytes", "Meilisearch Db Size In Bytes"))
+        register_int_gauge!(opts!("meilisearch_db_size_bytes", "Meilisearch DB Size In Bytes"))
             .expect("Can't create a metric");
+    pub static ref MEILISEARCH_USED_DB_SIZE_BYTES: IntGauge = register_int_gauge!(opts!(
+        "meilisearch_used_db_size_bytes",
+        "Meilisearch Used DB Size In Bytes"
+    ))
+    .expect("Can't create a metric");
     pub static ref MEILISEARCH_INDEX_COUNT: IntGauge =
         register_int_gauge!(opts!("meilisearch_index_count", "Meilisearch Index Count"))
             .expect("Can't create a metric");
@@ -26,11 +38,16 @@ lazy_static! {
         &["index"]
     )
     .expect("Can't create a metric");
-    pub static ref HTTP_RESPONSE_TIME_SECONDS: HistogramVec = register_histogram_vec!(
+    pub static ref MEILISEARCH_HTTP_RESPONSE_TIME_SECONDS: HistogramVec = register_histogram_vec!(
         "http_response_time_seconds",
         "HTTP response times",
         &["method", "path"],
         HTTP_RESPONSE_TIME_CUSTOM_BUCKETS.to_vec()
     )
     .expect("Can't create a metric");
+    pub static ref MEILISEARCH_NB_TASKS: IntGaugeVec = register_int_gauge_vec!(
+        opts!("meilisearch_nb_tasks", "Meilisearch Number of tasks"),
+        &["kind", "value"]
+    )
+    .expect("Can't create a metric");
 }

meilisearch/src/middleware.rs

@@ -52,11 +52,11 @@ where
         if is_registered_resource {
             let request_method = req.method().to_string();
             histogram_timer = Some(
-                crate::metrics::HTTP_RESPONSE_TIME_SECONDS
+                crate::metrics::MEILISEARCH_HTTP_RESPONSE_TIME_SECONDS
                     .with_label_values(&[&request_method, request_path])
                     .start_timer(),
             );
-            crate::metrics::HTTP_REQUESTS_TOTAL
+            crate::metrics::MEILISEARCH_HTTP_REQUESTS_TOTAL
                 .with_label_values(&[&request_method, request_path])
                 .inc();
         }

meilisearch/src/option.rs

@@ -48,6 +48,8 @@ const MEILI_IGNORE_DUMP_IF_DB_EXISTS: &str = "MEILI_IGNORE_DUMP_IF_DB_EXISTS";
 const MEILI_DUMP_DIR: &str = "MEILI_DUMP_DIR";
 const MEILI_LOG_LEVEL: &str = "MEILI_LOG_LEVEL";
 const MEILI_EXPERIMENTAL_ENABLE_METRICS: &str = "MEILI_EXPERIMENTAL_ENABLE_METRICS";
+const MEILI_EXPERIMENTAL_REDUCE_INDEXING_MEMORY_USAGE: &str =
+    "MEILI_EXPERIMENTAL_REDUCE_INDEXING_MEMORY_USAGE";
 
 const DEFAULT_CONFIG_FILE_PATH: &str = "./config.toml";
 const DEFAULT_DB_PATH: &str = "./data.ms";
@@ -293,6 +295,11 @@ pub struct Opt {
     #[serde(default)]
     pub experimental_enable_metrics: bool,
 
+    /// Experimental RAM reduction during indexing, do not use in production, see: <https://github.com/meilisearch/product/discussions/652>
+    #[clap(long, env = MEILI_EXPERIMENTAL_REDUCE_INDEXING_MEMORY_USAGE)]
+    #[serde(default)]
+    pub experimental_reduce_indexing_memory_usage: bool,
+
     #[serde(flatten)]
     #[clap(flatten)]
     pub indexer_options: IndexerOpts,
@@ -385,6 +392,7 @@ impl Opt {
             #[cfg(all(not(debug_assertions), feature = "analytics"))]
             no_analytics,
             experimental_enable_metrics: enable_metrics_route,
+            experimental_reduce_indexing_memory_usage: reduce_indexing_memory_usage,
         } = self;
         export_to_env_if_not_present(MEILI_DB_PATH, db_path);
         export_to_env_if_not_present(MEILI_HTTP_ADDR, http_addr);
@@ -426,6 +434,10 @@ impl Opt {
             MEILI_EXPERIMENTAL_ENABLE_METRICS,
             enable_metrics_route.to_string(),
         );
+        export_to_env_if_not_present(
+            MEILI_EXPERIMENTAL_REDUCE_INDEXING_MEMORY_USAGE,
+            reduce_indexing_memory_usage.to_string(),
+        );
         indexer_options.export_to_env();
     }
 

meilisearch/src/routes/indexes/documents.rs

@@ -29,7 +29,7 @@ use tempfile::tempfile;
 use tokio::fs::File;
 use tokio::io::{AsyncSeekExt, AsyncWriteExt, BufWriter};
 
-use crate::analytics::{Analytics, DocumentDeletionKind};
+use crate::analytics::{Analytics, DocumentDeletionKind, DocumentFetchKind};
 use crate::error::MeilisearchHttpError;
 use crate::error::PayloadError::ReceivePayload;
 use crate::extractors::authentication::policies::*;
@@ -97,10 +97,14 @@ pub async fn get_document(
     index_scheduler: GuardedData<ActionPolicy<{ actions::DOCUMENTS_GET }>, Data<IndexScheduler>>,
     document_param: web::Path<DocumentParam>,
     params: AwebQueryParameter<GetDocument, DeserrQueryParamError>,
+    req: HttpRequest,
+    analytics: web::Data<dyn Analytics>,
 ) -> Result<HttpResponse, ResponseError> {
     let DocumentParam { index_uid, document_id } = document_param.into_inner();
     let index_uid = IndexUid::try_from(index_uid)?;
 
+    analytics.get_fetch_documents(&DocumentFetchKind::PerDocumentId, &req);
+
     let GetDocument { fields } = params.into_inner();
     let attributes_to_retrieve = fields.merge_star_and_none();
 
@@ -161,16 +165,31 @@ pub async fn documents_by_query_post(
     index_scheduler: GuardedData<ActionPolicy<{ actions::DOCUMENTS_GET }>, Data<IndexScheduler>>,
     index_uid: web::Path<String>,
     body: AwebJson<BrowseQuery, DeserrJsonError>,
+    req: HttpRequest,
+    analytics: web::Data<dyn Analytics>,
 ) -> Result<HttpResponse, ResponseError> {
     debug!("called with body: {:?}", body);
 
-    documents_by_query(&index_scheduler, index_uid, body.into_inner())
+    let body = body.into_inner();
+
+    analytics.post_fetch_documents(
+        &DocumentFetchKind::Normal {
+            with_filter: body.filter.is_some(),
+            limit: body.limit,
+            offset: body.offset,
+        },
+        &req,
+    );
+
+    documents_by_query(&index_scheduler, index_uid, body)
 }
 
 pub async fn get_documents(
     index_scheduler: GuardedData<ActionPolicy<{ actions::DOCUMENTS_GET }>, Data<IndexScheduler>>,
     index_uid: web::Path<String>,
     params: AwebQueryParameter<BrowseQueryGet, DeserrQueryParamError>,
+    req: HttpRequest,
+    analytics: web::Data<dyn Analytics>,
 ) -> Result<HttpResponse, ResponseError> {
     debug!("called with params: {:?}", params);
 
@@ -191,6 +210,15 @@ pub async fn get_documents(
         filter,
     };
 
+    analytics.get_fetch_documents(
+        &DocumentFetchKind::Normal {
+            with_filter: query.filter.is_some(),
+            limit: query.limit,
+            offset: query.offset,
+        },
+        &req,
+    );
+
     documents_by_query(&index_scheduler, index_uid, query)
 }
 
@@ -458,7 +486,7 @@ pub async fn delete_documents_batch(
 #[derive(Debug, Deserr)]
 #[deserr(error = DeserrJsonError, rename_all = camelCase, deny_unknown_fields)]
 pub struct DocumentDeletionByFilter {
-    #[deserr(error = DeserrJsonError<InvalidDocumentDeleteFilter>)]
+    #[deserr(error = DeserrJsonError<InvalidDocumentFilter>, missing_field_error = DeserrJsonError::missing_document_filter)]
     filter: Value,
 }
 
@@ -480,8 +508,8 @@ pub async fn delete_documents_by_filter(
     || -> Result<_, ResponseError> {
         Ok(crate::search::parse_filter(&filter)?.ok_or(MeilisearchHttpError::EmptyFilter)?)
     }()
-    // and whatever was the error, the error code should always be an InvalidDocumentDeleteFilter
-    .map_err(|err| ResponseError::from_msg(err.message, Code::InvalidDocumentDeleteFilter))?;
+    // and whatever was the error, the error code should always be an InvalidDocumentFilter
+    .map_err(|err| ResponseError::from_msg(err.message, Code::InvalidDocumentFilter))?;
     let task = KindWithContent::DocumentDeletionByFilter { index_uid, filter_expr: filter };
 
     let task: SummarizedTaskView =
@@ -540,7 +568,12 @@ fn retrieve_documents<S: AsRef<str>>(
     };
 
     let candidates = if let Some(filter) = filter {
-        filter.evaluate(&rtxn, index)?
+        filter.evaluate(&rtxn, index).map_err(|err| match err {
+            milli::Error::UserError(milli::UserError::InvalidFilter(_)) => {
+                ResponseError::from_msg(err.to_string(), Code::InvalidDocumentFilter)
+            }
+            e => e.into(),
+        })?
     } else {
         index.documents_ids(&rtxn)?
     };

meilisearch/src/routes/indexes/search.rs

@@ -34,6 +34,8 @@ pub fn configure(cfg: &mut web::ServiceConfig) {
 pub struct SearchQueryGet {
     #[deserr(default, error = DeserrQueryParamError<InvalidSearchQ>)]
     q: Option<String>,
+    #[deserr(default, error = DeserrQueryParamError<InvalidSearchQ>)]
+    vector: Option<Vec<f32>>,
     #[deserr(default = Param(DEFAULT_SEARCH_OFFSET()), error = DeserrQueryParamError<InvalidSearchOffset>)]
     offset: Param<usize>,
     #[deserr(default = Param(DEFAULT_SEARCH_LIMIT()), error = DeserrQueryParamError<InvalidSearchLimit>)]
@@ -80,6 +82,7 @@ impl From<SearchQueryGet> for SearchQuery {
 
         Self {
             q: other.q,
+            vector: other.vector,
             offset: other.offset.0,
             limit: other.limit.0,
             page: other.page.as_deref().copied(),

meilisearch/src/routes/metrics.rs

@@ -17,7 +17,7 @@ pub fn configure(config: &mut web::ServiceConfig) {
 
 pub async fn get_metrics(
     index_scheduler: GuardedData<ActionPolicy<{ actions::METRICS_GET }>, Data<IndexScheduler>>,
-    auth_controller: GuardedData<ActionPolicy<{ actions::METRICS_GET }>, Data<AuthController>>,
+    auth_controller: Data<AuthController>,
 ) -> Result<HttpResponse, ResponseError> {
     let auth_filters = index_scheduler.filters();
     if !auth_filters.all_indexes_authorized() {
@@ -28,10 +28,10 @@ pub async fn get_metrics(
         return Err(error);
     }
 
-    let response =
-        create_all_stats((*index_scheduler).clone(), (*auth_controller).clone(), auth_filters)?;
+    let response = create_all_stats((*index_scheduler).clone(), auth_controller, auth_filters)?;
 
     crate::metrics::MEILISEARCH_DB_SIZE_BYTES.set(response.database_size as i64);
+    crate::metrics::MEILISEARCH_USED_DB_SIZE_BYTES.set(response.used_database_size as i64);
     crate::metrics::MEILISEARCH_INDEX_COUNT.set(response.indexes.len() as i64);
 
     for (index, value) in response.indexes.iter() {
@@ -40,6 +40,14 @@ pub async fn get_metrics(
             .set(value.number_of_documents as i64);
     }
 
+    for (kind, value) in index_scheduler.get_stats()? {
+        for (value, count) in value {
+            crate::metrics::MEILISEARCH_NB_TASKS
+                .with_label_values(&[&kind, &value])
+                .set(count as i64);
+        }
+    }
+
     let encoder = TextEncoder::new();
     let mut buffer = vec![];
     encoder.encode(&prometheus::gather(), &mut buffer).expect("Failed to encode metrics");

meilisearch/src/routes/mod.rs

@@ -231,6 +231,8 @@ pub async fn running() -> HttpResponse {
 #[serde(rename_all = "camelCase")]
 pub struct Stats {
     pub database_size: u64,
+    #[serde(skip)]
+    pub used_database_size: u64,
     #[serde(serialize_with = "time::serde::rfc3339::option::serialize")]
     pub last_update: Option<OffsetDateTime>,
     pub indexes: BTreeMap<String, indexes::IndexStats>,
@@ -259,6 +261,7 @@ pub fn create_all_stats(
     let mut last_task: Option<OffsetDateTime> = None;
     let mut indexes = BTreeMap::new();
     let mut database_size = 0;
+    let mut used_database_size = 0;
 
     for index_uid in index_scheduler.index_names()? {
         // Accumulate the size of all indexes, even unauthorized ones, so
@@ -266,6 +269,7 @@ pub fn create_all_stats(
         // See <https://github.com/meilisearch/meilisearch/pull/3541#discussion_r1126747643> for context.
         let stats = index_scheduler.index_stats(&index_uid)?;
         database_size += stats.inner_stats.database_size;
+        used_database_size += stats.inner_stats.used_database_size;
 
         if !filters.is_index_authorized(&index_uid) {
             continue;
@@ -278,10 +282,14 @@ pub fn create_all_stats(
     }
 
     database_size += index_scheduler.size()?;
+    used_database_size += index_scheduler.used_size()?;
     database_size += auth_controller.size()?;
-    database_size += index_scheduler.compute_update_file_size()?;
+    used_database_size += auth_controller.used_size()?;
+    let update_file_size = index_scheduler.compute_update_file_size()?;
+    database_size += update_file_size;
+    used_database_size += update_file_size;
 
-    let stats = Stats { database_size, last_update: last_task, indexes };
+    let stats = Stats { database_size, used_database_size, last_update: last_task, indexes };
     Ok(stats)
 }
 

meilisearch/src/routes/tasks.rs

@@ -99,7 +99,7 @@ pub struct DetailsView {
     #[serde(skip_serializing_if = "Option::is_none")]
     pub deleted_tasks: Option<Option<u64>>,
     #[serde(skip_serializing_if = "Option::is_none")]
-    pub original_filter: Option<String>,
+    pub original_filter: Option<Option<String>>,
     #[serde(skip_serializing_if = "Option::is_none")]
     pub dump_uid: Option<Option<String>>,
     #[serde(skip_serializing_if = "Option::is_none")]
@@ -131,12 +131,13 @@ impl From<Details> for DetailsView {
             } => DetailsView {
                 provided_ids: Some(received_document_ids),
                 deleted_documents: Some(deleted_documents),
+                original_filter: Some(None),
                 ..DetailsView::default()
             },
             Details::DocumentDeletionByFilter { original_filter, deleted_documents } => {
                 DetailsView {
                     provided_ids: Some(0),
-                    original_filter: Some(original_filter),
+                    original_filter: Some(Some(original_filter)),
                     deleted_documents: Some(deleted_documents),
                     ..DetailsView::default()
                 }
@@ -148,7 +149,7 @@ impl From<Details> for DetailsView {
                 DetailsView {
                     matched_tasks: Some(matched_tasks),
                     canceled_tasks: Some(canceled_tasks),
-                    original_filter: Some(original_filter),
+                    original_filter: Some(Some(original_filter)),
                     ..DetailsView::default()
                 }
             }
@@ -156,7 +157,7 @@ impl From<Details> for DetailsView {
                 DetailsView {
                     matched_tasks: Some(matched_tasks),
                     deleted_tasks: Some(deleted_tasks),
-                    original_filter: Some(original_filter),
+                    original_filter: Some(Some(original_filter)),
                     ..DetailsView::default()
                 }
             }
@@ -729,7 +730,7 @@ mod tests {
             let err = deserr_query_params::<TaskDeletionOrCancelationQuery>(params).unwrap_err();
             snapshot!(meili_snap::json_string!(err), @r###"
             {
-              "message": "Invalid value in parameter `types`: `createIndex` is not a valid task type. Available types are `documentAdditionOrUpdate`, `documentDeletion`, `documentDeletionByFilter`, `settingsUpdate`, `indexCreation`, `indexDeletion`, `indexUpdate`, `indexSwap`, `taskCancelation`, `taskDeletion`, `dumpCreation`, `snapshotCreation`.",
+              "message": "Invalid value in parameter `types`: `createIndex` is not a valid task type. Available types are `documentAdditionOrUpdate`, `documentDeletion`, `settingsUpdate`, `indexCreation`, `indexDeletion`, `indexUpdate`, `indexSwap`, `taskCancelation`, `taskDeletion`, `dumpCreation`, `snapshotCreation`.",
               "code": "invalid_task_types",
               "type": "invalid_request",
               "link": "https://docs.meilisearch.com/errors#invalid_task_types"

meilisearch/src/search.rs

@@ -9,13 +9,15 @@ use meilisearch_auth::IndexSearchRules;
 use meilisearch_types::deserr::DeserrJsonError;
 use meilisearch_types::error::deserr_codes::*;
 use meilisearch_types::index_uid::IndexUid;
+use meilisearch_types::milli::dot_product_similarity;
 use meilisearch_types::settings::DEFAULT_PAGINATION_MAX_TOTAL_HITS;
 use meilisearch_types::{milli, Document};
 use milli::tokenizer::TokenizerBuilder;
 use milli::{
     AscDesc, FieldId, FieldsIdsMap, Filter, FormatOptions, Index, MatchBounds, MatcherBuilder,
-    SortError, TermsMatchingStrategy, DEFAULT_VALUES_PER_FACET,
+    SortError, TermsMatchingStrategy, VectorOrArrayOfVectors, DEFAULT_VALUES_PER_FACET,
 };
+use ordered_float::OrderedFloat;
 use regex::Regex;
 use serde::Serialize;
 use serde_json::{json, Value};
@@ -31,11 +33,13 @@ pub const DEFAULT_CROP_MARKER: fn() -> String = || "…".to_string();
 pub const DEFAULT_HIGHLIGHT_PRE_TAG: fn() -> String = || "<em>".to_string();
 pub const DEFAULT_HIGHLIGHT_POST_TAG: fn() -> String = || "</em>".to_string();
 
-#[derive(Debug, Clone, Default, PartialEq, Eq, Deserr)]
+#[derive(Debug, Clone, Default, PartialEq, Deserr)]
 #[deserr(error = DeserrJsonError, rename_all = camelCase, deny_unknown_fields)]
 pub struct SearchQuery {
     #[deserr(default, error = DeserrJsonError<InvalidSearchQ>)]
     pub q: Option<String>,
+    #[deserr(default, error = DeserrJsonError<InvalidSearchQ>)]
+    pub vector: Option<Vec<f32>>,
     #[deserr(default = DEFAULT_SEARCH_OFFSET(), error = DeserrJsonError<InvalidSearchOffset>)]
     pub offset: usize,
     #[deserr(default = DEFAULT_SEARCH_LIMIT(), error = DeserrJsonError<InvalidSearchLimit>)]
@@ -80,13 +84,15 @@ impl SearchQuery {
 // This struct contains the fields of `SearchQuery` inline.
 // This is because neither deserr nor serde support `flatten` when using `deny_unknown_fields.
 // The `From<SearchQueryWithIndex>` implementation ensures both structs remain up to date.
-#[derive(Debug, Clone, PartialEq, Eq, Deserr)]
+#[derive(Debug, Clone, PartialEq, Deserr)]
 #[deserr(error = DeserrJsonError, rename_all = camelCase, deny_unknown_fields)]
 pub struct SearchQueryWithIndex {
     #[deserr(error = DeserrJsonError<InvalidIndexUid>, missing_field_error = DeserrJsonError::missing_index_uid)]
     pub index_uid: IndexUid,
     #[deserr(default, error = DeserrJsonError<InvalidSearchQ>)]
     pub q: Option<String>,
+    #[deserr(default, error = DeserrJsonError<InvalidSearchQ>)]
+    pub vector: Option<Vec<f32>>,
     #[deserr(default = DEFAULT_SEARCH_OFFSET(), error = DeserrJsonError<InvalidSearchOffset>)]
     pub offset: usize,
     #[deserr(default = DEFAULT_SEARCH_LIMIT(), error = DeserrJsonError<InvalidSearchLimit>)]
@@ -126,6 +132,7 @@ impl SearchQueryWithIndex {
         let SearchQueryWithIndex {
             index_uid,
             q,
+            vector,
             offset,
             limit,
             page,
@@ -147,6 +154,7 @@ impl SearchQueryWithIndex {
             index_uid,
             SearchQuery {
                 q,
+                vector,
                 offset,
                 limit,
                 page,
@@ -209,6 +217,8 @@ pub struct SearchHit {
 pub struct SearchResult {
     pub hits: Vec<SearchHit>,
     pub query: String,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub vector: Option<Vec<f32>>,
     pub processing_time_ms: u128,
     #[serde(flatten)]
     pub hits_info: HitsInfo,
@@ -270,6 +280,10 @@ pub fn perform_search(
 
     let mut search = index.search(&rtxn);
 
+    if let Some(ref vector) = query.vector {
+        search.vector(vector.clone());
+    }
+
     if let Some(ref query) = query.q {
         search.query(query);
     }
@@ -389,7 +403,6 @@ pub fn perform_search(
     formatter_builder.highlight_suffix(query.highlight_post_tag);
 
     let mut documents = Vec::new();
-
     let documents_iter = index.documents(&rtxn, documents_ids)?;
 
     for (_id, obkv) in documents_iter {
@@ -416,6 +429,12 @@ pub fn perform_search(
             insert_geo_distance(sort, &mut document);
         }
 
+        if let Some(vector) = query.vector.as_ref() {
+            if let Some(vectors) = extract_field("_vectors", &fields_ids_map, obkv)? {
+                insert_semantic_similarity(vector, vectors, &mut document);
+            }
+        }
+
         let hit = SearchHit { document, formatted, matches_position };
         documents.push(hit);
     }
@@ -465,7 +484,8 @@ pub fn perform_search(
     let result = SearchResult {
         hits: documents,
         hits_info,
-        query: query.q.clone().unwrap_or_default(),
+        query: query.q.unwrap_or_default(),
+        vector: query.vector,
         processing_time_ms: before_search.elapsed().as_millis(),
         facet_distribution,
         facet_stats,
@@ -489,6 +509,20 @@ fn insert_geo_distance(sorts: &[String], document: &mut Document) {
     }
 }
 
+fn insert_semantic_similarity(query: &[f32], vectors: Value, document: &mut Document) {
+    let vectors =
+        match serde_json::from_value(vectors).map(VectorOrArrayOfVectors::into_array_of_vectors) {
+            Ok(vectors) => vectors,
+            Err(_) => return,
+        };
+    let similarity = vectors
+        .into_iter()
+        .map(|v| OrderedFloat(dot_product_similarity(query, &v)))
+        .max()
+        .map(OrderedFloat::into_inner);
+    document.insert("_semanticSimilarity".to_string(), json!(similarity));
+}
+
 fn compute_formatted_options(
     attr_to_highlight: &HashSet<String>,
     attr_to_crop: &[String],
@@ -616,6 +650,22 @@ fn make_document(
     Ok(document)
 }
 
+/// Extract the JSON value under the field name specified
+/// but doesn't support nested objects.
+fn extract_field(
+    field_name: &str,
+    field_ids_map: &FieldsIdsMap,
+    obkv: obkv::KvReaderU16,
+) -> Result<Option<serde_json::Value>, MeilisearchHttpError> {
+    match field_ids_map.id(field_name) {
+        Some(fid) => match obkv.get(fid) {
+            Some(value) => Ok(serde_json::from_slice(value).map(Some)?),
+            None => Ok(None),
+        },
+        None => Ok(None),
+    }
+}
+
 fn format_fields<A: AsRef<[u8]>>(
     document: &Document,
     field_ids_map: &FieldsIdsMap,

meilisearch/tests/auth/authorization.rs

@@ -16,8 +16,11 @@ pub static AUTHORIZATIONS: Lazy<HashMap<(&'static str, &'static str), HashSet<&'
             ("GET",     "/indexes/products/search") =>                         hashset!{"search", "*"},
             ("POST",    "/indexes/products/documents") =>                      hashset!{"documents.add", "documents.*", "*"},
             ("GET",     "/indexes/products/documents") =>                      hashset!{"documents.get", "documents.*", "*"},
+            ("POST",    "/indexes/products/documents/fetch") =>                hashset!{"documents.get", "documents.*", "*"},
             ("GET",     "/indexes/products/documents/0") =>                    hashset!{"documents.get", "documents.*", "*"},
             ("DELETE",  "/indexes/products/documents/0") =>                    hashset!{"documents.delete", "documents.*", "*"},
+            ("POST",    "/indexes/products/documents/delete-batch") =>         hashset!{"documents.delete", "documents.*", "*"},
+            ("POST",    "/indexes/products/documents/delete") =>               hashset!{"documents.delete", "documents.*", "*"},
             ("GET",     "/tasks") =>                                           hashset!{"tasks.get", "tasks.*", "*"},
             ("DELETE",  "/tasks") =>                                           hashset!{"tasks.delete", "tasks.*", "*"},
             ("GET",     "/tasks?indexUid=products") =>                         hashset!{"tasks.get", "tasks.*", "*"},

meilisearch/tests/documents/add_documents.rs

@@ -1781,7 +1781,7 @@ async fn error_add_documents_payload_size() {
     snapshot!(json_string!(response, { ".duration" => "[duration]", ".enqueuedAt" => "[date]", ".startedAt" => "[date]", ".finishedAt" => "[date]" }),
         @r###"
     {
-      "message": "The provided payload reached the size limit.",
+      "message": "The provided payload reached the size limit. The maximum accepted payload size is 10.00 MiB.",
       "code": "payload_too_large",
       "type": "invalid_request",
       "link": "https://docs.meilisearch.com/errors#payload_too_large"

meilisearch/tests/documents/errors.rs

@@ -180,9 +180,9 @@ async fn get_all_documents_bad_filter() {
     snapshot!(json_string!(response), @r###"
     {
       "message": "Attribute `doggo` is not filterable. This index does not have configured filterable attributes.\n1:6 doggo=bernese",
-      "code": "invalid_search_filter",
+      "code": "invalid_document_filter",
       "type": "invalid_request",
-      "link": "https://docs.meilisearch.com/errors#invalid_search_filter"
+      "link": "https://docs.meilisearch.com/errors#invalid_document_filter"
     }
     "###);
 }
@@ -547,9 +547,9 @@ async fn delete_document_by_filter() {
     snapshot!(json_string!(response), @r###"
     {
       "message": "Invalid syntax for the filter parameter: `expected String, Array, found: true`.",
-      "code": "invalid_document_delete_filter",
+      "code": "invalid_document_filter",
       "type": "invalid_request",
-      "link": "https://docs.meilisearch.com/errors#invalid_document_delete_filter"
+      "link": "https://docs.meilisearch.com/errors#invalid_document_filter"
     }
     "###);
 
@@ -559,9 +559,9 @@ async fn delete_document_by_filter() {
     snapshot!(json_string!(response), @r###"
     {
       "message": "Was expecting an operation `=`, `!=`, `>=`, `>`, `<=`, `<`, `IN`, `NOT IN`, `TO`, `EXISTS`, `NOT EXISTS`, `IS NULL`, `IS NOT NULL`, `IS EMPTY`, `IS NOT EMPTY`, `_geoRadius`, or `_geoBoundingBox` at `hello`.\n1:6 hello",
-      "code": "invalid_document_delete_filter",
+      "code": "invalid_document_filter",
       "type": "invalid_request",
-      "link": "https://docs.meilisearch.com/errors#invalid_document_delete_filter"
+      "link": "https://docs.meilisearch.com/errors#invalid_document_filter"
     }
     "###);
 
@@ -571,9 +571,21 @@ async fn delete_document_by_filter() {
     snapshot!(json_string!(response), @r###"
     {
       "message": "Sending an empty filter is forbidden.",
-      "code": "invalid_document_delete_filter",
+      "code": "invalid_document_filter",
       "type": "invalid_request",
-      "link": "https://docs.meilisearch.com/errors#invalid_document_delete_filter"
+      "link": "https://docs.meilisearch.com/errors#invalid_document_filter"
+    }
+    "###);
+
+    // do not send any filter
+    let (response, code) = index.delete_document_by_filter(json!({})).await;
+    snapshot!(code, @"400 Bad Request");
+    snapshot!(json_string!(response), @r###"
+    {
+      "message": "Missing field `filter`",
+      "code": "missing_document_filter",
+      "type": "invalid_request",
+      "link": "https://docs.meilisearch.com/errors#missing_document_filter"
     }
     "###);
 
@@ -630,9 +642,9 @@ async fn delete_document_by_filter() {
       },
       "error": {
         "message": "Attribute `doggo` is not filterable. This index does not have configured filterable attributes.\n1:6 doggo = bernese",
-        "code": "invalid_search_filter",
+        "code": "invalid_document_filter",
         "type": "invalid_request",
-        "link": "https://docs.meilisearch.com/errors#invalid_search_filter"
+        "link": "https://docs.meilisearch.com/errors#invalid_document_filter"
       },
       "duration": "[duration]",
       "enqueuedAt": "[date]",
@@ -664,9 +676,9 @@ async fn delete_document_by_filter() {
       },
       "error": {
         "message": "Attribute `catto` is not filterable. Available filterable attributes are: `doggo`.\n1:6 catto = jorts",
-        "code": "invalid_search_filter",
+        "code": "invalid_document_filter",
         "type": "invalid_request",
-        "link": "https://docs.meilisearch.com/errors#invalid_search_filter"
+        "link": "https://docs.meilisearch.com/errors#invalid_document_filter"
       },
       "duration": "[duration]",
       "enqueuedAt": "[date]",
@@ -748,4 +760,27 @@ async fn fetch_document_by_filter() {
       "link": "https://docs.meilisearch.com/errors#invalid_document_filter"
     }
     "###);
+
+    let (response, code) = index.get_document_by_filter(json!({ "filter": "cool doggo" })).await;
+    snapshot!(code, @"400 Bad Request");
+    snapshot!(json_string!(response), @r###"
+    {
+      "message": "Was expecting an operation `=`, `!=`, `>=`, `>`, `<=`, `<`, `IN`, `NOT IN`, `TO`, `EXISTS`, `NOT EXISTS`, `IS NULL`, `IS NOT NULL`, `IS EMPTY`, `IS NOT EMPTY`, `_geoRadius`, or `_geoBoundingBox` at `cool doggo`.\n1:11 cool doggo",
+      "code": "invalid_document_filter",
+      "type": "invalid_request",
+      "link": "https://docs.meilisearch.com/errors#invalid_document_filter"
+    }
+    "###);
+
+    let (response, code) =
+        index.get_document_by_filter(json!({ "filter": "doggo = bernese" })).await;
+    snapshot!(code, @"400 Bad Request");
+    snapshot!(json_string!(response), @r###"
+    {
+      "message": "Attribute `doggo` is not filterable. Available filterable attributes are: `color`.\n1:6 doggo = bernese",
+      "code": "invalid_document_filter",
+      "type": "invalid_request",
+      "link": "https://docs.meilisearch.com/errors#invalid_document_filter"
+    }
+    "###);
 }

meilisearch/tests/search/errors.rs

@@ -946,7 +946,7 @@ async fn sort_unset_ranking_rule() {
     index.wait_task(1).await;
 
     let expected_response = json!({
-        "message": "The sort ranking rule must be specified in the ranking rules settings to use the sort parameter at search time.",
+        "message": "You must specify where `sort` is listed in the rankingRules setting to use the sort parameter at search time.",
         "code": "invalid_search_sort",
         "type": "invalid_request",
         "link": "https://docs.meilisearch.com/errors#invalid_search_sort"

meilisearch/tests/tasks/errors.rs

@@ -97,7 +97,7 @@ async fn task_bad_types() {
     snapshot!(code, @"400 Bad Request");
     snapshot!(json_string!(response), @r###"
     {
-      "message": "Invalid value in parameter `types`: `doggo` is not a valid task type. Available types are `documentAdditionOrUpdate`, `documentDeletion`, `documentDeletionByFilter`, `settingsUpdate`, `indexCreation`, `indexDeletion`, `indexUpdate`, `indexSwap`, `taskCancelation`, `taskDeletion`, `dumpCreation`, `snapshotCreation`.",
+      "message": "Invalid value in parameter `types`: `doggo` is not a valid task type. Available types are `documentAdditionOrUpdate`, `documentDeletion`, `settingsUpdate`, `indexCreation`, `indexDeletion`, `indexUpdate`, `indexSwap`, `taskCancelation`, `taskDeletion`, `dumpCreation`, `snapshotCreation`.",
       "code": "invalid_task_types",
       "type": "invalid_request",
       "link": "https://docs.meilisearch.com/errors#invalid_task_types"
@@ -108,7 +108,7 @@ async fn task_bad_types() {
     snapshot!(code, @"400 Bad Request");
     snapshot!(json_string!(response), @r###"
     {
-      "message": "Invalid value in parameter `types`: `doggo` is not a valid task type. Available types are `documentAdditionOrUpdate`, `documentDeletion`, `documentDeletionByFilter`, `settingsUpdate`, `indexCreation`, `indexDeletion`, `indexUpdate`, `indexSwap`, `taskCancelation`, `taskDeletion`, `dumpCreation`, `snapshotCreation`.",
+      "message": "Invalid value in parameter `types`: `doggo` is not a valid task type. Available types are `documentAdditionOrUpdate`, `documentDeletion`, `settingsUpdate`, `indexCreation`, `indexDeletion`, `indexUpdate`, `indexSwap`, `taskCancelation`, `taskDeletion`, `dumpCreation`, `snapshotCreation`.",
       "code": "invalid_task_types",
       "type": "invalid_request",
       "link": "https://docs.meilisearch.com/errors#invalid_task_types"
@@ -119,7 +119,7 @@ async fn task_bad_types() {
     snapshot!(code, @"400 Bad Request");
     snapshot!(json_string!(response), @r###"
     {
-      "message": "Invalid value in parameter `types`: `doggo` is not a valid task type. Available types are `documentAdditionOrUpdate`, `documentDeletion`, `documentDeletionByFilter`, `settingsUpdate`, `indexCreation`, `indexDeletion`, `indexUpdate`, `indexSwap`, `taskCancelation`, `taskDeletion`, `dumpCreation`, `snapshotCreation`.",
+      "message": "Invalid value in parameter `types`: `doggo` is not a valid task type. Available types are `documentAdditionOrUpdate`, `documentDeletion`, `settingsUpdate`, `indexCreation`, `indexDeletion`, `indexUpdate`, `indexSwap`, `taskCancelation`, `taskDeletion`, `dumpCreation`, `snapshotCreation`.",
       "code": "invalid_task_types",
       "type": "invalid_request",
       "link": "https://docs.meilisearch.com/errors#invalid_task_types"

meilisearch/tests/tasks/mod.rs

@@ -413,7 +413,7 @@ async fn test_summarized_document_addition_or_update() {
 }
 
 #[actix_web::test]
-async fn test_summarized_delete_batch() {
+async fn test_summarized_delete_documents_by_batch() {
     let server = Server::new().await;
     let index = server.index("test");
     index.delete_batch(vec![1, 2, 3]).await;
@@ -430,7 +430,8 @@ async fn test_summarized_delete_batch() {
       "canceledBy": null,
       "details": {
         "providedIds": 3,
-        "deletedDocuments": 0
+        "deletedDocuments": 0,
+        "originalFilter": null
       },
       "error": {
         "message": "Index `test` not found.",
@@ -460,7 +461,8 @@ async fn test_summarized_delete_batch() {
       "canceledBy": null,
       "details": {
         "providedIds": 1,
-        "deletedDocuments": 0
+        "deletedDocuments": 0,
+        "originalFilter": null
       },
       "error": null,
       "duration": "[duration]",
@@ -472,7 +474,100 @@ async fn test_summarized_delete_batch() {
 }
 
 #[actix_web::test]
-async fn test_summarized_delete_document() {
+async fn test_summarized_delete_documents_by_filter() {
+    let server = Server::new().await;
+    let index = server.index("test");
+
+    index.delete_document_by_filter(json!({ "filter": "doggo = bernese" })).await;
+    index.wait_task(0).await;
+    let (task, _) = index.get_task(0).await;
+    assert_json_snapshot!(task,
+        { ".duration" => "[duration]", ".enqueuedAt" => "[date]", ".startedAt" => "[date]", ".finishedAt" => "[date]" },
+        @r###"
+    {
+      "uid": 0,
+      "indexUid": "test",
+      "status": "failed",
+      "type": "documentDeletion",
+      "canceledBy": null,
+      "details": {
+        "providedIds": 0,
+        "deletedDocuments": 0,
+        "originalFilter": "\"doggo = bernese\""
+      },
+      "error": {
+        "message": "Index `test` not found.",
+        "code": "index_not_found",
+        "type": "invalid_request",
+        "link": "https://docs.meilisearch.com/errors#index_not_found"
+      },
+      "duration": "[duration]",
+      "enqueuedAt": "[date]",
+      "startedAt": "[date]",
+      "finishedAt": "[date]"
+    }
+    "###);
+
+    index.create(None).await;
+    index.delete_document_by_filter(json!({ "filter": "doggo = bernese" })).await;
+    index.wait_task(2).await;
+    let (task, _) = index.get_task(2).await;
+    assert_json_snapshot!(task,
+        { ".duration" => "[duration]", ".enqueuedAt" => "[date]", ".startedAt" => "[date]", ".finishedAt" => "[date]" },
+        @r###"
+    {
+      "uid": 2,
+      "indexUid": "test",
+      "status": "failed",
+      "type": "documentDeletion",
+      "canceledBy": null,
+      "details": {
+        "providedIds": 0,
+        "deletedDocuments": 0,
+        "originalFilter": "\"doggo = bernese\""
+      },
+      "error": {
+        "message": "Attribute `doggo` is not filterable. This index does not have configured filterable attributes.\n1:6 doggo = bernese",
+        "code": "invalid_document_filter",
+        "type": "invalid_request",
+        "link": "https://docs.meilisearch.com/errors#invalid_document_filter"
+      },
+      "duration": "[duration]",
+      "enqueuedAt": "[date]",
+      "startedAt": "[date]",
+      "finishedAt": "[date]"
+    }
+    "###);
+
+    index.update_settings(json!({ "filterableAttributes": ["doggo"] })).await;
+    index.delete_document_by_filter(json!({ "filter": "doggo = bernese" })).await;
+    index.wait_task(4).await;
+    let (task, _) = index.get_task(4).await;
+    assert_json_snapshot!(task,
+        { ".duration" => "[duration]", ".enqueuedAt" => "[date]", ".startedAt" => "[date]", ".finishedAt" => "[date]" },
+        @r###"
+    {
+      "uid": 4,
+      "indexUid": "test",
+      "status": "succeeded",
+      "type": "documentDeletion",
+      "canceledBy": null,
+      "details": {
+        "providedIds": 0,
+        "deletedDocuments": 0,
+        "originalFilter": "\"doggo = bernese\""
+      },
+      "error": null,
+      "duration": "[duration]",
+      "enqueuedAt": "[date]",
+      "startedAt": "[date]",
+      "finishedAt": "[date]"
+    }
+    "###);
+}
+
+#[actix_web::test]
+async fn test_summarized_delete_document_by_id() {
     let server = Server::new().await;
     let index = server.index("test");
     index.delete_document(1).await;
@@ -489,7 +584,8 @@ async fn test_summarized_delete_document() {
       "canceledBy": null,
       "details": {
         "providedIds": 1,
-        "deletedDocuments": 0
+        "deletedDocuments": 0,
+        "originalFilter": null
       },
       "error": {
         "message": "Index `test` not found.",
@@ -519,7 +615,8 @@ async fn test_summarized_delete_document() {
       "canceledBy": null,
       "details": {
         "providedIds": 1,
-        "deletedDocuments": 0
+        "deletedDocuments": 0,
+        "originalFilter": null
       },
       "error": null,
       "duration": "[duration]",

milli/Cargo.toml

@@ -15,6 +15,7 @@ license.workspace = true
 bimap = { version = "0.6.3", features = ["serde"] }
 bincode = "1.3.3"
 bstr = "1.4.0"
+bytemuck = { version = "1.13.1", features = ["extern_crate_alloc"] }
 byteorder = "1.4.3"
 charabia = { version = "0.7.2", default-features = false }
 concat-arrays = "0.1.2"
@@ -25,26 +26,39 @@ flatten-serde-json = { path = "../flatten-serde-json" }
 fst = "0.4.7"
 fxhash = "0.2.1"
 geoutils = "0.5.1"
-grenad = { version = "0.4.4", default-features = false, features = ["tempfile"] }
-heed = { git = "https://github.com/meilisearch/heed", tag = "v0.12.5", default-features = false, features = ["lmdb", "sync-read-txn"] }
+grenad = { version = "0.4.4", default-features = false, features = [
+    "tempfile",
+] }
+heed = { git = "https://github.com/meilisearch/heed", tag = "v0.12.6", default-features = false, features = [
+    "lmdb",
+    "sync-read-txn",
+] }
+hnsw = { version = "0.11.0", features = ["serde1"] }
 json-depth-checker = { path = "../json-depth-checker" }
 levenshtein_automata = { version = "0.2.1", features = ["fst_automaton"] }
 memmap2 = "0.5.10"
 obkv = "0.2.0"
 once_cell = "1.17.1"
 ordered-float = "3.6.0"
+rand_pcg = { version = "0.3.1", features = ["serde1"] }
 rayon = "1.7.0"
 roaring = "0.10.1"
 rstar = { version = "0.10.0", features = ["serde"] }
 serde = { version = "1.0.160", features = ["derive"] }
 serde_json = { version = "1.0.95", features = ["preserve_order"] }
 slice-group-by = "0.3.0"
-smallstr =  { version = "0.3.0", features = ["serde"] }
+space = "0.17.0"
+smallstr = { version = "0.3.0", features = ["serde"] }
 smallvec = "1.10.0"
 smartstring = "1.0.1"
 tempfile = "3.5.0"
 thiserror = "1.0.40"
-time = { version = "0.3.20", features = ["serde-well-known", "formatting", "parsing", "macros"] }
+time = { version = "0.3.20", features = [
+    "serde-well-known",
+    "formatting",
+    "parsing",
+    "macros",
+] }
 uuid = { version = "1.3.1", features = ["v4"] }
 
 filter-parser = { path = "../filter-parser" }
@@ -63,13 +77,13 @@ big_s = "1.0.2"
 insta = "1.29.0"
 maplit = "1.0.2"
 md5 = "0.7.0"
-rand = {version = "0.8.5", features = ["small_rng"] }
+rand = { version = "0.8.5", features = ["small_rng"] }
 
 [target.'cfg(fuzzing)'.dev-dependencies]
 fuzzcheck = "0.12.1"
 
 [features]
-all-tokenizations = [ "charabia/default" ]
+all-tokenizations = ["charabia/default"]
 
 # Use POSIX semaphores instead of SysV semaphores in LMDB
 # For more information on this feature, see heed's Cargo.toml

milli/examples/search.rs

@@ -52,6 +52,7 @@ fn main() -> Result<(), Box<dyn Error>> {
             let docs = execute_search(
                 &mut ctx,
                 &(!query.trim().is_empty()).then(|| query.trim().to_owned()),
+                &None,
                 TermsMatchingStrategy::Last,
                 false,
                 &None,

milli/src/distance.rs

@@ -0,0 +1,44 @@
+use serde::{Deserialize, Serialize};
+use space::Metric;
+
+#[derive(Debug, Default, Clone, Copy, Serialize, Deserialize)]
+pub struct DotProduct;
+
+impl Metric<Vec<f32>> for DotProduct {
+    type Unit = u32;
+
+    // TODO explain me this function, I don't understand why f32.to_bits is ordered.
+    // I tried to do this and it wasn't OK <https://stackoverflow.com/a/43305015/1941280>
+    //
+    // Following <https://docs.rs/space/0.17.0/space/trait.Metric.html>.
+    fn distance(&self, a: &Vec<f32>, b: &Vec<f32>) -> Self::Unit {
+        let dist = 1.0 - dot_product_similarity(a, b);
+        debug_assert!(!dist.is_nan());
+        dist.to_bits()
+    }
+}
+
+/// Returns the dot product similarity score that will between 0.0 and 1.0
+/// if both vectors are normalized. The higher the more similar the vectors are.
+pub fn dot_product_similarity(a: &[f32], b: &[f32]) -> f32 {
+    a.iter().zip(b).map(|(a, b)| a * b).sum()
+}
+
+#[derive(Debug, Default, Clone, Copy, Serialize, Deserialize)]
+pub struct Euclidean;
+
+impl Metric<Vec<f32>> for Euclidean {
+    type Unit = u32;
+
+    fn distance(&self, a: &Vec<f32>, b: &Vec<f32>) -> Self::Unit {
+        let dist = euclidean_squared_distance(a, b).sqrt();
+        debug_assert!(!dist.is_nan());
+        dist.to_bits()
+    }
+}
+
+/// Return the squared euclidean distance between both vectors that will
+/// between 0.0 and +inf. The smaller the nearer the vectors are.
+pub fn euclidean_squared_distance(a: &[f32], b: &[f32]) -> f32 {
+    a.iter().zip(b).map(|(a, b)| (a - b).powi(2)).sum()
+}

milli/src/error.rs

@@ -110,9 +110,13 @@ only composed of alphanumeric characters (a-z A-Z 0-9), hyphens (-) and undersco
     },
     #[error(transparent)]
     InvalidGeoField(#[from] GeoError),
+    #[error("Invalid vector dimensions: expected: `{}`, found: `{}`.", .expected, .found)]
+    InvalidVectorDimensions { expected: usize, found: usize },
+    #[error("The `_vectors` field in the document with the id: `{document_id}` is not an array. Was expecting an array of floats or an array of arrays of floats but instead got `{value}`.")]
+    InvalidVectorsType { document_id: Value, value: Value },
     #[error("{0}")]
     InvalidFilter(String),
-    #[error("Invalid type for filter subexpression: `expected {}, found: {1}`.", .0.join(", "))]
+    #[error("Invalid type for filter subexpression: expected: {}, found: {1}.", .0.join(", "))]
     InvalidFilterExpression(&'static [&'static str], Value),
     #[error("Attribute `{}` is not sortable. {}",
         .field,
@@ -126,7 +130,7 @@ only composed of alphanumeric characters (a-z A-Z 0-9), hyphens (-) and undersco
     InvalidSortableAttribute { field: String, valid_fields: BTreeSet<String> },
     #[error("{}", HeedError::BadOpenOptions)]
     InvalidLmdbOpenOptions,
-    #[error("The sort ranking rule must be specified in the ranking rules settings to use the sort parameter at search time.")]
+    #[error("You must specify where `sort` is listed in the rankingRules setting to use the sort parameter at search time.")]
     SortRankingRuleMissing,
     #[error("The database file is in an invalid state.")]
     InvalidStoreFile,

milli/src/heed_codec/roaring_bitmap/cbo_roaring_bitmap_codec.rs

@@ -49,7 +49,7 @@ impl CboRoaringBitmapCodec {
         } else {
             // Otherwise, it means we used the classic RoaringBitmapCodec and
             // that the header takes threshold integers.
-            RoaringBitmap::deserialize_from(bytes)
+            RoaringBitmap::deserialize_unchecked_from(bytes)
         }
     }
 
@@ -69,7 +69,7 @@ impl CboRoaringBitmapCodec {
                     vec.push(integer);
                 }
             } else {
-                roaring |= RoaringBitmap::deserialize_from(bytes.as_ref())?;
+                roaring |= RoaringBitmap::deserialize_unchecked_from(bytes.as_ref())?;
             }
         }
 

milli/src/heed_codec/roaring_bitmap/roaring_bitmap_codec.rs

@@ -8,7 +8,7 @@ impl heed::BytesDecode<'_> for RoaringBitmapCodec {
     type DItem = RoaringBitmap;
 
     fn bytes_decode(bytes: &[u8]) -> Option<Self::DItem> {
-        RoaringBitmap::deserialize_from(bytes).ok()
+        RoaringBitmap::deserialize_unchecked_from(bytes).ok()
     }
 }
 

milli/src/index.rs

@@ -8,10 +8,12 @@ use charabia::{Language, Script};
 use heed::flags::Flags;
 use heed::types::*;
 use heed::{CompactionOption, Database, PolyDatabase, RoTxn, RwTxn};
+use rand_pcg::Pcg32;
 use roaring::RoaringBitmap;
 use rstar::RTree;
 use time::OffsetDateTime;
 
+use crate::distance::DotProduct;
 use crate::error::{InternalError, UserError};
 use crate::facet::FacetType;
 use crate::fields_ids_map::FieldsIdsMap;
@@ -21,12 +23,14 @@ use crate::heed_codec::facet::{
 };
 use crate::heed_codec::{ScriptLanguageCodec, StrBEU16Codec, StrRefCodec};
 use crate::{
-    default_criteria, BEU32StrCodec, BoRoaringBitmapCodec, CboRoaringBitmapCodec, Criterion,
-    DocumentId, ExternalDocumentsIds, FacetDistribution, FieldDistribution, FieldId,
-    FieldIdWordCountCodec, GeoPoint, ObkvCodec, Result, RoaringBitmapCodec, RoaringBitmapLenCodec,
-    Search, U8StrStrCodec, BEU16, BEU32,
+    default_criteria, CboRoaringBitmapCodec, Criterion, DocumentId, ExternalDocumentsIds,
+    FacetDistribution, FieldDistribution, FieldId, FieldIdWordCountCodec, GeoPoint, ObkvCodec,
+    Result, RoaringBitmapCodec, RoaringBitmapLenCodec, Search, U8StrStrCodec, BEU16, BEU32,
 };
 
+/// The HNSW data-structure that we serialize, fill and search in.
+pub type Hnsw = hnsw::Hnsw<DotProduct, Vec<f32>, Pcg32, 12, 24>;
+
 pub const DEFAULT_MIN_WORD_LEN_ONE_TYPO: u8 = 5;
 pub const DEFAULT_MIN_WORD_LEN_TWO_TYPOS: u8 = 9;
 
@@ -43,6 +47,7 @@ pub mod main_key {
     pub const FIELDS_IDS_MAP_KEY: &str = "fields-ids-map";
     pub const GEO_FACETED_DOCUMENTS_IDS_KEY: &str = "geo-faceted-documents-ids";
     pub const GEO_RTREE_KEY: &str = "geo-rtree";
+    pub const VECTOR_HNSW_KEY: &str = "vector-hnsw";
     pub const HARD_EXTERNAL_DOCUMENTS_IDS_KEY: &str = "hard-external-documents-ids";
     pub const NUMBER_FACETED_DOCUMENTS_IDS_PREFIX: &str = "number-faceted-documents-ids";
     pub const PRIMARY_KEY_KEY: &str = "primary-key";
@@ -87,6 +92,7 @@ pub mod db_name {
     pub const FACET_ID_STRING_DOCIDS: &str = "facet-id-string-docids";
     pub const FIELD_ID_DOCID_FACET_F64S: &str = "field-id-docid-facet-f64s";
     pub const FIELD_ID_DOCID_FACET_STRINGS: &str = "field-id-docid-facet-strings";
+    pub const VECTOR_ID_DOCID: &str = "vector-id-docids";
     pub const DOCUMENTS: &str = "documents";
     pub const SCRIPT_LANGUAGE_DOCIDS: &str = "script_language_docids";
 }
@@ -111,9 +117,6 @@ pub struct Index {
     /// A prefix of word and all the documents ids containing this prefix, from attributes for which typos are not allowed.
     pub exact_word_prefix_docids: Database<Str, RoaringBitmapCodec>,
 
-    /// Maps a word and a document id (u32) to all the positions where the given word appears.
-    pub docid_word_positions: Database<BEU32StrCodec, BoRoaringBitmapCodec>,
-
     /// Maps the proximity between a pair of words with all the docids where this relation appears.
     pub word_pair_proximity_docids: Database<U8StrStrCodec, CboRoaringBitmapCodec>,
     /// Maps the proximity between a pair of word and prefix with all the docids where this relation appears.
@@ -153,6 +156,9 @@ pub struct Index {
     /// Maps the document id, the facet field id and the strings.
     pub field_id_docid_facet_strings: Database<FieldDocIdFacetStringCodec, Str>,
 
+    /// Maps a vector id to the document id that have it.
+    pub vector_id_docid: Database<OwnedType<BEU32>, OwnedType<BEU32>>,
+
     /// Maps the document id to the document as an obkv store.
     pub(crate) documents: Database<OwnedType<BEU32>, ObkvCodec>,
 }
@@ -166,37 +172,49 @@ impl Index {
     ) -> Result<Index> {
         use db_name::*;
 
-        options.max_dbs(23);
+        options.max_dbs(24);
         unsafe { options.flag(Flags::MdbAlwaysFreePages) };
 
         let env = options.open(path)?;
-        let main = env.create_poly_database(Some(MAIN))?;
-        let word_docids = env.create_database(Some(WORD_DOCIDS))?;
-        let exact_word_docids = env.create_database(Some(EXACT_WORD_DOCIDS))?;
-        let word_prefix_docids = env.create_database(Some(WORD_PREFIX_DOCIDS))?;
-        let exact_word_prefix_docids = env.create_database(Some(EXACT_WORD_PREFIX_DOCIDS))?;
-        let docid_word_positions = env.create_database(Some(DOCID_WORD_POSITIONS))?;
-        let word_pair_proximity_docids = env.create_database(Some(WORD_PAIR_PROXIMITY_DOCIDS))?;
-        let script_language_docids = env.create_database(Some(SCRIPT_LANGUAGE_DOCIDS))?;
+        let mut wtxn = env.write_txn()?;
+        let main = env.create_poly_database(&mut wtxn, Some(MAIN))?;
+        let word_docids = env.create_database(&mut wtxn, Some(WORD_DOCIDS))?;
+        let exact_word_docids = env.create_database(&mut wtxn, Some(EXACT_WORD_DOCIDS))?;
+        let word_prefix_docids = env.create_database(&mut wtxn, Some(WORD_PREFIX_DOCIDS))?;
+        let exact_word_prefix_docids =
+            env.create_database(&mut wtxn, Some(EXACT_WORD_PREFIX_DOCIDS))?;
+        let word_pair_proximity_docids =
+            env.create_database(&mut wtxn, Some(WORD_PAIR_PROXIMITY_DOCIDS))?;
+        let script_language_docids =
+            env.create_database(&mut wtxn, Some(SCRIPT_LANGUAGE_DOCIDS))?;
         let word_prefix_pair_proximity_docids =
-            env.create_database(Some(WORD_PREFIX_PAIR_PROXIMITY_DOCIDS))?;
+            env.create_database(&mut wtxn, Some(WORD_PREFIX_PAIR_PROXIMITY_DOCIDS))?;
         let prefix_word_pair_proximity_docids =
-            env.create_database(Some(PREFIX_WORD_PAIR_PROXIMITY_DOCIDS))?;
-        let word_position_docids = env.create_database(Some(WORD_POSITION_DOCIDS))?;
-        let word_fid_docids = env.create_database(Some(WORD_FIELD_ID_DOCIDS))?;
-        let field_id_word_count_docids = env.create_database(Some(FIELD_ID_WORD_COUNT_DOCIDS))?;
-        let word_prefix_position_docids = env.create_database(Some(WORD_PREFIX_POSITION_DOCIDS))?;
-        let word_prefix_fid_docids = env.create_database(Some(WORD_PREFIX_FIELD_ID_DOCIDS))?;
-        let facet_id_f64_docids = env.create_database(Some(FACET_ID_F64_DOCIDS))?;
-        let facet_id_string_docids = env.create_database(Some(FACET_ID_STRING_DOCIDS))?;
-        let facet_id_exists_docids = env.create_database(Some(FACET_ID_EXISTS_DOCIDS))?;
-        let facet_id_is_null_docids = env.create_database(Some(FACET_ID_IS_NULL_DOCIDS))?;
-        let facet_id_is_empty_docids = env.create_database(Some(FACET_ID_IS_EMPTY_DOCIDS))?;
-
-        let field_id_docid_facet_f64s = env.create_database(Some(FIELD_ID_DOCID_FACET_F64S))?;
+            env.create_database(&mut wtxn, Some(PREFIX_WORD_PAIR_PROXIMITY_DOCIDS))?;
+        let word_position_docids = env.create_database(&mut wtxn, Some(WORD_POSITION_DOCIDS))?;
+        let word_fid_docids = env.create_database(&mut wtxn, Some(WORD_FIELD_ID_DOCIDS))?;
+        let field_id_word_count_docids =
+            env.create_database(&mut wtxn, Some(FIELD_ID_WORD_COUNT_DOCIDS))?;
+        let word_prefix_position_docids =
+            env.create_database(&mut wtxn, Some(WORD_PREFIX_POSITION_DOCIDS))?;
+        let word_prefix_fid_docids =
+            env.create_database(&mut wtxn, Some(WORD_PREFIX_FIELD_ID_DOCIDS))?;
+        let facet_id_f64_docids = env.create_database(&mut wtxn, Some(FACET_ID_F64_DOCIDS))?;
+        let facet_id_string_docids =
+            env.create_database(&mut wtxn, Some(FACET_ID_STRING_DOCIDS))?;
+        let facet_id_exists_docids =
+            env.create_database(&mut wtxn, Some(FACET_ID_EXISTS_DOCIDS))?;
+        let facet_id_is_null_docids =
+            env.create_database(&mut wtxn, Some(FACET_ID_IS_NULL_DOCIDS))?;
+        let facet_id_is_empty_docids =
+            env.create_database(&mut wtxn, Some(FACET_ID_IS_EMPTY_DOCIDS))?;
+        let field_id_docid_facet_f64s =
+            env.create_database(&mut wtxn, Some(FIELD_ID_DOCID_FACET_F64S))?;
         let field_id_docid_facet_strings =
-            env.create_database(Some(FIELD_ID_DOCID_FACET_STRINGS))?;
-        let documents = env.create_database(Some(DOCUMENTS))?;
+            env.create_database(&mut wtxn, Some(FIELD_ID_DOCID_FACET_STRINGS))?;
+        let vector_id_docid = env.create_database(&mut wtxn, Some(VECTOR_ID_DOCID))?;
+        let documents = env.create_database(&mut wtxn, Some(DOCUMENTS))?;
+        wtxn.commit()?;
 
         Index::set_creation_dates(&env, main, created_at, updated_at)?;
 
@@ -207,7 +225,6 @@ impl Index {
             exact_word_docids,
             word_prefix_docids,
             exact_word_prefix_docids,
-            docid_word_positions,
             word_pair_proximity_docids,
             script_language_docids,
             word_prefix_pair_proximity_docids,
@@ -224,6 +241,7 @@ impl Index {
             facet_id_is_empty_docids,
             field_id_docid_facet_f64s,
             field_id_docid_facet_strings,
+            vector_id_docid,
             documents,
         })
     }
@@ -495,6 +513,26 @@ impl Index {
         }
     }
 
+    /* vector HNSW */
+
+    /// Writes the provided `hnsw`.
+    pub(crate) fn put_vector_hnsw(&self, wtxn: &mut RwTxn, hnsw: &Hnsw) -> heed::Result<()> {
+        self.main.put::<_, Str, SerdeBincode<Hnsw>>(wtxn, main_key::VECTOR_HNSW_KEY, hnsw)
+    }
+
+    /// Delete the `hnsw`.
+    pub(crate) fn delete_vector_hnsw(&self, wtxn: &mut RwTxn) -> heed::Result<bool> {
+        self.main.delete::<_, Str>(wtxn, main_key::VECTOR_HNSW_KEY)
+    }
+
+    /// Returns the `hnsw`.
+    pub fn vector_hnsw(&self, rtxn: &RoTxn) -> Result<Option<Hnsw>> {
+        match self.main.get::<_, Str, SerdeBincode<Hnsw>>(rtxn, main_key::VECTOR_HNSW_KEY)? {
+            Some(hnsw) => Ok(Some(hnsw)),
+            None => Ok(None),
+        }
+    }
+
     /* field distribution */
 
     /// Writes the field distribution which associates every field name with
@@ -1459,9 +1497,9 @@ pub(crate) mod tests {
 
         db_snap!(index, field_distribution,
             @r###"
-        age              1     
-        id               2     
-        name             2     
+        age              1      |
+        id               2      |
+        name             2      |
         "###
         );
 
@@ -1479,9 +1517,9 @@ pub(crate) mod tests {
 
         db_snap!(index, field_distribution,
             @r###"
-        age              1     
-        id               2     
-        name             2     
+        age              1      |
+        id               2      |
+        name             2      |
         "###
         );
 
@@ -1495,9 +1533,9 @@ pub(crate) mod tests {
 
         db_snap!(index, field_distribution,
             @r###"
-        has_dog          1     
-        id               2     
-        name             2     
+        has_dog          1      |
+        id               2      |
+        name             2      |
         "###
         );
     }

milli/src/lib.rs

@@ -5,57 +5,12 @@
 #[global_allocator]
 pub static ALLOC: mimalloc::MiMalloc = mimalloc::MiMalloc;
 
-// #[cfg(test)]
-// pub mod allocator {
-//     use std::alloc::{GlobalAlloc, System};
-//     use std::sync::atomic::{self, AtomicI64};
-
-//     #[global_allocator]
-//     pub static ALLOC: CountingAlloc = CountingAlloc {
-//         max_resident: AtomicI64::new(0),
-//         resident: AtomicI64::new(0),
-//         allocated: AtomicI64::new(0),
-//     };
-
-//     pub struct CountingAlloc {
-//         pub max_resident: AtomicI64,
-//         pub resident: AtomicI64,
-//         pub allocated: AtomicI64,
-//     }
-//     unsafe impl GlobalAlloc for CountingAlloc {
-//         unsafe fn alloc(&self, layout: std::alloc::Layout) -> *mut u8 {
-//             self.allocated.fetch_add(layout.size() as i64, atomic::Ordering::SeqCst);
-//             let old_resident =
-//                 self.resident.fetch_add(layout.size() as i64, atomic::Ordering::SeqCst);
-
-//             let resident = old_resident + layout.size() as i64;
-//             self.max_resident.fetch_max(resident, atomic::Ordering::SeqCst);
-
-//             // if layout.size() > 1_000_000 {
-//             //     eprintln!(
-//             //         "allocating {} with new resident size: {resident}",
-//             //         layout.size() / 1_000_000
-//             //     );
-//             //     // let trace = std::backtrace::Backtrace::capture();
-//             //     // let t = trace.to_string();
-//             //     // eprintln!("{t}");
-//             // }
-
-//             System.alloc(layout)
-//         }
-
-//         unsafe fn dealloc(&self, ptr: *mut u8, layout: std::alloc::Layout) {
-//             self.resident.fetch_sub(layout.size() as i64, atomic::Ordering::Relaxed);
-//             System.dealloc(ptr, layout)
-//         }
-//     }
-// }
-
 #[macro_use]
 pub mod documents;
 
 mod asc_desc;
 mod criterion;
+pub mod distance;
 mod error;
 mod external_documents_ids;
 pub mod facet;
@@ -75,6 +30,7 @@ use std::convert::{TryFrom, TryInto};
 use std::hash::BuildHasherDefault;
 
 use charabia::normalizer::{CharNormalizer, CompatibilityDecompositionNormalizer};
+pub use distance::{dot_product_similarity, euclidean_squared_distance};
 pub use filter_parser::{Condition, FilterCondition, Span, Token};
 use fxhash::{FxHasher32, FxHasher64};
 pub use grenad::CompressionType;
@@ -329,6 +285,35 @@ pub fn normalize_facet(original: &str) -> String {
     CompatibilityDecompositionNormalizer.normalize_str(original.trim()).to_lowercase()
 }
 
+/// Represents either a vector or an array of multiple vectors.
+#[derive(serde::Serialize, serde::Deserialize, Debug)]
+#[serde(transparent)]
+pub struct VectorOrArrayOfVectors {
+    #[serde(with = "either::serde_untagged")]
+    inner: either::Either<Vec<f32>, Vec<Vec<f32>>>,
+}
+
+impl VectorOrArrayOfVectors {
+    pub fn into_array_of_vectors(self) -> Vec<Vec<f32>> {
+        match self.inner {
+            either::Either::Left(vector) => vec![vector],
+            either::Either::Right(vectors) => vectors,
+        }
+    }
+}
+
+/// Normalize a vector by dividing the dimensions by the lenght of it.
+pub fn normalize_vector(mut vector: Vec<f32>) -> Vec<f32> {
+    let squared: f32 = vector.iter().map(|x| x * x).sum();
+    let length = squared.sqrt();
+    if length <= f32::EPSILON {
+        vector
+    } else {
+        vector.iter_mut().for_each(|x| *x /= length);
+        vector
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use serde_json::json;

milli/src/search/mod.rs

@@ -22,6 +22,7 @@ pub mod new;
 
 pub struct Search<'a> {
     query: Option<String>,
+    vector: Option<Vec<f32>>,
     // this should be linked to the String in the query
     filter: Option<Filter<'a>>,
     offset: usize,
@@ -39,6 +40,7 @@ impl<'a> Search<'a> {
     pub fn new(rtxn: &'a heed::RoTxn, index: &'a Index) -> Search<'a> {
         Search {
             query: None,
+            vector: None,
             filter: None,
             offset: 0,
             limit: 20,
@@ -57,6 +59,11 @@ impl<'a> Search<'a> {
         self
     }
 
+    pub fn vector(&mut self, vector: impl Into<Vec<f32>>) -> &mut Search<'a> {
+        self.vector = Some(vector.into());
+        self
+    }
+
     pub fn offset(&mut self, offset: usize) -> &mut Search<'a> {
         self.offset = offset;
         self
@@ -106,6 +113,7 @@ impl<'a> Search<'a> {
             execute_search(
                 &mut ctx,
                 &self.query,
+                &self.vector,
                 self.terms_matching_strategy,
                 self.exhaustive_number_hits,
                 &self.filter,
@@ -132,6 +140,7 @@ impl fmt::Debug for Search<'_> {
     fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
         let Search {
             query,
+            vector: _,
             filter,
             offset,
             limit,
@@ -145,6 +154,7 @@ impl fmt::Debug for Search<'_> {
         } = self;
         f.debug_struct("Search")
             .field("query", query)
+            .field("vector", &"[...]")
             .field("filter", filter)
             .field("offset", offset)
             .field("limit", limit)

milli/src/search/new/graph_based_ranking_rule.rs

@@ -46,7 +46,7 @@ use super::logger::SearchLogger;
 use super::query_graph::QueryNode;
 use super::ranking_rule_graph::{
     ConditionDocIdsCache, DeadEndsCache, ExactnessGraph, FidGraph, PositionGraph, ProximityGraph,
-    RankingRuleGraph, RankingRuleGraphTrait, TypoGraph,
+    RankingRuleGraph, RankingRuleGraphTrait, TypoGraph, WordsGraph,
 };
 use super::small_bitmap::SmallBitmap;
 use super::{QueryGraph, RankingRule, RankingRuleOutput, SearchContext};
@@ -54,6 +54,12 @@ use crate::search::new::query_term::LocatedQueryTermSubset;
 use crate::search::new::ranking_rule_graph::PathVisitor;
 use crate::{Result, TermsMatchingStrategy};
 
+pub type Words = GraphBasedRankingRule<WordsGraph>;
+impl GraphBasedRankingRule<WordsGraph> {
+    pub fn new(terms_matching_strategy: TermsMatchingStrategy) -> Self {
+        Self::new_with_id("words".to_owned(), Some(terms_matching_strategy))
+    }
+}
 pub type Proximity = GraphBasedRankingRule<ProximityGraph>;
 impl GraphBasedRankingRule<ProximityGraph> {
     pub fn new(terms_matching_strategy: Option<TermsMatchingStrategy>) -> Self {

milli/src/search/new/logger/visual.rs

@@ -4,7 +4,6 @@ use std::io::{BufWriter, Write};
 use std::path::{Path, PathBuf};
 use std::time::Instant;
 
-// use rand::random;
 use roaring::RoaringBitmap;
 
 use crate::search::new::interner::Interned;
@@ -13,6 +12,7 @@ use crate::search::new::query_term::LocatedQueryTermSubset;
 use crate::search::new::ranking_rule_graph::{
     Edge, FidCondition, FidGraph, PositionCondition, PositionGraph, ProximityCondition,
     ProximityGraph, RankingRuleGraph, RankingRuleGraphTrait, TypoCondition, TypoGraph,
+    WordsCondition, WordsGraph,
 };
 use crate::search::new::ranking_rules::BoxRankingRule;
 use crate::search::new::{QueryGraph, QueryNode, RankingRule, SearchContext, SearchLogger};
@@ -24,11 +24,12 @@ pub enum SearchEvents {
     RankingRuleSkipBucket { ranking_rule_idx: usize, bucket_len: u64 },
     RankingRuleEndIteration { ranking_rule_idx: usize, universe_len: u64 },
     ExtendResults { new: Vec<u32> },
-    WordsGraph { query_graph: QueryGraph },
     ProximityGraph { graph: RankingRuleGraph<ProximityGraph> },
     ProximityPaths { paths: Vec<Vec<Interned<ProximityCondition>>> },
     TypoGraph { graph: RankingRuleGraph<TypoGraph> },
     TypoPaths { paths: Vec<Vec<Interned<TypoCondition>>> },
+    WordsGraph { graph: RankingRuleGraph<WordsGraph> },
+    WordsPaths { paths: Vec<Vec<Interned<WordsCondition>>> },
     FidGraph { graph: RankingRuleGraph<FidGraph> },
     FidPaths { paths: Vec<Vec<Interned<FidCondition>>> },
     PositionGraph { graph: RankingRuleGraph<PositionGraph> },
@@ -139,8 +140,11 @@ impl SearchLogger<QueryGraph> for VisualSearchLogger {
         let Some(location) = self.location.last() else { return };
         match location {
             Location::Words => {
-                if let Some(query_graph) = state.downcast_ref::<QueryGraph>() {
-                    self.events.push(SearchEvents::WordsGraph { query_graph: query_graph.clone() });
+                if let Some(graph) = state.downcast_ref::<RankingRuleGraph<WordsGraph>>() {
+                    self.events.push(SearchEvents::WordsGraph { graph: graph.clone() });
+                }
+                if let Some(paths) = state.downcast_ref::<Vec<Vec<Interned<WordsCondition>>>>() {
+                    self.events.push(SearchEvents::WordsPaths { paths: paths.clone() });
                 }
             }
             Location::Typo => {
@@ -329,7 +333,6 @@ impl<'ctx> DetailedLoggerFinish<'ctx> {
             SearchEvents::ExtendResults { new } => {
                 self.write_extend_results(new)?;
             }
-            SearchEvents::WordsGraph { query_graph } => self.write_words_graph(query_graph)?,
             SearchEvents::ProximityGraph { graph } => self.write_rr_graph(&graph)?,
             SearchEvents::ProximityPaths { paths } => {
                 self.write_rr_graph_paths::<ProximityGraph>(paths)?;
@@ -338,6 +341,10 @@ impl<'ctx> DetailedLoggerFinish<'ctx> {
             SearchEvents::TypoPaths { paths } => {
                 self.write_rr_graph_paths::<TypoGraph>(paths)?;
             }
+            SearchEvents::WordsGraph { graph } => self.write_rr_graph(&graph)?,
+            SearchEvents::WordsPaths { paths } => {
+                self.write_rr_graph_paths::<WordsGraph>(paths)?;
+            }
             SearchEvents::FidGraph { graph } => self.write_rr_graph(&graph)?,
             SearchEvents::FidPaths { paths } => {
                 self.write_rr_graph_paths::<FidGraph>(paths)?;
@@ -455,7 +462,7 @@ fill: \"#B6E2D3\"
                 shape: class
                 max_nbr_typo: {}",
                     term_subset.description(ctx),
-                    term_subset.max_nbr_typos(ctx)
+                    term_subset.max_typo_cost(ctx)
                 )?;
 
                 for w in term_subset.all_single_words_except_prefix_db(ctx)? {
@@ -482,13 +489,6 @@ fill: \"#B6E2D3\"
         }
         Ok(())
     }
-    fn write_words_graph(&mut self, qg: QueryGraph) -> Result<()> {
-        self.make_new_file_for_internal_state_if_needed()?;
-
-        self.write_query_graph(&qg)?;
-
-        Ok(())
-    }
     fn write_rr_graph<R: RankingRuleGraphTrait>(
         &mut self,
         graph: &RankingRuleGraph<R>,

milli/src/search/new/matches/mod.rs

@@ -509,6 +509,7 @@ mod tests {
             let crate::search::PartialSearchResult { located_query_terms, .. } = execute_search(
                 &mut ctx,
                 &Some(query.to_string()),
+                &None,
                 crate::TermsMatchingStrategy::default(),
                 false,
                 &None,

milli/src/search/new/mod.rs

@@ -15,11 +15,7 @@ mod resolve_query_graph;
 mod small_bitmap;
 
 mod exact_attribute;
-// TODO: documentation + comments
-// implementation is currently an adaptation of the previous implementation to fit with the new model
 mod sort;
-// TODO: documentation + comments
-mod words;
 
 #[cfg(test)]
 mod tests;
@@ -32,6 +28,7 @@ use db_cache::DatabaseCache;
 use exact_attribute::ExactAttribute;
 use graph_based_ranking_rule::{Exactness, Fid, Position, Proximity, Typo};
 use heed::RoTxn;
+use hnsw::Searcher;
 use interner::{DedupInterner, Interner};
 pub use logger::visual::VisualSearchLogger;
 pub use logger::{DefaultSearchLogger, SearchLogger};
@@ -43,13 +40,17 @@ use ranking_rules::{
 use resolve_query_graph::{compute_query_graph_docids, PhraseDocIdsCache};
 use roaring::RoaringBitmap;
 use sort::Sort;
-use words::Words;
+use space::Neighbor;
 
 use self::geo_sort::GeoSort;
 pub use self::geo_sort::Strategy as GeoSortStrategy;
+use self::graph_based_ranking_rule::Words;
 use self::interner::Interned;
 use crate::search::new::distinct::apply_distinct_rule;
-use crate::{AscDesc, DocumentId, Filter, Index, Member, Result, TermsMatchingStrategy, UserError};
+use crate::{
+    normalize_vector, AscDesc, DocumentId, Filter, Index, Member, Result, TermsMatchingStrategy,
+    UserError, BEU32,
+};
 
 /// A structure used throughout the execution of a search query.
 pub struct SearchContext<'ctx> {
@@ -202,6 +203,11 @@ fn get_ranking_rules_for_query_graph_search<'ctx>(
     let mut sorted_fields = HashSet::new();
     let mut geo_sorted = false;
 
+    // Don't add the `words` ranking rule if the term matching strategy is `All`
+    if matches!(terms_matching_strategy, TermsMatchingStrategy::All) {
+        words = true;
+    }
+
     let mut ranking_rules: Vec<BoxRankingRule<QueryGraph>> = vec![];
     let settings_ranking_rules = ctx.index.criteria(ctx.txn)?;
     for rr in settings_ranking_rules {
@@ -348,6 +354,7 @@ fn resolve_sort_criteria<'ctx, Query: RankingRuleQueryTrait>(
 pub fn execute_search(
     ctx: &mut SearchContext,
     query: &Option<String>,
+    vector: &Option<Vec<f32>>,
     terms_matching_strategy: TermsMatchingStrategy,
     exhaustive_number_hits: bool,
     filters: &Option<Filter>,
@@ -397,8 +404,8 @@ pub fn execute_search(
         None
     };
     let bucket_sort_output = if let Some(query_terms) = query_terms {
-        let graph = QueryGraph::from_query(ctx, &query_terms)?;
-        located_query_terms = Some(query_terms);
+        let (graph, new_located_query_terms) = QueryGraph::from_query(ctx, &query_terms)?;
+        located_query_terms = Some(new_located_query_terms);
 
         let ranking_rules = get_ranking_rules_for_query_graph_search(
             ctx,
@@ -427,6 +434,34 @@ pub fn execute_search(
 
     let BucketSortOutput { docids, mut all_candidates } = bucket_sort_output;
 
+    let docids = match vector {
+        Some(vector) => {
+            // return the nearest documents that are also part of the candidates.
+            let mut searcher = Searcher::new();
+            let hnsw = ctx.index.vector_hnsw(ctx.txn)?.unwrap_or_default();
+            let ef = hnsw.len().min(100);
+            let mut dest = vec![Neighbor { index: 0, distance: 0 }; ef];
+            let vector = normalize_vector(vector.clone());
+            let neighbors = hnsw.nearest(&vector, ef, &mut searcher, &mut dest[..]);
+
+            let mut docids = Vec::new();
+            for Neighbor { index, distance: _ } in neighbors.iter() {
+                let index = BEU32::new(*index as u32);
+                let docid = ctx.index.vector_id_docid.get(ctx.txn, &index)?.unwrap().get();
+                if universe.contains(docid) {
+                    docids.push(docid);
+                    if docids.len() == (from + length) {
+                        break;
+                    }
+                }
+            }
+
+            docids.into_iter().skip(from).take(length).collect()
+        }
+        // return the search docids if the vector field is not specified
+        None => docids,
+    };
+
     // The candidates is the universe unless the exhaustive number of hits
     // is requested and a distinct attribute is set.
     if exhaustive_number_hits {

milli/src/search/new/query_graph.rs

@@ -88,12 +88,15 @@ pub struct QueryGraph {
 }
 
 impl QueryGraph {
-    /// Build the query graph from the parsed user search query.
+    /// Build the query graph from the parsed user search query, return an updated list of the located query terms
+    /// which contains ngrams.
     pub fn from_query(
         ctx: &mut SearchContext,
         // NOTE: the terms here must be consecutive
         terms: &[LocatedQueryTerm],
-    ) -> Result<QueryGraph> {
+    ) -> Result<(QueryGraph, Vec<LocatedQueryTerm>)> {
+        let mut new_located_query_terms = terms.to_vec();
+
         let nbr_typos = number_of_typos_allowed(ctx)?;
 
         let mut nodes_data: Vec<QueryNodeData> = vec![QueryNodeData::Start, QueryNodeData::End];
@@ -107,10 +110,11 @@ impl QueryGraph {
         let original_terms_len = terms.len();
         for term_idx in 0..original_terms_len {
             let mut new_nodes = vec![];
+
             let new_node_idx = add_node(
                 &mut nodes_data,
                 QueryNodeData::Term(LocatedQueryTermSubset {
-                    term_subset: QueryTermSubset::full(Interned::from_raw(term_idx as u16)),
+                    term_subset: QueryTermSubset::full(terms[term_idx].value),
                     positions: terms[term_idx].positions.clone(),
                     term_ids: term_idx as u8..=term_idx as u8,
                 }),
@@ -121,6 +125,7 @@ impl QueryGraph {
                 if let Some(ngram) =
                     query_term::make_ngram(ctx, &terms[term_idx - 1..=term_idx], &nbr_typos)?
                 {
+                    new_located_query_terms.push(ngram.clone());
                     let ngram_idx = add_node(
                         &mut nodes_data,
                         QueryNodeData::Term(LocatedQueryTermSubset {
@@ -136,6 +141,7 @@ impl QueryGraph {
                 if let Some(ngram) =
                     query_term::make_ngram(ctx, &terms[term_idx - 2..=term_idx], &nbr_typos)?
                 {
+                    new_located_query_terms.push(ngram.clone());
                     let ngram_idx = add_node(
                         &mut nodes_data,
                         QueryNodeData::Term(LocatedQueryTermSubset {
@@ -167,7 +173,7 @@ impl QueryGraph {
         let mut graph = QueryGraph { root_node, end_node, nodes };
         graph.build_initial_edges();
 
-        Ok(graph)
+        Ok((graph, new_located_query_terms))
     }
 
     /// Remove the given nodes, connecting all their predecessors to all their successors.

milli/src/search/new/query_term/compute_derivations.rs

@@ -28,16 +28,14 @@ pub enum ZeroOrOneTypo {
 impl Interned<QueryTerm> {
     pub fn compute_fully_if_needed(self, ctx: &mut SearchContext) -> Result<()> {
         let s = ctx.term_interner.get_mut(self);
-        if s.max_nbr_typos == 0 {
-            s.one_typo = Lazy::Init(OneTypoTerm::default());
-            s.two_typo = Lazy::Init(TwoTypoTerm::default());
-        } else if s.max_nbr_typos == 1 && s.one_typo.is_uninit() {
+        if s.max_levenshtein_distance <= 1 && s.one_typo.is_uninit() {
             assert!(s.two_typo.is_uninit());
+            // Initialize one_typo subterm even if max_nbr_typo is 0 because of split words
             self.initialize_one_typo_subterm(ctx)?;
             let s = ctx.term_interner.get_mut(self);
             assert!(s.one_typo.is_init());
             s.two_typo = Lazy::Init(TwoTypoTerm::default());
-        } else if s.max_nbr_typos > 1 && s.two_typo.is_uninit() {
+        } else if s.max_levenshtein_distance > 1 && s.two_typo.is_uninit() {
             assert!(s.two_typo.is_uninit());
             self.initialize_one_and_two_typo_subterm(ctx)?;
             let s = ctx.term_interner.get_mut(self);
@@ -187,7 +185,7 @@ pub fn partially_initialized_term_from_word(
                 original: ctx.word_interner.insert(word.to_owned()),
                 ngram_words: None,
                 is_prefix: false,
-                max_nbr_typos: 0,
+                max_levenshtein_distance: 0,
                 zero_typo: <_>::default(),
                 one_typo: Lazy::Init(<_>::default()),
                 two_typo: Lazy::Init(<_>::default()),
@@ -258,7 +256,7 @@ pub fn partially_initialized_term_from_word(
     Ok(QueryTerm {
         original: word_interned,
         ngram_words: None,
-        max_nbr_typos: max_typo,
+        max_levenshtein_distance: max_typo,
         is_prefix,
         zero_typo,
         one_typo: Lazy::Uninit,
@@ -277,7 +275,16 @@ fn find_split_words(ctx: &mut SearchContext, word: &str) -> Result<Option<Intern
 impl Interned<QueryTerm> {
     fn initialize_one_typo_subterm(self, ctx: &mut SearchContext) -> Result<()> {
         let self_mut = ctx.term_interner.get_mut(self);
-        let QueryTerm { original, is_prefix, one_typo, .. } = self_mut;
+
+        let allows_split_words = self_mut.allows_split_words();
+        let QueryTerm {
+            original,
+            is_prefix,
+            one_typo,
+            max_levenshtein_distance: max_nbr_typos,
+            ..
+        } = self_mut;
+
         let original = *original;
         let is_prefix = *is_prefix;
         // let original_str = ctx.word_interner.get(*original).to_owned();
@@ -286,26 +293,33 @@ impl Interned<QueryTerm> {
         }
         let mut one_typo_words = BTreeSet::new();
 
-        find_zero_one_typo_derivations(ctx, original, is_prefix, |derived_word, nbr_typos| {
-            match nbr_typos {
-                ZeroOrOneTypo::Zero => {}
-                ZeroOrOneTypo::One => {
-                    if one_typo_words.len() < limits::MAX_ONE_TYPO_COUNT {
-                        one_typo_words.insert(derived_word);
-                    } else {
-                        return Ok(ControlFlow::Break(()));
+        if *max_nbr_typos > 0 {
+            find_zero_one_typo_derivations(ctx, original, is_prefix, |derived_word, nbr_typos| {
+                match nbr_typos {
+                    ZeroOrOneTypo::Zero => {}
+                    ZeroOrOneTypo::One => {
+                        if one_typo_words.len() < limits::MAX_ONE_TYPO_COUNT {
+                            one_typo_words.insert(derived_word);
+                        } else {
+                            return Ok(ControlFlow::Break(()));
+                        }
                     }
                 }
-            }
-            Ok(ControlFlow::Continue(()))
-        })?;
-        let original_str = ctx.word_interner.get(original).to_owned();
-        let split_words = find_split_words(ctx, original_str.as_str())?;
+                Ok(ControlFlow::Continue(()))
+            })?;
+        }
+
+        let split_words = if allows_split_words {
+            let original_str = ctx.word_interner.get(original).to_owned();
+            find_split_words(ctx, original_str.as_str())?
+        } else {
+            None
+        };
 
         let self_mut = ctx.term_interner.get_mut(self);
 
         // Only add the split words to the derivations if:
-        // 1. the term is not an ngram; OR
+        // 1. the term is neither an ngram nor a phrase; OR
         // 2. the term is an ngram, but the split words are different from the ngram's component words
         let split_words = if let Some((ngram_words, split_words)) =
             self_mut.ngram_words.as_ref().zip(split_words.as_ref())
@@ -327,7 +341,13 @@ impl Interned<QueryTerm> {
     }
     fn initialize_one_and_two_typo_subterm(self, ctx: &mut SearchContext) -> Result<()> {
         let self_mut = ctx.term_interner.get_mut(self);
-        let QueryTerm { original, is_prefix, two_typo, .. } = self_mut;
+        let QueryTerm {
+            original,
+            is_prefix,
+            two_typo,
+            max_levenshtein_distance: max_nbr_typos,
+            ..
+        } = self_mut;
         let original_str = ctx.word_interner.get(*original).to_owned();
         if two_typo.is_init() {
             return Ok(());
@@ -335,34 +355,37 @@ impl Interned<QueryTerm> {
         let mut one_typo_words = BTreeSet::new();
         let mut two_typo_words = BTreeSet::new();
 
-        find_zero_one_two_typo_derivations(
-            *original,
-            *is_prefix,
-            ctx.index.words_fst(ctx.txn)?,
-            &mut ctx.word_interner,
-            |derived_word, nbr_typos| {
-                if one_typo_words.len() >= limits::MAX_ONE_TYPO_COUNT
-                    && two_typo_words.len() >= limits::MAX_TWO_TYPOS_COUNT
-                {
-                    // No chance we will add either one- or two-typo derivations anymore, stop iterating.
-                    return Ok(ControlFlow::Break(()));
-                }
-                match nbr_typos {
-                    NumberOfTypos::Zero => {}
-                    NumberOfTypos::One => {
-                        if one_typo_words.len() < limits::MAX_ONE_TYPO_COUNT {
-                            one_typo_words.insert(derived_word);
+        if *max_nbr_typos > 0 {
+            find_zero_one_two_typo_derivations(
+                *original,
+                *is_prefix,
+                ctx.index.words_fst(ctx.txn)?,
+                &mut ctx.word_interner,
+                |derived_word, nbr_typos| {
+                    if one_typo_words.len() >= limits::MAX_ONE_TYPO_COUNT
+                        && two_typo_words.len() >= limits::MAX_TWO_TYPOS_COUNT
+                    {
+                        // No chance we will add either one- or two-typo derivations anymore, stop iterating.
+                        return Ok(ControlFlow::Break(()));
+                    }
+                    match nbr_typos {
+                        NumberOfTypos::Zero => {}
+                        NumberOfTypos::One => {
+                            if one_typo_words.len() < limits::MAX_ONE_TYPO_COUNT {
+                                one_typo_words.insert(derived_word);
+                            }
+                        }
+                        NumberOfTypos::Two => {
+                            if two_typo_words.len() < limits::MAX_TWO_TYPOS_COUNT {
+                                two_typo_words.insert(derived_word);
+                            }
                         }
                     }
-                    NumberOfTypos::Two => {
-                        if two_typo_words.len() < limits::MAX_TWO_TYPOS_COUNT {
-                            two_typo_words.insert(derived_word);
-                        }
-                    }
-                }
-                Ok(ControlFlow::Continue(()))
-            },
-        )?;
+                    Ok(ControlFlow::Continue(()))
+                },
+            )?;
+        }
+
         let split_words = find_split_words(ctx, original_str.as_str())?;
         let self_mut = ctx.term_interner.get_mut(self);
 

milli/src/search/new/query_term/mod.rs

@@ -43,7 +43,7 @@ pub struct QueryTermSubset {
 pub struct QueryTerm {
     original: Interned<String>,
     ngram_words: Option<Vec<Interned<String>>>,
-    max_nbr_typos: u8,
+    max_levenshtein_distance: u8,
     is_prefix: bool,
     zero_typo: ZeroTypoTerm,
     // May not be computed yet
@@ -342,10 +342,16 @@ impl QueryTermSubset {
         }
         None
     }
-    pub fn max_nbr_typos(&self, ctx: &SearchContext) -> u8 {
+    pub fn max_typo_cost(&self, ctx: &SearchContext) -> u8 {
         let t = ctx.term_interner.get(self.original);
-        match t.max_nbr_typos {
-            0 => 0,
+        match t.max_levenshtein_distance {
+            0 => {
+                if t.allows_split_words() {
+                    1
+                } else {
+                    0
+                }
+            }
             1 => {
                 if self.one_typo_subset.is_empty() {
                     0
@@ -438,6 +444,9 @@ impl QueryTerm {
 
         self.zero_typo.is_empty() && one_typo.is_empty() && two_typo.is_empty()
     }
+    fn allows_split_words(&self) -> bool {
+        self.zero_typo.phrase.is_none()
+    }
 }
 
 impl Interned<QueryTerm> {

milli/src/search/new/query_term/parse_query.rs

@@ -77,13 +77,9 @@ pub fn located_query_terms_from_tokens(
                 }
             }
             TokenKind::Separator(separator_kind) => {
-                match separator_kind {
-                    SeparatorKind::Hard => {
-                        position += 1;
-                    }
-                    SeparatorKind::Soft => {
-                        position += 0;
-                    }
+                // add penalty for hard separators
+                if let SeparatorKind::Hard = separator_kind {
+                    position = position.wrapping_add(7);
                 }
 
                 phrase = 'phrase: {
@@ -217,7 +213,7 @@ pub fn make_ngram(
         original: ngram_str_interned,
         ngram_words: Some(words_interned),
         is_prefix,
-        max_nbr_typos,
+        max_levenshtein_distance: max_nbr_typos,
         zero_typo: term.zero_typo,
         one_typo: Lazy::Uninit,
         two_typo: Lazy::Uninit,
@@ -271,7 +267,7 @@ impl PhraseBuilder {
                 QueryTerm {
                     original: ctx.word_interner.insert(phrase_desc),
                     ngram_words: None,
-                    max_nbr_typos: 0,
+                    max_levenshtein_distance: 0,
                     is_prefix: false,
                     zero_typo: ZeroTypoTerm {
                         phrase: Some(phrase),
@@ -288,3 +284,36 @@ impl PhraseBuilder {
         })
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use charabia::TokenizerBuilder;
+
+    use super::*;
+    use crate::index::tests::TempIndex;
+
+    fn temp_index_with_documents() -> TempIndex {
+        let temp_index = TempIndex::new();
+        temp_index
+            .add_documents(documents!([
+                { "id": 1, "name": "split this world westfali westfalia the Ŵôřlḑôle" },
+                { "id": 2, "name": "Westfália" },
+                { "id": 3, "name": "Ŵôřlḑôle" },
+            ]))
+            .unwrap();
+        temp_index
+    }
+
+    #[test]
+    fn start_with_hard_separator() -> Result<()> {
+        let tokenizer = TokenizerBuilder::new().build();
+        let tokens = tokenizer.tokenize(".");
+        let index = temp_index_with_documents();
+        let rtxn = index.read_txn()?;
+        let mut ctx = SearchContext::new(&index, &rtxn);
+        // panics with `attempt to add with overflow` before <https://github.com/meilisearch/meilisearch/issues/3785>
+        let located_query_terms = located_query_terms_from_tokens(&mut ctx, tokens, None)?;
+        assert!(located_query_terms.is_empty());
+        Ok(())
+    }
+}

milli/src/search/new/ranking_rule_graph/cheapest_paths.rs

@@ -205,18 +205,12 @@ impl<G: RankingRuleGraphTrait> VisitorState<G> {
 impl<G: RankingRuleGraphTrait> RankingRuleGraph<G> {
     pub fn find_all_costs_to_end(&self) -> MappedInterner<QueryNode, Vec<u64>> {
         let mut costs_to_end = self.query_graph.nodes.map(|_| vec![]);
-        let mut enqueued = SmallBitmap::new(self.query_graph.nodes.len());
 
-        let mut node_stack = VecDeque::new();
-
-        *costs_to_end.get_mut(self.query_graph.end_node) = vec![0];
-
-        for prev_node in self.query_graph.nodes.get(self.query_graph.end_node).predecessors.iter() {
-            node_stack.push_back(prev_node);
-            enqueued.insert(prev_node);
-        }
-
-        while let Some(cur_node) = node_stack.pop_front() {
+        self.traverse_breadth_first_backward(self.query_graph.end_node, |cur_node| {
+            if cur_node == self.query_graph.end_node {
+                *costs_to_end.get_mut(self.query_graph.end_node) = vec![0];
+                return;
+            }
             let mut self_costs = Vec::<u64>::new();
 
             let cur_node_edges = &self.edges_of_node.get(cur_node);
@@ -232,13 +226,7 @@ impl<G: RankingRuleGraphTrait> RankingRuleGraph<G> {
             self_costs.dedup();
 
             *costs_to_end.get_mut(cur_node) = self_costs;
-            for prev_node in self.query_graph.nodes.get(cur_node).predecessors.iter() {
-                if !enqueued.contains(prev_node) {
-                    node_stack.push_back(prev_node);
-                    enqueued.insert(prev_node);
-                }
-            }
-        }
+        });
         costs_to_end
     }
 
@@ -247,17 +235,12 @@ impl<G: RankingRuleGraphTrait> RankingRuleGraph<G> {
         node_with_removed_outgoing_conditions: Interned<QueryNode>,
         costs: &mut MappedInterner<QueryNode, Vec<u64>>,
     ) {
-        let mut enqueued = SmallBitmap::new(self.query_graph.nodes.len());
-        let mut node_stack = VecDeque::new();
-
-        enqueued.insert(node_with_removed_outgoing_conditions);
-        node_stack.push_back(node_with_removed_outgoing_conditions);
-
-        'main_loop: while let Some(cur_node) = node_stack.pop_front() {
+        // Traverse the graph backward from the target node, recomputing the cost for each of its predecessors.
+        // We first check that no other node is contributing the same total cost to a predecessor before removing
+        // the cost from the predecessor.
+        self.traverse_breadth_first_backward(node_with_removed_outgoing_conditions, |cur_node| {
             let mut costs_to_remove = FxHashSet::default();
-            for c in costs.get(cur_node) {
-                costs_to_remove.insert(*c);
-            }
+            costs_to_remove.extend(costs.get(cur_node).iter().copied());
 
             let cur_node_edges = &self.edges_of_node.get(cur_node);
             for edge_idx in cur_node_edges.iter() {
@@ -265,22 +248,75 @@ impl<G: RankingRuleGraphTrait> RankingRuleGraph<G> {
                 for cost in costs.get(edge.dest_node).iter() {
                     costs_to_remove.remove(&(*cost + edge.cost as u64));
                     if costs_to_remove.is_empty() {
-                        continue 'main_loop;
+                        return;
                     }
                 }
             }
             if costs_to_remove.is_empty() {
-                continue 'main_loop;
+                return;
             }
             let mut new_costs = BTreeSet::from_iter(costs.get(cur_node).iter().copied());
             for c in costs_to_remove {
                 new_costs.remove(&c);
             }
             *costs.get_mut(cur_node) = new_costs.into_iter().collect();
+        });
+    }
 
+    /// Traverse the graph backwards from the given node such that every time
+    /// a node is visited, we are guaranteed that all its successors either:
+    /// 1. have already been visited; OR
+    /// 2. were not reachable from the given node
+    pub fn traverse_breadth_first_backward(
+        &self,
+        from: Interned<QueryNode>,
+        mut visit: impl FnMut(Interned<QueryNode>),
+    ) {
+        let mut reachable = SmallBitmap::for_interned_values_in(&self.query_graph.nodes);
+        {
+            // go backward to get the set of all reachable nodes from the given node
+            // the nodes that are not reachable will be set as `visited`
+            let mut stack = VecDeque::new();
+            let mut enqueued = SmallBitmap::for_interned_values_in(&self.query_graph.nodes);
+            enqueued.insert(from);
+            stack.push_back(from);
+            while let Some(n) = stack.pop_front() {
+                if reachable.contains(n) {
+                    continue;
+                }
+                reachable.insert(n);
+                for prev_node in self.query_graph.nodes.get(n).predecessors.iter() {
+                    if !enqueued.contains(prev_node) && !reachable.contains(prev_node) {
+                        stack.push_back(prev_node);
+                        enqueued.insert(prev_node);
+                    }
+                }
+            }
+        };
+        let mut unreachable_or_visited =
+            SmallBitmap::for_interned_values_in(&self.query_graph.nodes);
+        for (n, _) in self.query_graph.nodes.iter() {
+            if !reachable.contains(n) {
+                unreachable_or_visited.insert(n);
+            }
+        }
+
+        let mut enqueued = SmallBitmap::for_interned_values_in(&self.query_graph.nodes);
+        let mut stack = VecDeque::new();
+
+        enqueued.insert(from);
+        stack.push_back(from);
+
+        while let Some(cur_node) = stack.pop_front() {
+            if !self.query_graph.nodes.get(cur_node).successors.is_subset(&unreachable_or_visited) {
+                stack.push_back(cur_node);
+                continue;
+            }
+            unreachable_or_visited.insert(cur_node);
+            visit(cur_node);
             for prev_node in self.query_graph.nodes.get(cur_node).predecessors.iter() {
-                if !enqueued.contains(prev_node) {
-                    node_stack.push_back(prev_node);
+                if !enqueued.contains(prev_node) && !unreachable_or_visited.contains(prev_node) {
+                    stack.push_back(prev_node);
                     enqueued.insert(prev_node);
                 }
             }

milli/src/search/new/ranking_rule_graph/mod.rs

@@ -20,6 +20,8 @@ mod position;
 mod proximity;
 /// Implementation of the `typo` ranking rule
 mod typo;
+/// Implementation of the `words` ranking rule
+mod words;
 
 use std::collections::BTreeSet;
 use std::hash::Hash;
@@ -33,6 +35,7 @@ pub use position::{PositionCondition, PositionGraph};
 pub use proximity::{ProximityCondition, ProximityGraph};
 use roaring::RoaringBitmap;
 pub use typo::{TypoCondition, TypoGraph};
+pub use words::{WordsCondition, WordsGraph};
 
 use super::interner::{DedupInterner, FixedSizeInterner, Interned, MappedInterner};
 use super::query_term::LocatedQueryTermSubset;

milli/src/search/new/ranking_rule_graph/position/mod.rs

@@ -111,23 +111,16 @@ impl RankingRuleGraphTrait for PositionGraph {
 
 fn cost_from_position(sum_positions: u32) -> u32 {
     match sum_positions {
-        0 | 1 | 2 | 3 => sum_positions,
-        4 | 5 => 4,
-        6 | 7 => 5,
-        8 | 9 => 6,
-        10 | 11 => 7,
-        12 | 13 => 8,
-        14 | 15 => 9,
-        16 | 17..=24 => 10,
-        25..=32 => 11,
-        33..=64 => 12,
-        65..=128 => 13,
-        129..=256 => 14,
-        257..=512 => 15,
-        513..=1024 => 16,
-        1025..=2048 => 17,
-        2049..=4096 => 18,
-        4097..=8192 => 19,
-        _ => 20,
+        0 => 0,
+        1 => 1,
+        2..=4 => 2,
+        5..=7 => 3,
+        8..=11 => 4,
+        12..=16 => 5,
+        17..=24 => 6,
+        25..=64 => 7,
+        65..=256 => 8,
+        257..=1024 => 9,
+        _ => 10,
     }
 }

milli/src/search/new/ranking_rule_graph/typo/mod.rs

@@ -50,7 +50,7 @@ impl RankingRuleGraphTrait for TypoGraph {
         // 3-gram -> equivalent to 2 typos
         let base_cost = if term.term_ids.len() == 1 { 0 } else { term.term_ids.len() as u32 };
 
-        for nbr_typos in 0..=term.term_subset.max_nbr_typos(ctx) {
+        for nbr_typos in 0..=term.term_subset.max_typo_cost(ctx) {
             let mut term = term.clone();
             match nbr_typos {
                 0 => {

milli/src/search/new/ranking_rule_graph/words/mod.rs

@@ -0,0 +1,49 @@
+use roaring::RoaringBitmap;
+
+use super::{ComputedCondition, RankingRuleGraphTrait};
+use crate::search::new::interner::{DedupInterner, Interned};
+use crate::search::new::query_term::LocatedQueryTermSubset;
+use crate::search::new::resolve_query_graph::compute_query_term_subset_docids;
+use crate::search::new::SearchContext;
+use crate::Result;
+
+#[derive(Clone, PartialEq, Eq, Hash)]
+pub struct WordsCondition {
+    term: LocatedQueryTermSubset,
+}
+
+pub enum WordsGraph {}
+
+impl RankingRuleGraphTrait for WordsGraph {
+    type Condition = WordsCondition;
+
+    fn resolve_condition(
+        ctx: &mut SearchContext,
+        condition: &Self::Condition,
+        universe: &RoaringBitmap,
+    ) -> Result<ComputedCondition> {
+        let WordsCondition { term, .. } = condition;
+        // maybe compute_query_term_subset_docids should accept a universe as argument
+        let mut docids = compute_query_term_subset_docids(ctx, &term.term_subset)?;
+        docids &= universe;
+
+        Ok(ComputedCondition {
+            docids,
+            universe_len: universe.len(),
+            start_term_subset: None,
+            end_term_subset: term.clone(),
+        })
+    }
+
+    fn build_edges(
+        _ctx: &mut SearchContext,
+        conditions_interner: &mut DedupInterner<Self::Condition>,
+        _from: Option<&LocatedQueryTermSubset>,
+        to_term: &LocatedQueryTermSubset,
+    ) -> Result<Vec<(u32, Interned<Self::Condition>)>> {
+        Ok(vec![(
+            to_term.term_ids.len() as u32,
+            conditions_interner.insert(WordsCondition { term: to_term.clone() }),
+        )])
+    }
+}

milli/src/search/new/tests/attribute_position.rs

@@ -138,7 +138,7 @@ fn test_attribute_position_simple() {
     s.terms_matching_strategy(TermsMatchingStrategy::All);
     s.query("quick brown");
     let SearchResult { documents_ids, .. } = s.execute().unwrap();
-    insta::assert_snapshot!(format!("{documents_ids:?}"), @"[10, 11, 12, 13, 3, 4, 2, 1, 0, 6, 8, 7, 9, 5]");
+    insta::assert_snapshot!(format!("{documents_ids:?}"), @"[10, 11, 12, 13, 2, 3, 4, 1, 0, 6, 8, 7, 9, 5]");
 }
 #[test]
 fn test_attribute_position_repeated() {
@@ -163,7 +163,7 @@ fn test_attribute_position_different_fields() {
     s.terms_matching_strategy(TermsMatchingStrategy::All);
     s.query("quick brown");
     let SearchResult { documents_ids, .. } = s.execute().unwrap();
-    insta::assert_snapshot!(format!("{documents_ids:?}"), @"[10, 11, 12, 13, 3, 4, 2, 1, 0, 6, 8, 7, 9, 5]");
+    insta::assert_snapshot!(format!("{documents_ids:?}"), @"[10, 11, 12, 13, 2, 3, 4, 1, 0, 6, 8, 7, 9, 5]");
 }
 
 #[test]
@@ -176,5 +176,5 @@ fn test_attribute_position_ngrams() {
     s.terms_matching_strategy(TermsMatchingStrategy::All);
     s.query("quick brown");
     let SearchResult { documents_ids, .. } = s.execute().unwrap();
-    insta::assert_snapshot!(format!("{documents_ids:?}"), @"[10, 11, 12, 13, 3, 4, 2, 1, 0, 6, 8, 7, 9, 5]");
+    insta::assert_snapshot!(format!("{documents_ids:?}"), @"[10, 11, 12, 13, 2, 3, 4, 1, 0, 6, 8, 7, 9, 5]");
 }

milli/src/search/new/tests/ngram_split_words.rs

@@ -3,9 +3,9 @@ This module tests the following properties:
 
 1. Two consecutive words from a query can be combined into a "2gram"
 2. Three consecutive words from a query can be combined into a "3gram"
-3. A word from the query can be split into two consecutive words (split words)
+3. A word from the query can be split into two consecutive words (split words), no matter how short it is
 4. A 2gram can be split into two words
-5. A 3gram cannot be split into two words
+5. A 3gram can be split into two words
 6. 2grams can contain up to 1 typo
 7. 3grams cannot have typos
 8. 2grams and 3grams can be prefix tolerant
@@ -14,6 +14,7 @@ This module tests the following properties:
 11. Disabling typo tolerance does not disable ngram tolerance
 12. Prefix tolerance is disabled for the last word if a space follows it
 13. Ngrams cannot be formed by combining a phrase and a word or two phrases
+14. Split words are not disabled by the `disableOnAttribute` or `disableOnWords` typo settings
 */
 
 use crate::index::tests::TempIndex;
@@ -56,6 +57,10 @@ fn create_index() -> TempIndex {
             {
                 "id": 5,
                 "text": "sunflowering is not a verb"
+            },
+            {
+                "id": 6,
+                "text": "xy z"
             }
         ]))
         .unwrap();
@@ -263,10 +268,11 @@ fn test_disable_split_words() {
     s.query("sunflower ");
     let SearchResult { documents_ids, .. } = s.execute().unwrap();
     // no document containing `sun flower`
-    insta::assert_snapshot!(format!("{documents_ids:?}"), @"[3]");
+    insta::assert_snapshot!(format!("{documents_ids:?}"), @"[1, 3]");
     let texts = collect_field_values(&index, &txn, "text", &documents_ids);
     insta::assert_debug_snapshot!(texts, @r###"
     [
+        "\"the sun flower is tall\"",
         "\"the sunflower is tall\"",
     ]
     "###);
@@ -307,10 +313,11 @@ fn test_3gram_no_split_words() {
     let SearchResult { documents_ids, .. } = s.execute().unwrap();
 
     // no document with `sun flower`
-    insta::assert_snapshot!(format!("{documents_ids:?}"), @"[2, 3, 5]");
+    insta::assert_snapshot!(format!("{documents_ids:?}"), @"[1, 2, 3, 5]");
     let texts = collect_field_values(&index, &txn, "text", &documents_ids);
     insta::assert_debug_snapshot!(texts, @r###"
     [
+        "\"the sun flower is tall\"",
         "\"the sunflowers are pretty\"",
         "\"the sunflower is tall\"",
         "\"sunflowering is not a verb\"",
@@ -369,3 +376,50 @@ fn test_no_ngram_phrases() {
     ]
     "###);
 }
+
+#[test]
+fn test_short_split_words() {
+    let index = create_index();
+    let txn = index.read_txn().unwrap();
+
+    let mut s = Search::new(&txn, &index);
+    s.terms_matching_strategy(TermsMatchingStrategy::All);
+    s.query("xyz");
+    let SearchResult { documents_ids, .. } = s.execute().unwrap();
+
+    insta::assert_snapshot!(format!("{documents_ids:?}"), @"[6]");
+    let texts = collect_field_values(&index, &txn, "text", &documents_ids);
+    insta::assert_debug_snapshot!(texts, @r###"
+    [
+        "\"xy z\"",
+    ]
+    "###);
+}
+
+#[test]
+fn test_split_words_never_disabled() {
+    let index = create_index();
+
+    index
+        .update_settings(|s| {
+            s.set_exact_words(["sunflower"].iter().map(ToString::to_string).collect());
+            s.set_exact_attributes(["text"].iter().map(ToString::to_string).collect());
+        })
+        .unwrap();
+
+    let txn = index.read_txn().unwrap();
+
+    let mut s = Search::new(&txn, &index);
+    s.terms_matching_strategy(TermsMatchingStrategy::All);
+    s.query("the sunflower is tall");
+    let SearchResult { documents_ids, .. } = s.execute().unwrap();
+
+    insta::assert_snapshot!(format!("{documents_ids:?}"), @"[1, 3]");
+    let texts = collect_field_values(&index, &txn, "text", &documents_ids);
+    insta::assert_debug_snapshot!(texts, @r###"
+    [
+        "\"the sun flower is tall\"",
+        "\"the sunflower is tall\"",
+    ]
+    "###);
+}

milli/src/search/new/tests/typo.rs

@@ -9,7 +9,7 @@ This module tests the following properties:
 6. A typo on the first letter of a word counts as two typos
 7. Phrases are not typo tolerant
 8. 2grams can have 1 typo if they are larger than `min_word_len_two_typos`
-9. 3grams are not typo tolerant
+9. 3grams are not typo tolerant (but they can be split into two words)
 10. The `typo` ranking rule assumes the role of the `words` ranking rule implicitly
 if `words` doesn't exist before it.
 11. The `typo` ranking rule places documents with the same number of typos in the same bucket
@@ -287,16 +287,17 @@ fn test_typo_exact_word() {
     ]
     "###);
 
-    // exact words do not disable prefix (sunflowering OK, but no sunflowar or sun flower)
+    // exact words do not disable prefix (sunflowering OK, but no sunflowar)
     let mut s = Search::new(&txn, &index);
     s.terms_matching_strategy(TermsMatchingStrategy::All);
     s.query("network interconnection sunflower");
     let SearchResult { documents_ids, .. } = s.execute().unwrap();
-    insta::assert_snapshot!(format!("{documents_ids:?}"), @"[16, 18]");
+    insta::assert_snapshot!(format!("{documents_ids:?}"), @"[16, 17, 18]");
     let texts = collect_field_values(&index, &txn, "text", &documents_ids);
     insta::assert_debug_snapshot!(texts, @r###"
     [
         "\"network interconnection sunflower\"",
+        "\"network interconnection sun flower\"",
         "\"network interconnection sunflowering\"",
     ]
     "###);

milli/src/search/new/words.rs

@@ -1,87 +0,0 @@
-use roaring::RoaringBitmap;
-
-use super::logger::SearchLogger;
-use super::query_graph::QueryNode;
-use super::resolve_query_graph::compute_query_graph_docids;
-use super::small_bitmap::SmallBitmap;
-use super::{QueryGraph, RankingRule, RankingRuleOutput, SearchContext};
-use crate::{Result, TermsMatchingStrategy};
-
-pub struct Words {
-    exhausted: bool, // TODO: remove
-    query_graph: Option<QueryGraph>,
-    nodes_to_remove: Vec<SmallBitmap<QueryNode>>,
-    terms_matching_strategy: TermsMatchingStrategy,
-}
-impl Words {
-    pub fn new(terms_matching_strategy: TermsMatchingStrategy) -> Self {
-        Self {
-            exhausted: true,
-            query_graph: None,
-            nodes_to_remove: vec![],
-            terms_matching_strategy,
-        }
-    }
-}
-
-impl<'ctx> RankingRule<'ctx, QueryGraph> for Words {
-    fn id(&self) -> String {
-        "words".to_owned()
-    }
-    fn start_iteration(
-        &mut self,
-        ctx: &mut SearchContext<'ctx>,
-        _logger: &mut dyn SearchLogger<QueryGraph>,
-        _universe: &RoaringBitmap,
-        parent_query_graph: &QueryGraph,
-    ) -> Result<()> {
-        self.exhausted = false;
-        self.query_graph = Some(parent_query_graph.clone());
-        self.nodes_to_remove = match self.terms_matching_strategy {
-            TermsMatchingStrategy::Last => {
-                let mut ns = parent_query_graph.removal_order_for_terms_matching_strategy_last(ctx);
-                ns.reverse();
-                ns
-            }
-            TermsMatchingStrategy::All => {
-                vec![]
-            }
-        };
-        Ok(())
-    }
-
-    fn next_bucket(
-        &mut self,
-        ctx: &mut SearchContext<'ctx>,
-        logger: &mut dyn SearchLogger<QueryGraph>,
-        universe: &RoaringBitmap,
-    ) -> Result<Option<RankingRuleOutput<QueryGraph>>> {
-        if self.exhausted {
-            return Ok(None);
-        }
-        let Some(query_graph) = &mut self.query_graph else { panic!() };
-        logger.log_internal_state(query_graph);
-
-        let this_bucket = compute_query_graph_docids(ctx, query_graph, universe)?;
-
-        let child_query_graph = query_graph.clone();
-
-        if self.nodes_to_remove.is_empty() {
-            self.exhausted = true;
-        } else {
-            let nodes_to_remove = self.nodes_to_remove.pop().unwrap();
-            query_graph.remove_nodes_keep_edges(&nodes_to_remove.iter().collect::<Vec<_>>());
-        }
-        Ok(Some(RankingRuleOutput { query: child_query_graph, candidates: this_bucket }))
-    }
-
-    fn end_iteration(
-        &mut self,
-        _ctx: &mut SearchContext<'ctx>,
-        _logger: &mut dyn SearchLogger<QueryGraph>,
-    ) {
-        self.exhausted = true;
-        self.nodes_to_remove = vec![];
-        self.query_graph = None;
-    }
-}

milli/src/snapshot_tests.rs

@@ -89,7 +89,6 @@ Create a snapshot test of the given database.
     - `exact_word_docids`
     - `word_prefix_docids`
     - `exact_word_prefix_docids`
-    - `docid_word_positions`
     - `word_pair_proximity_docids`
     - `word_prefix_pair_proximity_docids`
     - `word_position_docids`
@@ -217,11 +216,6 @@ pub fn snap_exact_word_prefix_docids(index: &Index) -> String {
         &format!("{s:<16} {}", display_bitmap(&b))
     })
 }
-pub fn snap_docid_word_positions(index: &Index) -> String {
-    make_db_snap_from_iter!(index, docid_word_positions, |((idx, s), b)| {
-        &format!("{idx:<6} {s:<16} {}", display_bitmap(&b))
-    })
-}
 pub fn snap_word_pair_proximity_docids(index: &Index) -> String {
     make_db_snap_from_iter!(index, word_pair_proximity_docids, |((proximity, word1, word2), b)| {
         &format!("{proximity:<2} {word1:<16} {word2:<16} {}", display_bitmap(&b))
@@ -324,7 +318,7 @@ pub fn snap_field_distributions(index: &Index) -> String {
     let rtxn = index.read_txn().unwrap();
     let mut snap = String::new();
     for (field, count) in index.field_distribution(&rtxn).unwrap() {
-        writeln!(&mut snap, "{field:<16} {count:<6}").unwrap();
+        writeln!(&mut snap, "{field:<16} {count:<6} |").unwrap();
     }
     snap
 }
@@ -334,7 +328,7 @@ pub fn snap_fields_ids_map(index: &Index) -> String {
     let mut snap = String::new();
     for field_id in fields_ids_map.ids() {
         let name = fields_ids_map.name(field_id).unwrap();
-        writeln!(&mut snap, "{field_id:<3} {name:<16}").unwrap();
+        writeln!(&mut snap, "{field_id:<3} {name:<16} |").unwrap();
     }
     snap
 }
@@ -477,9 +471,6 @@ macro_rules! full_snap_of_db {
     ($index:ident, exact_word_prefix_docids) => {{
         $crate::snapshot_tests::snap_exact_word_prefix_docids(&$index)
     }};
-    ($index:ident, docid_word_positions) => {{
-        $crate::snapshot_tests::snap_docid_word_positions(&$index)
-    }};
     ($index:ident, word_pair_proximity_docids) => {{
         $crate::snapshot_tests::snap_word_pair_proximity_docids(&$index)
     }};

milli/src/snapshots/index.rs/initial_field_distribution/1/field_distribution.snap

@@ -1,7 +1,7 @@
 ---
 source: milli/src/index.rs
 ---
-age              1     
-id               2     
-name             2     
+age              1      |
+id               2      |
+name             2      |
 

milli/src/snapshots/index.rs/initial_field_distribution/field_distribution.snap

@@ -1,7 +1,7 @@
 ---
 source: milli/src/index.rs
 ---
-age              1     
-id               2     
-name             2     
+age              1      |
+id               2      |
+name             2      |
 

milli/src/update/clear_documents.rs

@@ -23,7 +23,6 @@ impl<'t, 'u, 'i> ClearDocuments<'t, 'u, 'i> {
             exact_word_docids,
             word_prefix_docids,
             exact_word_prefix_docids,
-            docid_word_positions,
             word_pair_proximity_docids,
             word_prefix_pair_proximity_docids,
             prefix_word_pair_proximity_docids,
@@ -40,6 +39,7 @@ impl<'t, 'u, 'i> ClearDocuments<'t, 'u, 'i> {
             facet_id_is_empty_docids,
             field_id_docid_facet_f64s,
             field_id_docid_facet_strings,
+            vector_id_docid,
             documents,
         } = self.index;
 
@@ -58,6 +58,7 @@ impl<'t, 'u, 'i> ClearDocuments<'t, 'u, 'i> {
         self.index.put_field_distribution(self.wtxn, &FieldDistribution::default())?;
         self.index.delete_geo_rtree(self.wtxn)?;
         self.index.delete_geo_faceted_documents_ids(self.wtxn)?;
+        self.index.delete_vector_hnsw(self.wtxn)?;
 
         // We clean all the faceted documents ids.
         for field_id in faceted_fields {
@@ -80,7 +81,6 @@ impl<'t, 'u, 'i> ClearDocuments<'t, 'u, 'i> {
         exact_word_docids.clear(self.wtxn)?;
         word_prefix_docids.clear(self.wtxn)?;
         exact_word_prefix_docids.clear(self.wtxn)?;
-        docid_word_positions.clear(self.wtxn)?;
         word_pair_proximity_docids.clear(self.wtxn)?;
         word_prefix_pair_proximity_docids.clear(self.wtxn)?;
         prefix_word_pair_proximity_docids.clear(self.wtxn)?;
@@ -97,6 +97,7 @@ impl<'t, 'u, 'i> ClearDocuments<'t, 'u, 'i> {
         facet_id_string_docids.clear(self.wtxn)?;
         field_id_docid_facet_f64s.clear(self.wtxn)?;
         field_id_docid_facet_strings.clear(self.wtxn)?;
+        vector_id_docid.clear(self.wtxn)?;
         documents.clear(self.wtxn)?;
 
         Ok(number_of_documents)
@@ -141,7 +142,6 @@ mod tests {
 
         assert!(index.word_docids.is_empty(&rtxn).unwrap());
         assert!(index.word_prefix_docids.is_empty(&rtxn).unwrap());
-        assert!(index.docid_word_positions.is_empty(&rtxn).unwrap());
         assert!(index.word_pair_proximity_docids.is_empty(&rtxn).unwrap());
         assert!(index.field_id_word_count_docids.is_empty(&rtxn).unwrap());
         assert!(index.word_prefix_pair_proximity_docids.is_empty(&rtxn).unwrap());

milli/src/update/delete_documents.rs

@@ -1,11 +1,13 @@
 use std::collections::btree_map::Entry;
-use std::collections::{HashMap, HashSet};
+use std::collections::{BTreeSet, HashMap, HashSet};
 
 use fst::IntoStreamer;
 use heed::types::{ByteSlice, DecodeIgnore, Str, UnalignedSlice};
 use heed::{BytesDecode, BytesEncode, Database, RwIter};
+use hnsw::Searcher;
 use roaring::RoaringBitmap;
 use serde::{Deserialize, Serialize};
+use space::KnnPoints;
 use time::OffsetDateTime;
 
 use super::facet::delete::FacetsDelete;
@@ -14,9 +16,9 @@ use crate::error::InternalError;
 use crate::facet::FacetType;
 use crate::heed_codec::facet::FieldDocIdFacetCodec;
 use crate::heed_codec::CboRoaringBitmapCodec;
+use crate::index::Hnsw;
 use crate::{
-    ExternalDocumentsIds, FieldId, FieldIdMapMissingEntry, Index, Result, RoaringBitmapCodec,
-    SmallString32, BEU32,
+    ExternalDocumentsIds, FieldId, FieldIdMapMissingEntry, Index, Result, RoaringBitmapCodec, BEU32,
 };
 
 pub struct DeleteDocuments<'t, 'u, 'i> {
@@ -232,7 +234,6 @@ impl<'t, 'u, 'i> DeleteDocuments<'t, 'u, 'i> {
             exact_word_docids,
             word_prefix_docids,
             exact_word_prefix_docids,
-            docid_word_positions,
             word_pair_proximity_docids,
             field_id_word_count_docids,
             word_prefix_pair_proximity_docids,
@@ -249,25 +250,12 @@ impl<'t, 'u, 'i> DeleteDocuments<'t, 'u, 'i> {
             facet_id_exists_docids,
             facet_id_is_null_docids,
             facet_id_is_empty_docids,
+            vector_id_docid,
             documents,
         } = self.index;
-
-        // Retrieve the words contained in the documents.
-        let mut words = Vec::new();
+        // Remove from the documents database
         for docid in &self.to_delete_docids {
             documents.delete(self.wtxn, &BEU32::new(docid))?;
-
-            // We iterate through the words positions of the document id, retrieve the word and delete the positions.
-            // We create an iterator to be able to get the content and delete the key-value itself.
-            // It's faster to acquire a cursor to get and delete, as we avoid traversing the LMDB B-Tree two times but only once.
-            let mut iter = docid_word_positions.prefix_iter_mut(self.wtxn, &(docid, ""))?;
-            while let Some(result) = iter.next() {
-                let ((_docid, word), _positions) = result?;
-                // This boolean will indicate if we must remove this word from the words FST.
-                words.push((SmallString32::from(word), false));
-                // safety: we don't keep references from inside the LMDB database.
-                unsafe { iter.del_current()? };
-            }
         }
         // We acquire the current external documents ids map...
         // Note that its soft-deleted document ids field will be equal to the `to_delete_docids`
@@ -278,42 +266,27 @@ impl<'t, 'u, 'i> DeleteDocuments<'t, 'u, 'i> {
         let new_external_documents_ids = new_external_documents_ids.into_static();
         self.index.put_external_documents_ids(self.wtxn, &new_external_documents_ids)?;
 
-        // Maybe we can improve the get performance of the words
-        // if we sort the words first, keeping the LMDB pages in cache.
-        words.sort_unstable();
-
+        let mut words_to_keep = BTreeSet::default();
+        let mut words_to_delete = BTreeSet::default();
         // We iterate over the words and delete the documents ids
         // from the word docids database.
-        for (word, must_remove) in &mut words {
-            remove_from_word_docids(
-                self.wtxn,
-                word_docids,
-                word.as_str(),
-                must_remove,
-                &self.to_delete_docids,
-            )?;
-
-            remove_from_word_docids(
-                self.wtxn,
-                exact_word_docids,
-                word.as_str(),
-                must_remove,
-                &self.to_delete_docids,
-            )?;
-        }
+        remove_from_word_docids(
+            self.wtxn,
+            word_docids,
+            &self.to_delete_docids,
+            &mut words_to_keep,
+            &mut words_to_delete,
+        )?;
+        remove_from_word_docids(
+            self.wtxn,
+            exact_word_docids,
+            &self.to_delete_docids,
+            &mut words_to_keep,
+            &mut words_to_delete,
+        )?;
 
         // We construct an FST set that contains the words to delete from the words FST.
-        let words_to_delete =
-            words.iter().filter_map(
-                |(word, must_remove)| {
-                    if *must_remove {
-                        Some(word.as_str())
-                    } else {
-                        None
-                    }
-                },
-            );
-        let words_to_delete = fst::Set::from_iter(words_to_delete)?;
+        let words_to_delete = fst::Set::from_iter(words_to_delete.difference(&words_to_keep))?;
 
         let new_words_fst = {
             // We retrieve the current words FST from the database.
@@ -467,6 +440,30 @@ impl<'t, 'u, 'i> DeleteDocuments<'t, 'u, 'i> {
             &self.to_delete_docids,
         )?;
 
+        // An ugly and slow way to remove the vectors from the HNSW
+        // It basically reconstructs the HNSW from scratch without editing the current one.
+        let current_hnsw = self.index.vector_hnsw(self.wtxn)?.unwrap_or_default();
+        if !current_hnsw.is_empty() {
+            let mut new_hnsw = Hnsw::default();
+            let mut searcher = Searcher::new();
+            let mut new_vector_id_docids = Vec::new();
+
+            for result in vector_id_docid.iter(self.wtxn)? {
+                let (vector_id, docid) = result?;
+                if !self.to_delete_docids.contains(docid.get()) {
+                    let vector = current_hnsw.get_point(vector_id.get() as usize).clone();
+                    let vector_id = new_hnsw.insert(vector, &mut searcher);
+                    new_vector_id_docids.push((vector_id as u32, docid));
+                }
+            }
+
+            vector_id_docid.clear(self.wtxn)?;
+            for (vector_id, docid) in new_vector_id_docids {
+                vector_id_docid.put(self.wtxn, &BEU32::new(vector_id), &docid)?;
+            }
+            self.index.put_vector_hnsw(self.wtxn, &new_hnsw)?;
+        }
+
         self.index.put_soft_deleted_documents_ids(self.wtxn, &RoaringBitmap::new())?;
 
         Ok(DetailedDocumentDeletionResult {
@@ -532,23 +529,24 @@ fn remove_from_word_prefix_docids(
 fn remove_from_word_docids(
     txn: &mut heed::RwTxn,
     db: &heed::Database<Str, RoaringBitmapCodec>,
-    word: &str,
-    must_remove: &mut bool,
     to_remove: &RoaringBitmap,
+    words_to_keep: &mut BTreeSet<String>,
+    words_to_remove: &mut BTreeSet<String>,
 ) -> Result<()> {
     // We create an iterator to be able to get the content and delete the word docids.
     // It's faster to acquire a cursor to get and delete or put, as we avoid traversing
     // the LMDB B-Tree two times but only once.
-    let mut iter = db.prefix_iter_mut(txn, word)?;
-    if let Some((key, mut docids)) = iter.next().transpose()? {
-        if key == word {
-            let previous_len = docids.len();
-            docids -= to_remove;
-            if docids.is_empty() {
-                // safety: we don't keep references from inside the LMDB database.
-                unsafe { iter.del_current()? };
-                *must_remove = true;
-            } else if docids.len() != previous_len {
+    let mut iter = db.iter_mut(txn)?;
+    while let Some((key, mut docids)) = iter.next().transpose()? {
+        let previous_len = docids.len();
+        docids -= to_remove;
+        if docids.is_empty() {
+            // safety: we don't keep references from inside the LMDB database.
+            unsafe { iter.del_current()? };
+            words_to_remove.insert(key.to_owned());
+        } else {
+            words_to_keep.insert(key.to_owned());
+            if docids.len() != previous_len {
                 let key = key.to_owned();
                 // safety: we don't keep references from inside the LMDB database.
                 unsafe { iter.put_current(&key, &docids)? };
@@ -627,7 +625,7 @@ mod tests {
 
     use super::*;
     use crate::index::tests::TempIndex;
-    use crate::{db_snap, Filter};
+    use crate::{db_snap, Filter, Search};
 
     fn delete_documents<'t>(
         wtxn: &mut RwTxn<'t, '_>,
@@ -1199,4 +1197,52 @@ mod tests {
             DeletionStrategy::AlwaysSoft,
         );
     }
+
+    #[test]
+    fn delete_words_exact_attributes() {
+        let index = TempIndex::new();
+
+        index
+            .update_settings(|settings| {
+                settings.set_primary_key(S("id"));
+                settings.set_searchable_fields(vec![S("text"), S("exact")]);
+                settings.set_exact_attributes(vec![S("exact")].into_iter().collect());
+            })
+            .unwrap();
+
+        index
+            .add_documents(documents!([
+                { "id": 0, "text": "hello" },
+                { "id": 1, "exact": "hello"}
+            ]))
+            .unwrap();
+        db_snap!(index, word_docids, 1, @r###"
+        hello            [0, ]
+        "###);
+        db_snap!(index, exact_word_docids, 1, @r###"
+        hello            [1, ]
+        "###);
+        db_snap!(index, words_fst, 1, @"300000000000000001084cfcfc2ce1000000016000000090ea47f");
+
+        let mut wtxn = index.write_txn().unwrap();
+        let deleted_internal_ids =
+            delete_documents(&mut wtxn, &index, &["1"], DeletionStrategy::AlwaysHard);
+        wtxn.commit().unwrap();
+
+        db_snap!(index, word_docids, 2, @r###"
+        hello            [0, ]
+        "###);
+        db_snap!(index, exact_word_docids, 2, @"");
+        db_snap!(index, words_fst, 2, @"300000000000000001084cfcfc2ce1000000016000000090ea47f");
+
+        insta::assert_snapshot!(format!("{deleted_internal_ids:?}"), @"[1]");
+        let txn = index.read_txn().unwrap();
+        let words = index.words_fst(&txn).unwrap().into_stream().into_strs().unwrap();
+        insta::assert_snapshot!(format!("{words:?}"), @r###"["hello"]"###);
+
+        let mut s = Search::new(&txn, &index);
+        s.query("hello");
+        let crate::SearchResult { documents_ids, .. } = s.execute().unwrap();
+        insta::assert_snapshot!(format!("{documents_ids:?}"), @"[0]");
+    }
 }

milli/src/update/facet/mod.rs

@@ -261,7 +261,9 @@ pub(crate) mod test_helpers {
             let options = options.map_size(4096 * 4 * 1000 * 100);
             let tempdir = tempfile::TempDir::new().unwrap();
             let env = options.open(tempdir.path()).unwrap();
-            let content = env.create_database(None).unwrap();
+            let mut wtxn = env.write_txn().unwrap();
+            let content = env.create_database(&mut wtxn, None).unwrap();
+            wtxn.commit().unwrap();
 
             FacetIndex {
                 content,

milli/src/update/index_documents/extract/extract_fid_word_count_docids.rs

@@ -1,6 +1,6 @@
 use std::collections::HashMap;
 use std::fs::File;
-use std::{cmp, io};
+use std::io;
 
 use grenad::Sorter;
 
@@ -54,11 +54,10 @@ pub fn extract_fid_word_count_docids<R: io::Read + io::Seek>(
         }
 
         for position in read_u32_ne_bytes(value) {
-            let (field_id, position) = relative_from_absolute_position(position);
-            let word_count = position as u32 + 1;
+            let (field_id, _) = relative_from_absolute_position(position);
 
             let value = document_fid_wordcount.entry(field_id as FieldId).or_insert(0);
-            *value = cmp::max(*value, word_count);
+            *value += 1;
         }
     }
 
@@ -83,7 +82,7 @@ fn drain_document_fid_wordcount_into_sorter(
     let mut key_buffer = Vec::new();
 
     for (fid, count) in document_fid_wordcount.drain() {
-        if count <= 10 {
+        if count <= 30 {
             key_buffer.clear();
             key_buffer.extend_from_slice(&fid.to_be_bytes());
             key_buffer.push(count as u8);

milli/src/update/index_documents/extract/extract_vector_points.rs

@@ -0,0 +1,69 @@
+use std::convert::TryFrom;
+use std::fs::File;
+use std::io;
+
+use bytemuck::cast_slice;
+use serde_json::{from_slice, Value};
+
+use super::helpers::{create_writer, writer_into_reader, GrenadParameters};
+use crate::error::UserError;
+use crate::{FieldId, InternalError, Result, VectorOrArrayOfVectors};
+
+/// Extracts the embedding vector contained in each document under the `_vectors` field.
+///
+/// Returns the generated grenad reader containing the docid as key associated to the Vec<f32>
+#[logging_timer::time]
+pub fn extract_vector_points<R: io::Read + io::Seek>(
+    obkv_documents: grenad::Reader<R>,
+    indexer: GrenadParameters,
+    primary_key_id: FieldId,
+    vectors_fid: FieldId,
+) -> Result<grenad::Reader<File>> {
+    let mut writer = create_writer(
+        indexer.chunk_compression_type,
+        indexer.chunk_compression_level,
+        tempfile::tempfile()?,
+    );
+
+    let mut cursor = obkv_documents.into_cursor()?;
+    while let Some((docid_bytes, value)) = cursor.move_on_next()? {
+        let obkv = obkv::KvReader::new(value);
+
+        // since we only needs the primary key when we throw an error we create this getter to
+        // lazily get it when needed
+        let document_id = || -> Value {
+            let document_id = obkv.get(primary_key_id).unwrap();
+            serde_json::from_slice(document_id).unwrap()
+        };
+
+        // first we retrieve the _vectors field
+        if let Some(vectors) = obkv.get(vectors_fid) {
+            // extract the vectors
+            let vectors = match from_slice(vectors) {
+                Ok(vectors) => VectorOrArrayOfVectors::into_array_of_vectors(vectors),
+                Err(_) => {
+                    return Err(UserError::InvalidVectorsType {
+                        document_id: document_id(),
+                        value: from_slice(vectors).map_err(InternalError::SerdeJson)?,
+                    }
+                    .into())
+                }
+            };
+
+            for (i, vector) in vectors.into_iter().enumerate() {
+                match u16::try_from(i) {
+                    Ok(i) => {
+                        let mut key = docid_bytes.to_vec();
+                        key.extend_from_slice(&i.to_ne_bytes());
+                        let bytes = cast_slice(&vector);
+                        writer.insert(key, bytes)?;
+                    }
+                    Err(_) => continue,
+                }
+            }
+        }
+        // else => the `_vectors` object was `null`, there is nothing to do
+    }
+
+    writer_into_reader(writer)
+}

milli/src/update/index_documents/extract/mod.rs

@@ -4,6 +4,7 @@ mod extract_facet_string_docids;
 mod extract_fid_docid_facet_values;
 mod extract_fid_word_count_docids;
 mod extract_geo_points;
+mod extract_vector_points;
 mod extract_word_docids;
 mod extract_word_fid_docids;
 mod extract_word_pair_proximity_docids;
@@ -22,6 +23,7 @@ use self::extract_facet_string_docids::extract_facet_string_docids;
 use self::extract_fid_docid_facet_values::{extract_fid_docid_facet_values, ExtractedFacetValues};
 use self::extract_fid_word_count_docids::extract_fid_word_count_docids;
 use self::extract_geo_points::extract_geo_points;
+use self::extract_vector_points::extract_vector_points;
 use self::extract_word_docids::extract_word_docids;
 use self::extract_word_fid_docids::extract_word_fid_docids;
 use self::extract_word_pair_proximity_docids::extract_word_pair_proximity_docids;
@@ -45,6 +47,7 @@ pub(crate) fn data_from_obkv_documents(
     faceted_fields: HashSet<FieldId>,
     primary_key_id: FieldId,
     geo_fields_ids: Option<(FieldId, FieldId)>,
+    vectors_field_id: Option<FieldId>,
     stop_words: Option<fst::Set<&[u8]>>,
     max_positions_per_attributes: Option<u32>,
     exact_attributes: HashSet<FieldId>,
@@ -69,6 +72,7 @@ pub(crate) fn data_from_obkv_documents(
                     &faceted_fields,
                     primary_key_id,
                     geo_fields_ids,
+                    vectors_field_id,
                     &stop_words,
                     max_positions_per_attributes,
                 )
@@ -279,6 +283,7 @@ fn send_and_extract_flattened_documents_data(
     faceted_fields: &HashSet<FieldId>,
     primary_key_id: FieldId,
     geo_fields_ids: Option<(FieldId, FieldId)>,
+    vectors_field_id: Option<FieldId>,
     stop_words: &Option<fst::Set<&[u8]>>,
     max_positions_per_attributes: Option<u32>,
 ) -> Result<(
@@ -307,6 +312,25 @@ fn send_and_extract_flattened_documents_data(
         });
     }
 
+    if let Some(vectors_field_id) = vectors_field_id {
+        let documents_chunk_cloned = flattened_documents_chunk.clone();
+        let lmdb_writer_sx_cloned = lmdb_writer_sx.clone();
+        rayon::spawn(move || {
+            let result = extract_vector_points(
+                documents_chunk_cloned,
+                indexer,
+                primary_key_id,
+                vectors_field_id,
+            );
+            let _ = match result {
+                Ok(vector_points) => {
+                    lmdb_writer_sx_cloned.send(Ok(TypedChunk::VectorPoints(vector_points)))
+                }
+                Err(error) => lmdb_writer_sx_cloned.send(Err(error)),
+            };
+        });
+    }
+
     let (docid_word_positions_chunk, docid_fid_facet_values_chunks): (Result<_>, Result<_>) =
         rayon::join(
             || {
@@ -325,8 +349,6 @@ fn send_and_extract_flattened_documents_data(
                 // send docid_word_positions_chunk to DB writer
                 let docid_word_positions_chunk =
                     unsafe { as_cloneable_grenad(&docid_word_positions_chunk)? };
-                let _ = lmdb_writer_sx
-                    .send(Ok(TypedChunk::DocidWordPositions(docid_word_positions_chunk.clone())));
 
                 let _ =
                     lmdb_writer_sx.send(Ok(TypedChunk::ScriptLanguageDocids(script_language_pair)));

milli/src/update/index_documents/helpers/merge_functions.rs

@@ -4,7 +4,6 @@ use std::result::Result as StdResult;
 
 use roaring::RoaringBitmap;
 
-use super::read_u32_ne_bytes;
 use crate::heed_codec::CboRoaringBitmapCodec;
 use crate::update::index_documents::transform::Operation;
 use crate::Result;
@@ -22,10 +21,6 @@ pub fn concat_u32s_array<'a>(_key: &[u8], values: &[Cow<'a, [u8]>]) -> Result<Co
     }
 }
 
-pub fn roaring_bitmap_from_u32s_array(slice: &[u8]) -> RoaringBitmap {
-    read_u32_ne_bytes(slice).collect()
-}
-
 pub fn serialize_roaring_bitmap(bitmap: &RoaringBitmap, buffer: &mut Vec<u8>) -> io::Result<()> {
     buffer.clear();
     buffer.reserve(bitmap.serialized_size());

milli/src/update/index_documents/helpers/mod.rs

@@ -14,8 +14,8 @@ pub use grenad_helpers::{
 };
 pub use merge_functions::{
     concat_u32s_array, keep_first, keep_latest_obkv, merge_cbo_roaring_bitmaps,
-    merge_obkvs_and_operations, merge_roaring_bitmaps, merge_two_obkvs,
-    roaring_bitmap_from_u32s_array, serialize_roaring_bitmap, MergeFn,
+    merge_obkvs_and_operations, merge_roaring_bitmaps, merge_two_obkvs, serialize_roaring_bitmap,
+    MergeFn,
 };
 
 use crate::MAX_WORD_LENGTH;

milli/src/update/index_documents/mod.rs

@@ -304,6 +304,8 @@ where
             }
             None => None,
         };
+        // get the fid of the `_vectors` field.
+        let vectors_field_id = self.index.fields_ids_map(self.wtxn)?.id("_vectors");
 
         let stop_words = self.index.stop_words(self.wtxn)?;
         let exact_attributes = self.index.exact_attributes_ids(self.wtxn)?;
@@ -340,6 +342,7 @@ where
                     faceted_fields,
                     primary_key_id,
                     geo_fields_ids,
+                    vectors_field_id,
                     stop_words,
                     max_positions_per_attributes,
                     exact_attributes,
@@ -2471,11 +2474,11 @@ mod tests {
               {
                 "id": 3,
                 "text": "a a a a a a a a a a a a a a a a a
-                a a a a a a a a a a a a a a a a a a a a a a a a a a 
-                a a a a a a a a a a a a a a a a a a a a a a a a a a 
-                a a a a a a a a a a a a a a a a a a a a a a a a a a 
-                a a a a a a a a a a a a a a a a a a a a a a a a a a 
-                a a a a a a a a a a a a a a a a a a a a a a a a a a 
+                a a a a a a a a a a a a a a a a a a a a a a a a a a
+                a a a a a a a a a a a a a a a a a a a a a a a a a a
+                a a a a a a a a a a a a a a a a a a a a a a a a a a
+                a a a a a a a a a a a a a a a a a a a a a a a a a a
+                a a a a a a a a a a a a a a a a a a a a a a a a a a
                 a a a a a a a a a a a a a a a a a a a a a "
              }
             ]))
@@ -2513,6 +2516,5 @@ mod tests {
 
         db_snap!(index, word_fid_docids, 3, @"4c2e2a1832e5802796edc1638136d933");
         db_snap!(index, word_position_docids, 3, @"74f556b91d161d997a89468b4da1cb8f");
-        db_snap!(index, docid_word_positions, 3, @"5287245332627675740b28bd46e1cde1");
     }
 }

milli/src/update/index_documents/typed_chunk.rs

@@ -4,27 +4,29 @@ use std::convert::TryInto;
 use std::fs::File;
 use std::io;
 
+use bytemuck::allocation::pod_collect_to_vec;
 use charabia::{Language, Script};
 use grenad::MergerBuilder;
 use heed::types::ByteSlice;
-use heed::{BytesDecode, RwTxn};
+use heed::RwTxn;
+use hnsw::Searcher;
 use roaring::RoaringBitmap;
+use space::KnnPoints;
 
 use super::helpers::{
-    self, merge_ignore_values, roaring_bitmap_from_u32s_array, serialize_roaring_bitmap,
-    valid_lmdb_key, CursorClonableMmap,
+    self, merge_ignore_values, serialize_roaring_bitmap, valid_lmdb_key, CursorClonableMmap,
 };
 use super::{ClonableMmap, MergeFn};
+use crate::error::UserError;
 use crate::facet::FacetType;
 use crate::update::facet::FacetsUpdate;
-use crate::update::index_documents::helpers::as_cloneable_grenad;
+use crate::update::index_documents::helpers::{as_cloneable_grenad, try_split_array_at};
 use crate::{
-    lat_lng_to_xyz, BoRoaringBitmapCodec, CboRoaringBitmapCodec, DocumentId, GeoPoint, Index,
-    Result,
+    lat_lng_to_xyz, normalize_vector, CboRoaringBitmapCodec, DocumentId, GeoPoint, Index, Result,
+    BEU32,
 };
 
 pub(crate) enum TypedChunk {
-    DocidWordPositions(grenad::Reader<CursorClonableMmap>),
     FieldIdDocidFacetStrings(grenad::Reader<CursorClonableMmap>),
     FieldIdDocidFacetNumbers(grenad::Reader<CursorClonableMmap>),
     Documents(grenad::Reader<CursorClonableMmap>),
@@ -43,6 +45,7 @@ pub(crate) enum TypedChunk {
     FieldIdFacetIsNullDocids(grenad::Reader<File>),
     FieldIdFacetIsEmptyDocids(grenad::Reader<File>),
     GeoPoints(grenad::Reader<File>),
+    VectorPoints(grenad::Reader<File>),
     ScriptLanguageDocids(HashMap<(Script, Language), RoaringBitmap>),
 }
 
@@ -56,29 +59,6 @@ pub(crate) fn write_typed_chunk_into_index(
 ) -> Result<(RoaringBitmap, bool)> {
     let mut is_merged_database = false;
     match typed_chunk {
-        TypedChunk::DocidWordPositions(docid_word_positions_iter) => {
-            write_entries_into_database(
-                docid_word_positions_iter,
-                &index.docid_word_positions,
-                wtxn,
-                index_is_empty,
-                |value, buffer| {
-                    // ensure that values are unique and ordered
-                    let positions = roaring_bitmap_from_u32s_array(value);
-                    BoRoaringBitmapCodec::serialize_into(&positions, buffer);
-                    Ok(buffer)
-                },
-                |new_values, db_values, buffer| {
-                    let new_values = roaring_bitmap_from_u32s_array(new_values);
-                    let positions = match BoRoaringBitmapCodec::bytes_decode(db_values) {
-                        Some(db_values) => new_values | db_values,
-                        None => new_values, // should not happen
-                    };
-                    BoRoaringBitmapCodec::serialize_into(&positions, buffer);
-                    Ok(())
-                },
-            )?;
-        }
         TypedChunk::Documents(obkv_documents_iter) => {
             let mut cursor = obkv_documents_iter.into_cursor()?;
             while let Some((key, value)) = cursor.move_on_next()? {
@@ -249,6 +229,40 @@ pub(crate) fn write_typed_chunk_into_index(
             index.put_geo_rtree(wtxn, &rtree)?;
             index.put_geo_faceted_documents_ids(wtxn, &geo_faceted_docids)?;
         }
+        TypedChunk::VectorPoints(vector_points) => {
+            let mut hnsw = index.vector_hnsw(wtxn)?.unwrap_or_default();
+            let mut searcher = Searcher::new();
+
+            let mut expected_dimensions = match index.vector_id_docid.iter(wtxn)?.next() {
+                Some(result) => {
+                    let (vector_id, _) = result?;
+                    Some(hnsw.get_point(vector_id.get() as usize).len())
+                }
+                None => None,
+            };
+
+            let mut cursor = vector_points.into_cursor()?;
+            while let Some((key, value)) = cursor.move_on_next()? {
+                // convert the key back to a u32 (4 bytes)
+                let (left, _index) = try_split_array_at(key).unwrap();
+                let docid = DocumentId::from_be_bytes(left);
+                // convert the vector back to a Vec<f32>
+                let vector: Vec<f32> = pod_collect_to_vec(value);
+
+                // TODO Move this error in the vector extractor
+                let found = vector.len();
+                let expected = *expected_dimensions.get_or_insert(found);
+                if expected != found {
+                    return Err(UserError::InvalidVectorDimensions { expected, found })?;
+                }
+
+                let vector = normalize_vector(vector);
+                let vector_id = hnsw.insert(vector, &mut searcher) as u32;
+                index.vector_id_docid.put(wtxn, &BEU32::new(vector_id), &BEU32::new(docid))?;
+            }
+            log::debug!("There are {} entries in the HNSW so far", hnsw.len());
+            index.put_vector_hnsw(wtxn, &hnsw)?;
+        }
         TypedChunk::ScriptLanguageDocids(hash_pair) => {
             let mut buffer = Vec::new();
             for (key, value) in hash_pair {