Update the proximity precision for the settings delta

Co-authored-By: Maxime Legendre <maxime@meilisearch.com>

.github/workflows/test-suite.yml

@@ -154,8 +154,8 @@ jobs:
           apt-get update && apt-get install -y curl
           apt-get install build-essential -y
       - uses: dtolnay/rust-toolchain@1.89
-      - name: Cache dependencies
-        uses: Swatinem/rust-cache@v2.8.0
+      - name: Run df -h
+        run: df -h
       - name: Run tests in debug
         uses: actions-rs/cargo@v1
         with:

crates/milli/src/fields_ids_map/metadata.rs

@@ -18,6 +18,8 @@ use crate::{
 pub struct Metadata {
     /// The weight as defined in the FieldidsWeightsMap of the searchable attribute if it is searchable.
     pub searchable: Option<Weight>,
+    /// The field is part of the exact attributes.
+    pub exact: bool,
     /// The field is part of the sortable attributes.
     pub sortable: bool,
     /// The field is defined as the distinct attribute.
@@ -209,6 +211,7 @@ impl Metadata {
 #[derive(Debug, Clone)]
 pub struct MetadataBuilder {
     searchable_attributes: Option<Vec<String>>,
+    exact_searchable_attributes: Vec<String>,
     filterable_attributes: Vec<FilterableAttributesRule>,
     sortable_attributes: HashSet<String>,
     localized_attributes: Option<Vec<LocalizedAttributesRule>>,
@@ -220,15 +223,18 @@ impl MetadataBuilder {
     pub fn from_index(index: &Index, rtxn: &RoTxn) -> Result<Self> {
         let searchable_attributes = index
             .user_defined_searchable_fields(rtxn)?
-            .map(|fields| fields.into_iter().map(|s| s.to_string()).collect());
+            .map(|fields| fields.into_iter().map(String::from).collect());
+        let exact_searchable_attributes =
+            index.exact_attributes(rtxn)?.into_iter().map(String::from).collect();
         let filterable_attributes = index.filterable_attributes_rules(rtxn)?;
         let sortable_attributes = index.sortable_fields(rtxn)?;
         let localized_attributes = index.localized_attributes_rules(rtxn)?;
-        let distinct_attribute = index.distinct_field(rtxn)?.map(|s| s.to_string());
+        let distinct_attribute = index.distinct_field(rtxn)?.map(String::from);
         let asc_desc_attributes = index.asc_desc_fields(rtxn)?;
 
         Ok(Self::new(
             searchable_attributes,
+            exact_searchable_attributes,
             filterable_attributes,
             sortable_attributes,
             localized_attributes,
@@ -242,6 +248,7 @@ impl MetadataBuilder {
     /// This is used for testing, prefer using `MetadataBuilder::from_index` instead.
     pub fn new(
         searchable_attributes: Option<Vec<String>>,
+        exact_searchable_attributes: Vec<String>,
         filterable_attributes: Vec<FilterableAttributesRule>,
         sortable_attributes: HashSet<String>,
         localized_attributes: Option<Vec<LocalizedAttributesRule>>,
@@ -256,6 +263,7 @@ impl MetadataBuilder {
 
         Self {
             searchable_attributes,
+            exact_searchable_attributes,
             filterable_attributes,
             sortable_attributes,
             localized_attributes,
@@ -269,6 +277,7 @@ impl MetadataBuilder {
             // Vectors fields are not searchable, filterable, distinct or asc_desc
             return Metadata {
                 searchable: None,
+                exact: false,
                 sortable: false,
                 distinct: false,
                 asc_desc: false,
@@ -296,6 +305,7 @@ impl MetadataBuilder {
             // Geo fields are not searchable, distinct or asc_desc
             return Metadata {
                 searchable: None,
+                exact: false,
                 sortable,
                 distinct: false,
                 asc_desc: false,
@@ -309,6 +319,7 @@ impl MetadataBuilder {
             debug_assert!(!sortable, "geojson fields should not be sortable");
             return Metadata {
                 searchable: None,
+                exact: false,
                 sortable,
                 distinct: false,
                 asc_desc: false,
@@ -329,6 +340,8 @@ impl MetadataBuilder {
             None => Some(0),
         };
 
+        let exact = self.exact_searchable_attributes.iter().any(|attr| is_faceted_by(field, attr));
+
         let distinct =
             self.distinct_attribute.as_ref().is_some_and(|distinct_field| field == distinct_field);
         let asc_desc = self.asc_desc_attributes.contains(field);
@@ -343,6 +356,7 @@ impl MetadataBuilder {
 
         Metadata {
             searchable,
+            exact,
             sortable,
             distinct,
             asc_desc,

crates/milli/src/update/new/extract/searchable/extract_word_docids.rs

@@ -8,17 +8,26 @@ use bumpalo::Bump;
 
 use super::match_searchable_field;
 use super::tokenize_document::{tokenizer_builder, DocumentTokenizer};
+use crate::fields_ids_map::metadata::Metadata;
 use crate::update::new::document::DocumentContext;
 use crate::update::new::extract::cache::BalancedCaches;
 use crate::update::new::extract::perm_json_p::contained_in;
+use crate::update::new::extract::searchable::has_searchable_children;
 use crate::update::new::indexer::document_changes::{
     extract, DocumentChanges, Extractor, IndexingContext,
 };
+use crate::update::new::indexer::settings_changes::{
+    settings_change_extract, DocumentsIndentifiers, SettingsChangeExtractor,
+};
 use crate::update::new::ref_cell_ext::RefCellExt as _;
 use crate::update::new::steps::IndexingStep;
 use crate::update::new::thread_local::{FullySend, MostlySend, ThreadLocal};
-use crate::update::new::DocumentChange;
-use crate::{bucketed_position, DocumentId, FieldId, Result, MAX_POSITION_PER_ATTRIBUTE};
+use crate::update::new::{DocumentChange, DocumentIdentifiers};
+use crate::update::settings::SettingsDelta;
+use crate::{
+    bucketed_position, DocumentId, FieldId, PatternMatch, Result, UserError,
+    MAX_POSITION_PER_ATTRIBUTE,
+};
 
 const MAX_COUNTED_WORDS: usize = 30;
 
@@ -34,6 +43,15 @@ pub struct WordDocidsBalancedCaches<'extractor> {
 
 unsafe impl MostlySend for WordDocidsBalancedCaches<'_> {}
 
+/// Whether to extract or skip fields during word extraction.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+enum FieldDbExtraction {
+    /// Extract the word and put it in to the fid-based databases.
+    Extract,
+    /// Do not store the word in the fid-based databases.
+    Skip,
+}
+
 impl<'extractor> WordDocidsBalancedCaches<'extractor> {
     pub fn new_in(buckets: usize, max_memory: Option<usize>, alloc: &'extractor Bump) -> Self {
         Self {
@@ -47,12 +65,14 @@ impl<'extractor> WordDocidsBalancedCaches<'extractor> {
         }
     }
 
+    #[allow(clippy::too_many_arguments)]
     fn insert_add_u32(
         &mut self,
         field_id: FieldId,
         position: u16,
         word: &str,
         exact: bool,
+        field_db_extraction: FieldDbExtraction,
         docid: u32,
         bump: &Bump,
     ) -> Result<()> {
@@ -66,11 +86,13 @@ impl<'extractor> WordDocidsBalancedCaches<'extractor> {
         let buffer_size = word_bytes.len() + 1 + size_of::<FieldId>();
         let mut buffer = BumpVec::with_capacity_in(buffer_size, bump);
 
-        buffer.clear();
-        buffer.extend_from_slice(word_bytes);
-        buffer.push(0);
-        buffer.extend_from_slice(&field_id.to_be_bytes());
-        self.word_fid_docids.insert_add_u32(&buffer, docid)?;
+        if field_db_extraction == FieldDbExtraction::Extract {
+            buffer.clear();
+            buffer.extend_from_slice(word_bytes);
+            buffer.push(0);
+            buffer.extend_from_slice(&field_id.to_be_bytes());
+            self.word_fid_docids.insert_add_u32(&buffer, docid)?;
+        }
 
         let position = bucketed_position(position);
         buffer.clear();
@@ -83,21 +105,26 @@ impl<'extractor> WordDocidsBalancedCaches<'extractor> {
             self.flush_fid_word_count(&mut buffer)?;
         }
 
-        self.fid_word_count
-            .entry(field_id)
-            .and_modify(|(_current_count, new_count)| *new_count.get_or_insert(0) += 1)
-            .or_insert((None, Some(1)));
+        if field_db_extraction == FieldDbExtraction::Extract {
+            self.fid_word_count
+                .entry(field_id)
+                .and_modify(|(_current_count, new_count)| *new_count.get_or_insert(0) += 1)
+                .or_insert((None, Some(1)));
+        }
+
         self.current_docid = Some(docid);
 
         Ok(())
     }
 
+    #[allow(clippy::too_many_arguments)]
     fn insert_del_u32(
         &mut self,
         field_id: FieldId,
         position: u16,
         word: &str,
         exact: bool,
+        field_db_extraction: FieldDbExtraction,
         docid: u32,
         bump: &Bump,
     ) -> Result<()> {
@@ -111,11 +138,13 @@ impl<'extractor> WordDocidsBalancedCaches<'extractor> {
         let buffer_size = word_bytes.len() + 1 + size_of::<FieldId>();
         let mut buffer = BumpVec::with_capacity_in(buffer_size, bump);
 
-        buffer.clear();
-        buffer.extend_from_slice(word_bytes);
-        buffer.push(0);
-        buffer.extend_from_slice(&field_id.to_be_bytes());
-        self.word_fid_docids.insert_del_u32(&buffer, docid)?;
+        if field_db_extraction == FieldDbExtraction::Extract {
+            buffer.clear();
+            buffer.extend_from_slice(word_bytes);
+            buffer.push(0);
+            buffer.extend_from_slice(&field_id.to_be_bytes());
+            self.word_fid_docids.insert_del_u32(&buffer, docid)?;
+        }
 
         let position = bucketed_position(position);
         buffer.clear();
@@ -128,10 +157,12 @@ impl<'extractor> WordDocidsBalancedCaches<'extractor> {
             self.flush_fid_word_count(&mut buffer)?;
         }
 
-        self.fid_word_count
-            .entry(field_id)
-            .and_modify(|(current_count, _new_count)| *current_count.get_or_insert(0) += 1)
-            .or_insert((Some(1), None));
+        if field_db_extraction == FieldDbExtraction::Extract {
+            self.fid_word_count
+                .entry(field_id)
+                .and_modify(|(current_count, _new_count)| *current_count.get_or_insert(0) += 1)
+                .or_insert((Some(1), None));
+        }
 
         self.current_docid = Some(docid);
 
@@ -325,6 +356,24 @@ impl WordDocidsExtractors {
             exact_attributes.iter().any(|attr| contained_in(fname, attr))
                 || disabled_typos_terms.is_exact(word)
         };
+
+        let mut should_tokenize = |field_name: &str| {
+            let Some((field_id, meta)) = new_fields_ids_map.id_with_metadata_or_insert(field_name)
+            else {
+                return Err(UserError::AttributeLimitReached.into());
+            };
+
+            let pattern_match = if meta.is_searchable() {
+                PatternMatch::Match
+            } else {
+                // TODO: should be a match on the field_name using `match_field_legacy` function,
+                //       but for legacy reasons we iterate over all the fields to fill the field_id_map.
+                PatternMatch::Parent
+            };
+
+            Ok((field_id, pattern_match))
+        };
+
         match document_change {
             DocumentChange::Deletion(inner) => {
                 let mut token_fn = |fname: &str, fid, pos, word: &str| {
@@ -333,13 +382,14 @@ impl WordDocidsExtractors {
                         pos,
                         word,
                         is_exact(fname, word),
+                        FieldDbExtraction::Extract,
                         inner.docid(),
                         doc_alloc,
                     )
                 };
                 document_tokenizer.tokenize_document(
                     inner.current(rtxn, index, context.db_fields_ids_map)?,
-                    new_fields_ids_map,
+                    &mut should_tokenize,
                     &mut token_fn,
                 )?;
             }
@@ -361,13 +411,14 @@ impl WordDocidsExtractors {
                         pos,
                         word,
                         is_exact(fname, word),
+                        FieldDbExtraction::Extract,
                         inner.docid(),
                         doc_alloc,
                     )
                 };
                 document_tokenizer.tokenize_document(
                     inner.current(rtxn, index, context.db_fields_ids_map)?,
-                    new_fields_ids_map,
+                    &mut should_tokenize,
                     &mut token_fn,
                 )?;
 
@@ -377,13 +428,14 @@ impl WordDocidsExtractors {
                         pos,
                         word,
                         is_exact(fname, word),
+                        FieldDbExtraction::Extract,
                         inner.docid(),
                         doc_alloc,
                     )
                 };
                 document_tokenizer.tokenize_document(
                     inner.merged(rtxn, index, context.db_fields_ids_map)?,
-                    new_fields_ids_map,
+                    &mut should_tokenize,
                     &mut token_fn,
                 )?;
             }
@@ -394,13 +446,14 @@ impl WordDocidsExtractors {
                         pos,
                         word,
                         is_exact(fname, word),
+                        FieldDbExtraction::Extract,
                         inner.docid(),
                         doc_alloc,
                     )
                 };
                 document_tokenizer.tokenize_document(
                     inner.inserted(),
-                    new_fields_ids_map,
+                    &mut should_tokenize,
                     &mut token_fn,
                 )?;
             }
@@ -411,3 +464,292 @@ impl WordDocidsExtractors {
         cached_sorter.flush_fid_word_count(&mut buffer)
     }
 }
+
+pub struct WordDocidsSettingsExtractorsData<'a, SD> {
+    tokenizer: DocumentTokenizer<'a>,
+    max_memory_by_thread: Option<usize>,
+    buckets: usize,
+    settings_delta: &'a SD,
+}
+
+impl<'extractor, SD: SettingsDelta + Sync> SettingsChangeExtractor<'extractor>
+    for WordDocidsSettingsExtractorsData<'_, SD>
+{
+    type Data = RefCell<Option<WordDocidsBalancedCaches<'extractor>>>;
+
+    fn init_data<'doc>(&'doc self, extractor_alloc: &'extractor Bump) -> crate::Result<Self::Data> {
+        Ok(RefCell::new(Some(WordDocidsBalancedCaches::new_in(
+            self.buckets,
+            self.max_memory_by_thread,
+            extractor_alloc,
+        ))))
+    }
+
+    fn process<'doc>(
+        &'doc self,
+        documents: impl Iterator<Item = crate::Result<DocumentIdentifiers<'doc>>>,
+        context: &'doc DocumentContext<Self::Data>,
+    ) -> crate::Result<()> {
+        for document in documents {
+            let document = document?;
+            SettingsChangeWordDocidsExtractors::extract_document_from_settings_change(
+                document,
+                context,
+                &self.tokenizer,
+                self.settings_delta,
+            )?;
+        }
+        Ok(())
+    }
+}
+
+pub struct SettingsChangeWordDocidsExtractors;
+
+impl SettingsChangeWordDocidsExtractors {
+    pub fn run_extraction<'fid, 'indexer, 'index, 'extractor, SD, MSP>(
+        settings_delta: &SD,
+        documents: &'indexer DocumentsIndentifiers<'indexer>,
+        indexing_context: IndexingContext<'fid, 'indexer, 'index, MSP>,
+        extractor_allocs: &'extractor mut ThreadLocal<FullySend<Bump>>,
+        step: IndexingStep,
+    ) -> Result<WordDocidsCaches<'extractor>>
+    where
+        SD: SettingsDelta + Sync,
+        MSP: Fn() -> bool + Sync,
+    {
+        // Warning: this is duplicated code from extract_word_pair_proximity_docids.rs
+        // TODO we need to read the new AND old settings to support changing global parameters
+        let rtxn = indexing_context.index.read_txn()?;
+        let stop_words = indexing_context.index.stop_words(&rtxn)?;
+        let allowed_separators = indexing_context.index.allowed_separators(&rtxn)?;
+        let allowed_separators: Option<Vec<_>> =
+            allowed_separators.as_ref().map(|s| s.iter().map(String::as_str).collect());
+        let dictionary = indexing_context.index.dictionary(&rtxn)?;
+        let dictionary: Option<Vec<_>> =
+            dictionary.as_ref().map(|s| s.iter().map(String::as_str).collect());
+        let mut builder = tokenizer_builder(
+            stop_words.as_ref(),
+            allowed_separators.as_deref(),
+            dictionary.as_deref(),
+        );
+        let tokenizer = builder.build();
+        let localized_attributes_rules =
+            indexing_context.index.localized_attributes_rules(&rtxn)?.unwrap_or_default();
+        let document_tokenizer = DocumentTokenizer {
+            tokenizer: &tokenizer,
+            localized_attributes_rules: &localized_attributes_rules,
+            max_positions_per_attributes: MAX_POSITION_PER_ATTRIBUTE,
+        };
+        let extractor_data = WordDocidsSettingsExtractorsData {
+            tokenizer: document_tokenizer,
+            max_memory_by_thread: indexing_context.grenad_parameters.max_memory_by_thread(),
+            buckets: rayon::current_num_threads(),
+            settings_delta,
+        };
+        let datastore = ThreadLocal::new();
+        {
+            let span = tracing::debug_span!(target: "indexing::documents::extract", "vectors");
+            let _entered = span.enter();
+
+            settings_change_extract(
+                documents,
+                &extractor_data,
+                indexing_context,
+                extractor_allocs,
+                &datastore,
+                step,
+            )?;
+        }
+
+        let mut merger = WordDocidsCaches::new();
+        for cache in datastore.into_iter().flat_map(RefCell::into_inner) {
+            merger.push(cache)?;
+        }
+
+        Ok(merger)
+    }
+
+    /// Extracts document words from a settings change.
+    fn extract_document_from_settings_change<SD: SettingsDelta>(
+        document: DocumentIdentifiers<'_>,
+        context: &DocumentContext<RefCell<Option<WordDocidsBalancedCaches>>>,
+        document_tokenizer: &DocumentTokenizer,
+        settings_delta: &SD,
+    ) -> Result<()> {
+        let mut cached_sorter_ref = context.data.borrow_mut_or_yield();
+        let cached_sorter = cached_sorter_ref.as_mut().unwrap();
+        let doc_alloc = &context.doc_alloc;
+
+        let new_fields_ids_map = settings_delta.new_fields_ids_map();
+        let old_fields_ids_map = context.index.fields_ids_map_with_metadata(&context.rtxn)?;
+        let old_searchable = settings_delta.old_searchable_attributes().as_ref();
+        let new_searchable = settings_delta.new_searchable_attributes().as_ref();
+
+        let current_document = document.current(
+            &context.rtxn,
+            context.index,
+            old_fields_ids_map.as_fields_ids_map(),
+        )?;
+
+        #[derive(Debug, Clone, Copy, PartialEq)]
+        enum ActionToOperate {
+            ReindexAllFields,
+            // TODO improve by listing field prefixes
+            IndexAddedFields,
+            SkipDocument,
+        }
+
+        let mut action = ActionToOperate::SkipDocument;
+        // Here we do a preliminary check to determine the action to take.
+        // This check doesn't trigger the tokenizer as we never return
+        // PatternMatch::Match.
+        document_tokenizer.tokenize_document(
+            current_document,
+            &mut |field_name| {
+                let fid = new_fields_ids_map.id(field_name).expect("All fields IDs must exist");
+
+                // If the document must be reindexed, early return NoMatch to stop the scanning process.
+                if action == ActionToOperate::ReindexAllFields {
+                    return Ok((fid, PatternMatch::NoMatch));
+                }
+
+                let old_field_metadata = old_fields_ids_map.metadata(fid).unwrap();
+                let new_field_metadata = new_fields_ids_map.metadata(fid).unwrap();
+
+                action = match (old_field_metadata, new_field_metadata) {
+                    // At least one field is added or removed from the exact fields => ReindexAllFields
+                    (Metadata { exact: old_exact, .. }, Metadata { exact: new_exact, .. })
+                        if old_exact != new_exact =>
+                    {
+                        ActionToOperate::ReindexAllFields
+                    }
+                    // At least one field is removed from the searchable fields => ReindexAllFields
+                    (Metadata { searchable: Some(_), .. }, Metadata { searchable: None, .. }) => {
+                        ActionToOperate::ReindexAllFields
+                    }
+                    // At least one field is added in the searchable fields => IndexAddedFields
+                    (Metadata { searchable: None, .. }, Metadata { searchable: Some(_), .. }) => {
+                        // We can safely overwrite the action, because we early return when action is ReindexAllFields.
+                        ActionToOperate::IndexAddedFields
+                    }
+                    _ => action,
+                };
+
+                Ok((fid, PatternMatch::Parent))
+            },
+            &mut |_, _, _, _| Ok(()),
+        )?;
+
+        // Early return when we don't need to index the document
+        if action == ActionToOperate::SkipDocument {
+            return Ok(());
+        }
+
+        let mut should_tokenize = |field_name: &str| {
+            let field_id = new_fields_ids_map.id(field_name).expect("All fields IDs must exist");
+            let old_field_metadata = old_fields_ids_map.metadata(field_id).unwrap();
+            let new_field_metadata = new_fields_ids_map.metadata(field_id).unwrap();
+
+            let pattern_match = match action {
+                ActionToOperate::ReindexAllFields => {
+                    if old_field_metadata.is_searchable() || new_field_metadata.is_searchable() {
+                        PatternMatch::Match
+                    // If any old or new field is searchable then we need to iterate over all fields
+                    // else if any field matches we need to iterate over all fields
+                    } else if has_searchable_children(
+                        field_name,
+                        old_searchable.zip(new_searchable).map(|(old, new)| old.iter().chain(new)),
+                    ) {
+                        PatternMatch::Parent
+                    } else {
+                        PatternMatch::NoMatch
+                    }
+                }
+                ActionToOperate::IndexAddedFields => {
+                    // Was not searchable but now is
+                    if !old_field_metadata.is_searchable() && new_field_metadata.is_searchable() {
+                        PatternMatch::Match
+                    // If the field is now a parent of a searchable field
+                    } else if has_searchable_children(field_name, new_searchable) {
+                        PatternMatch::Parent
+                    } else {
+                        PatternMatch::NoMatch
+                    }
+                }
+                ActionToOperate::SkipDocument => unreachable!(),
+            };
+
+            Ok((field_id, pattern_match))
+        };
+
+        let old_disabled_typos_terms = settings_delta.old_disabled_typos_terms();
+        let new_disabled_typos_terms = settings_delta.new_disabled_typos_terms();
+        let mut token_fn = |_field_name: &str, field_id, pos, word: &str| {
+            let old_field_metadata = old_fields_ids_map.metadata(field_id).unwrap();
+            let new_field_metadata = new_fields_ids_map.metadata(field_id).unwrap();
+
+            match (old_field_metadata, new_field_metadata) {
+                (
+                    Metadata { searchable: Some(_), exact: old_exact, .. },
+                    Metadata { searchable: None, .. },
+                ) => cached_sorter.insert_del_u32(
+                    field_id,
+                    pos,
+                    word,
+                    old_exact || old_disabled_typos_terms.is_exact(word),
+                    // We deleted the field globally
+                    FieldDbExtraction::Skip,
+                    document.docid(),
+                    doc_alloc,
+                ),
+                (
+                    Metadata { searchable: None, .. },
+                    Metadata { searchable: Some(_), exact: new_exact, .. },
+                ) => cached_sorter.insert_add_u32(
+                    field_id,
+                    pos,
+                    word,
+                    new_exact || new_disabled_typos_terms.is_exact(word),
+                    FieldDbExtraction::Extract,
+                    document.docid(),
+                    doc_alloc,
+                ),
+                (Metadata { searchable: None, .. }, Metadata { searchable: None, .. }) => {
+                    unreachable!()
+                }
+                (Metadata { exact: old_exact, .. }, Metadata { exact: new_exact, .. }) => {
+                    cached_sorter.insert_del_u32(
+                        field_id,
+                        pos,
+                        word,
+                        old_exact || old_disabled_typos_terms.is_exact(word),
+                        // The field has already been extracted
+                        FieldDbExtraction::Skip,
+                        document.docid(),
+                        doc_alloc,
+                    )?;
+                    cached_sorter.insert_add_u32(
+                        field_id,
+                        pos,
+                        word,
+                        new_exact || new_disabled_typos_terms.is_exact(word),
+                        // The field has already been extracted
+                        FieldDbExtraction::Skip,
+                        document.docid(),
+                        doc_alloc,
+                    )
+                }
+            }
+        };
+
+        // TODO we must tokenize twice when we change global parameters like stop words,
+        //      the language settings, dictionary, separators, non-separators...
+        document_tokenizer.tokenize_document(
+            current_document,
+            &mut should_tokenize,
+            &mut token_fn,
+        )?;
+
+        Ok(())
+    }
+}

crates/milli/src/update/new/extract/searchable/extract_word_pair_proximity_docids.rs

@@ -6,17 +6,24 @@ use bumpalo::Bump;
 
 use super::match_searchable_field;
 use super::tokenize_document::{tokenizer_builder, DocumentTokenizer};
+use crate::fields_ids_map::metadata::Metadata;
+use crate::proximity::ProximityPrecision::*;
 use crate::proximity::{index_proximity, MAX_DISTANCE};
 use crate::update::new::document::{Document, DocumentContext};
 use crate::update::new::extract::cache::BalancedCaches;
 use crate::update::new::indexer::document_changes::{
     extract, DocumentChanges, Extractor, IndexingContext,
 };
+use crate::update::new::indexer::settings_change_extract;
+use crate::update::new::indexer::settings_changes::{
+    DocumentsIndentifiers, SettingsChangeExtractor,
+};
 use crate::update::new::ref_cell_ext::RefCellExt as _;
 use crate::update::new::steps::IndexingStep;
 use crate::update::new::thread_local::{FullySend, ThreadLocal};
-use crate::update::new::DocumentChange;
-use crate::{FieldId, GlobalFieldsIdsMap, Result, MAX_POSITION_PER_ATTRIBUTE};
+use crate::update::new::{DocumentChange, DocumentIdentifiers};
+use crate::update::settings::SettingsDelta;
+use crate::{FieldId, PatternMatch, Result, UserError, MAX_POSITION_PER_ATTRIBUTE};
 
 pub struct WordPairProximityDocidsExtractorData<'a> {
     tokenizer: DocumentTokenizer<'a>,
@@ -116,7 +123,7 @@ impl WordPairProximityDocidsExtractor {
     // and to store the docids of the documents that have a number of words in a given field
     // equal to or under than MAX_COUNTED_WORDS.
     fn extract_document_change(
-        context: &DocumentContext<RefCell<BalancedCaches>>,
+        context: &DocumentContext<RefCell<BalancedCaches<'_>>>,
         document_tokenizer: &DocumentTokenizer,
         searchable_attributes: Option<&[&str]>,
         document_change: DocumentChange,
@@ -147,8 +154,12 @@ impl WordPairProximityDocidsExtractor {
                 process_document_tokens(
                     document,
                     document_tokenizer,
-                    new_fields_ids_map,
                     &mut word_positions,
+                    &mut |field_name| {
+                        new_fields_ids_map
+                            .id_with_metadata_or_insert(field_name)
+                            .ok_or(UserError::AttributeLimitReached.into())
+                    },
                     &mut |(w1, w2), prox| {
                         del_word_pair_proximity.push(((w1, w2), prox));
                     },
@@ -170,8 +181,12 @@ impl WordPairProximityDocidsExtractor {
                 process_document_tokens(
                     document,
                     document_tokenizer,
-                    new_fields_ids_map,
                     &mut word_positions,
+                    &mut |field_name| {
+                        new_fields_ids_map
+                            .id_with_metadata_or_insert(field_name)
+                            .ok_or(UserError::AttributeLimitReached.into())
+                    },
                     &mut |(w1, w2), prox| {
                         del_word_pair_proximity.push(((w1, w2), prox));
                     },
@@ -180,8 +195,12 @@ impl WordPairProximityDocidsExtractor {
                 process_document_tokens(
                     document,
                     document_tokenizer,
-                    new_fields_ids_map,
                     &mut word_positions,
+                    &mut |field_name| {
+                        new_fields_ids_map
+                            .id_with_metadata_or_insert(field_name)
+                            .ok_or(UserError::AttributeLimitReached.into())
+                    },
                     &mut |(w1, w2), prox| {
                         add_word_pair_proximity.push(((w1, w2), prox));
                     },
@@ -192,8 +211,12 @@ impl WordPairProximityDocidsExtractor {
                 process_document_tokens(
                     document,
                     document_tokenizer,
-                    new_fields_ids_map,
                     &mut word_positions,
+                    &mut |field_name| {
+                        new_fields_ids_map
+                            .id_with_metadata_or_insert(field_name)
+                            .ok_or(UserError::AttributeLimitReached.into())
+                    },
                     &mut |(w1, w2), prox| {
                         add_word_pair_proximity.push(((w1, w2), prox));
                     },
@@ -257,8 +280,8 @@ fn drain_word_positions(
 fn process_document_tokens<'doc>(
     document: impl Document<'doc>,
     document_tokenizer: &DocumentTokenizer,
-    fields_ids_map: &mut GlobalFieldsIdsMap,
     word_positions: &mut VecDeque<(Rc<str>, u16)>,
+    field_id_and_metadata: &mut impl FnMut(&str) -> Result<(FieldId, Metadata)>,
     word_pair_proximity: &mut impl FnMut((Rc<str>, Rc<str>), u8),
 ) -> Result<()> {
     let mut field_id = None;
@@ -279,8 +302,248 @@ fn process_document_tokens<'doc>(
         word_positions.push_back((Rc::from(word), pos));
         Ok(())
     };
-    document_tokenizer.tokenize_document(document, fields_ids_map, &mut token_fn)?;
+
+    let mut should_tokenize = |field_name: &str| {
+        let (field_id, meta) = field_id_and_metadata(field_name)?;
+
+        let pattern_match = if meta.is_searchable() {
+            PatternMatch::Match
+        } else {
+            // TODO: should be a match on the field_name using `match_field_legacy` function,
+            //       but for legacy reasons we iterate over all the fields to fill the field_id_map.
+            PatternMatch::Parent
+        };
+
+        Ok((field_id, pattern_match))
+    };
+
+    document_tokenizer.tokenize_document(document, &mut should_tokenize, &mut token_fn)?;
 
     drain_word_positions(word_positions, word_pair_proximity);
     Ok(())
 }
+
+pub struct WordPairProximityDocidsSettingsExtractorsData<'a, SD> {
+    tokenizer: DocumentTokenizer<'a>,
+    max_memory_by_thread: Option<usize>,
+    buckets: usize,
+    settings_delta: &'a SD,
+}
+
+impl<'extractor, SD: SettingsDelta + Sync> SettingsChangeExtractor<'extractor>
+    for WordPairProximityDocidsSettingsExtractorsData<'_, SD>
+{
+    type Data = RefCell<BalancedCaches<'extractor>>;
+
+    fn init_data<'doc>(&'doc self, extractor_alloc: &'extractor Bump) -> crate::Result<Self::Data> {
+        Ok(RefCell::new(BalancedCaches::new_in(
+            self.buckets,
+            self.max_memory_by_thread,
+            extractor_alloc,
+        )))
+    }
+
+    fn process<'doc>(
+        &'doc self,
+        documents: impl Iterator<Item = crate::Result<DocumentIdentifiers<'doc>>>,
+        context: &'doc DocumentContext<Self::Data>,
+    ) -> crate::Result<()> {
+        for document in documents {
+            let document = document?;
+            SettingsChangeWordPairProximityDocidsExtractors::extract_document_from_settings_change(
+                document,
+                context,
+                &self.tokenizer,
+                self.settings_delta,
+            )?;
+        }
+        Ok(())
+    }
+}
+
+pub struct SettingsChangeWordPairProximityDocidsExtractors;
+
+impl SettingsChangeWordPairProximityDocidsExtractors {
+    pub fn run_extraction<'fid, 'indexer, 'index, 'extractor, SD, MSP>(
+        settings_delta: &SD,
+        documents: &'indexer DocumentsIndentifiers<'indexer>,
+        indexing_context: IndexingContext<'fid, 'indexer, 'index, MSP>,
+        extractor_allocs: &'extractor mut ThreadLocal<FullySend<Bump>>,
+        step: IndexingStep,
+    ) -> Result<Vec<BalancedCaches<'extractor>>>
+    where
+        SD: SettingsDelta + Sync,
+        MSP: Fn() -> bool + Sync,
+    {
+        // Warning: this is duplicated code from extract_word_docids.rs
+        let rtxn = indexing_context.index.read_txn()?;
+        let stop_words = indexing_context.index.stop_words(&rtxn)?;
+        let allowed_separators = indexing_context.index.allowed_separators(&rtxn)?;
+        let allowed_separators: Option<Vec<_>> =
+            allowed_separators.as_ref().map(|s| s.iter().map(String::as_str).collect());
+        let dictionary = indexing_context.index.dictionary(&rtxn)?;
+        let dictionary: Option<Vec<_>> =
+            dictionary.as_ref().map(|s| s.iter().map(String::as_str).collect());
+        let mut builder = tokenizer_builder(
+            stop_words.as_ref(),
+            allowed_separators.as_deref(),
+            dictionary.as_deref(),
+        );
+        let tokenizer = builder.build();
+        let localized_attributes_rules =
+            indexing_context.index.localized_attributes_rules(&rtxn)?.unwrap_or_default();
+        let document_tokenizer = DocumentTokenizer {
+            tokenizer: &tokenizer,
+            localized_attributes_rules: &localized_attributes_rules,
+            max_positions_per_attributes: MAX_POSITION_PER_ATTRIBUTE,
+        };
+        let extractor_data = WordPairProximityDocidsSettingsExtractorsData {
+            tokenizer: document_tokenizer,
+            max_memory_by_thread: indexing_context.grenad_parameters.max_memory_by_thread(),
+            buckets: rayon::current_num_threads(),
+            settings_delta,
+        };
+        let datastore = ThreadLocal::new();
+        {
+            let span = tracing::trace_span!(target: "indexing::documents::extract", "word_pair_proximity_docids_extraction");
+            let _entered = span.enter();
+
+            settings_change_extract(
+                documents,
+                &extractor_data,
+                indexing_context,
+                extractor_allocs,
+                &datastore,
+                step,
+            )?;
+        }
+
+        Ok(datastore.into_iter().map(RefCell::into_inner).collect())
+    }
+
+    /// Extracts document words from a settings change.
+    fn extract_document_from_settings_change<SD: SettingsDelta>(
+        document: DocumentIdentifiers<'_>,
+        context: &DocumentContext<RefCell<BalancedCaches<'_>>>,
+        document_tokenizer: &DocumentTokenizer,
+        settings_delta: &SD,
+    ) -> Result<()> {
+        let mut cached_sorter = context.data.borrow_mut_or_yield();
+        let doc_alloc = &context.doc_alloc;
+
+        let new_fields_ids_map = settings_delta.new_fields_ids_map();
+        let old_fields_ids_map = settings_delta.old_fields_ids_map();
+        let old_proximity_precision = *settings_delta.old_proximity_precision();
+        let new_proximity_precision = *settings_delta.new_proximity_precision();
+
+        let current_document = document.current(
+            &context.rtxn,
+            context.index,
+            old_fields_ids_map.as_fields_ids_map(),
+        )?;
+
+        #[derive(Debug, Clone, Copy, PartialEq)]
+        enum ActionToOperate {
+            ReindexAllFields,
+            SkipDocument,
+        }
+
+        // TODO prefix_fid delete_old_fid_based_databases
+        let mut action = match (old_proximity_precision, new_proximity_precision) {
+            (ByAttribute, ByWord) => ActionToOperate::ReindexAllFields,
+            (_, _) => ActionToOperate::SkipDocument,
+        };
+
+        // Here we do a preliminary check to determine the action to take.
+        // This check doesn't trigger the tokenizer as we never return
+        // PatternMatch::Match.
+        if action != ActionToOperate::ReindexAllFields {
+            document_tokenizer.tokenize_document(
+                current_document,
+                &mut |field_name| {
+                    let fid = new_fields_ids_map.id(field_name).expect("All fields IDs must exist");
+
+                    // If the document must be reindexed, early return NoMatch to stop the scanning process.
+                    if action == ActionToOperate::ReindexAllFields {
+                        return Ok((fid, PatternMatch::NoMatch));
+                    }
+
+                    let old_field_metadata = old_fields_ids_map.metadata(fid).unwrap();
+                    let new_field_metadata = new_fields_ids_map.metadata(fid).unwrap();
+
+                    action = match (old_field_metadata, new_field_metadata) {
+                        // At least one field is removed or added from the searchable fields
+                        (
+                            Metadata { searchable: Some(_), .. },
+                            Metadata { searchable: None, .. },
+                        )
+                        | (
+                            Metadata { searchable: None, .. },
+                            Metadata { searchable: Some(_), .. },
+                        ) => ActionToOperate::ReindexAllFields,
+                        _ => action,
+                    };
+
+                    Ok((fid, PatternMatch::Parent))
+                },
+                &mut |_, _, _, _| Ok(()),
+            )?;
+        }
+
+        // Early return when we don't need to index the document
+        if action == ActionToOperate::SkipDocument {
+            return Ok(());
+        }
+
+        let mut del_word_pair_proximity = bumpalo::collections::Vec::new_in(doc_alloc);
+        let mut add_word_pair_proximity = bumpalo::collections::Vec::new_in(doc_alloc);
+
+        // is a vecdequeue, and will be smol, so can stay on the heap for now
+        let mut word_positions: VecDeque<(Rc<str>, u16)> =
+            VecDeque::with_capacity(MAX_DISTANCE as usize);
+
+        process_document_tokens(
+            current_document,
+            // TODO Tokenize must be based on old settings
+            document_tokenizer,
+            &mut word_positions,
+            &mut |field_name| {
+                Ok(old_fields_ids_map.id_with_metadata(field_name).expect("All fields must exist"))
+            },
+            &mut |(w1, w2), prox| {
+                del_word_pair_proximity.push(((w1, w2), prox));
+            },
+        )?;
+
+        process_document_tokens(
+            current_document,
+            // TODO Tokenize must be based on new settings
+            document_tokenizer,
+            &mut word_positions,
+            &mut |field_name| {
+                Ok(new_fields_ids_map.id_with_metadata(field_name).expect("All fields must exist"))
+            },
+            &mut |(w1, w2), prox| {
+                add_word_pair_proximity.push(((w1, w2), prox));
+            },
+        )?;
+
+        let mut key_buffer = bumpalo::collections::Vec::new_in(doc_alloc);
+
+        del_word_pair_proximity.sort_unstable();
+        del_word_pair_proximity.dedup_by(|(k1, _), (k2, _)| k1 == k2);
+        for ((w1, w2), prox) in del_word_pair_proximity.iter() {
+            let key = build_key(*prox, w1, w2, &mut key_buffer);
+            cached_sorter.insert_del_u32(key, document.docid())?;
+        }
+
+        add_word_pair_proximity.sort_unstable();
+        add_word_pair_proximity.dedup_by(|(k1, _), (k2, _)| k1 == k2);
+        for ((w1, w2), prox) in add_word_pair_proximity.iter() {
+            let key = build_key(*prox, w1, w2, &mut key_buffer);
+            cached_sorter.insert_add_u32(key, document.docid())?;
+        }
+
+        Ok(())
+    }
+}

crates/milli/src/update/new/extract/searchable/mod.rs

@@ -2,8 +2,12 @@ mod extract_word_docids;
 mod extract_word_pair_proximity_docids;
 mod tokenize_document;
 
-pub use extract_word_docids::{WordDocidsCaches, WordDocidsExtractors};
-pub use extract_word_pair_proximity_docids::WordPairProximityDocidsExtractor;
+pub use extract_word_docids::{
+    SettingsChangeWordDocidsExtractors, WordDocidsCaches, WordDocidsExtractors,
+};
+pub use extract_word_pair_proximity_docids::{
+    SettingsChangeWordPairProximityDocidsExtractors, WordPairProximityDocidsExtractor,
+};
 
 use crate::attribute_patterns::{match_field_legacy, PatternMatch};
 
@@ -27,3 +31,15 @@ pub fn match_searchable_field(
 
     selection
 }
+
+fn has_searchable_children<I, A>(field_name: &str, searchable: Option<I>) -> bool
+where
+    I: IntoIterator<Item = A>,
+    A: AsRef<str>,
+{
+    searchable.is_none_or(|fields| {
+        fields
+            .into_iter()
+            .any(|attr| match_field_legacy(attr.as_ref(), field_name) != PatternMatch::Parent)
+    })
+}

crates/milli/src/update/new/extract/searchable/tokenize_document.rs

@@ -8,10 +8,7 @@ use crate::update::new::document::Document;
 use crate::update::new::extract::perm_json_p::{
     seek_leaf_values_in_array, seek_leaf_values_in_object, Depth,
 };
-use crate::{
-    FieldId, GlobalFieldsIdsMap, InternalError, LocalizedAttributesRule, Result, UserError,
-    MAX_WORD_LENGTH,
-};
+use crate::{FieldId, InternalError, LocalizedAttributesRule, Result, MAX_WORD_LENGTH};
 
 // todo: should be crate::proximity::MAX_DISTANCE but it has been forgotten
 const MAX_DISTANCE: u32 = 8;
@@ -26,22 +23,16 @@ impl DocumentTokenizer<'_> {
     pub fn tokenize_document<'doc>(
         &self,
         document: impl Document<'doc>,
-        field_id_map: &mut GlobalFieldsIdsMap,
+        should_tokenize: &mut impl FnMut(&str) -> Result<(FieldId, PatternMatch)>,
         token_fn: &mut impl FnMut(&str, FieldId, u16, &str) -> Result<()>,
     ) -> Result<()> {
         let mut field_position = HashMap::new();
         let mut tokenize_field = |field_name: &str, _depth, value: &Value| {
-            let Some((field_id, meta)) = field_id_map.id_with_metadata_or_insert(field_name) else {
-                return Err(UserError::AttributeLimitReached.into());
-            };
-
-            if meta.is_searchable() {
+            let (field_id, pattern_match) = should_tokenize(field_name)?;
+            if pattern_match == PatternMatch::Match {
                 self.tokenize_field(field_id, field_name, value, token_fn, &mut field_position)?;
             }
-
-            // todo: should be a match on the field_name using `match_field_legacy` function,
-            // but for legacy reasons we iterate over all the fields to fill the field_id_map.
-            Ok(PatternMatch::Match)
+            Ok(pattern_match)
         };
 
         for entry in document.iter_top_level_fields() {
@@ -192,7 +183,7 @@ mod test {
     use super::*;
     use crate::fields_ids_map::metadata::{FieldIdMapWithMetadata, MetadataBuilder};
     use crate::update::new::document::{DocumentFromVersions, Versions};
-    use crate::FieldsIdsMap;
+    use crate::{FieldsIdsMap, GlobalFieldsIdsMap, UserError};
 
     #[test]
     fn test_tokenize_document() {
@@ -231,6 +222,7 @@ mod test {
                 Default::default(),
                 Default::default(),
                 Default::default(),
+                Default::default(),
                 None,
                 None,
                 Default::default(),
@@ -251,15 +243,19 @@ mod test {
         let document = Versions::single(document);
         let document = DocumentFromVersions::new(&document);
 
+        let mut should_tokenize = |field_name: &str| {
+            let Some(field_id) = global_fields_ids_map.id_or_insert(field_name) else {
+                return Err(UserError::AttributeLimitReached.into());
+            };
+
+            Ok((field_id, PatternMatch::Match))
+        };
+
         document_tokenizer
-            .tokenize_document(
-                document,
-                &mut global_fields_ids_map,
-                &mut |_fname, fid, pos, word| {
-                    words.insert([fid, pos], word.to_string());
-                    Ok(())
-                },
-            )
+            .tokenize_document(document, &mut should_tokenize, &mut |_fname, fid, pos, word| {
+                words.insert([fid, pos], word.to_string());
+                Ok(())
+            })
             .unwrap();
 
         snapshot!(format!("{:#?}", words), @r###"

crates/milli/src/update/new/extract/vectors/mod.rs

@@ -1,5 +1,6 @@
 use std::cell::RefCell;
 use std::fmt::Debug;
+use std::sync::RwLock;
 
 use bumpalo::collections::Vec as BVec;
 use bumpalo::Bump;
@@ -27,7 +28,10 @@ use crate::vector::extractor::{
 use crate::vector::session::{EmbedSession, Input, Metadata, OnEmbed};
 use crate::vector::settings::ReindexAction;
 use crate::vector::{Embedding, RuntimeEmbedder, RuntimeEmbedders, RuntimeFragment};
-use crate::{DocumentId, FieldDistribution, InternalError, Result, ThreadPoolNoAbort, UserError};
+use crate::{
+    DocumentId, FieldDistribution, GlobalFieldsIdsMap, InternalError, Result, ThreadPoolNoAbort,
+    UserError,
+};
 
 pub struct EmbeddingExtractor<'a, 'b> {
     embedders: &'a RuntimeEmbedders,
@@ -321,6 +325,15 @@ impl<'extractor, SD: SettingsDelta + Sync> SettingsChangeExtractor<'extractor>
         let old_embedders = self.settings_delta.old_embedders();
         let unused_vectors_distribution = UnusedVectorsDistributionBump::new_in(&context.doc_alloc);
 
+        // We get a reference to the new and old fields ids maps but
+        // note that those are local versions where updates to them
+        // will not be reflected in the database. It's not an issue
+        // because new settings do not generate new fields.
+        let new_fields_ids_map = RwLock::new(self.settings_delta.new_fields_ids_map().clone());
+        let new_fields_ids_map = RefCell::new(GlobalFieldsIdsMap::new(&new_fields_ids_map));
+        let old_fields_ids_map = RwLock::new(self.settings_delta.old_fields_ids_map().clone());
+        let old_fields_ids_map = RefCell::new(GlobalFieldsIdsMap::new(&old_fields_ids_map));
+
         let mut all_chunks = BVec::with_capacity_in(embedders.len(), &context.doc_alloc);
         let embedder_configs = context.index.embedding_configs();
         for (embedder_name, action) in self.settings_delta.embedder_actions().iter() {
@@ -396,6 +409,7 @@ impl<'extractor, SD: SettingsDelta + Sync> SettingsChangeExtractor<'extractor>
                         if !must_regenerate {
                             continue;
                         }
+
                         // we need to regenerate the prompts for the document
                         chunks.settings_change_autogenerated(
                             document.docid(),
@@ -406,7 +420,8 @@ impl<'extractor, SD: SettingsDelta + Sync> SettingsChangeExtractor<'extractor>
                                 context.db_fields_ids_map,
                             )?,
                             self.settings_delta,
-                            context.new_fields_ids_map,
+                            &old_fields_ids_map,
+                            &new_fields_ids_map,
                             &unused_vectors_distribution,
                             old_is_user_provided,
                             fragments_changed,
@@ -442,7 +457,8 @@ impl<'extractor, SD: SettingsDelta + Sync> SettingsChangeExtractor<'extractor>
                                     context.db_fields_ids_map,
                                 )?,
                                 self.settings_delta,
-                                context.new_fields_ids_map,
+                                &old_fields_ids_map,
+                                &new_fields_ids_map,
                                 &unused_vectors_distribution,
                                 old_is_user_provided,
                                 true,
@@ -638,7 +654,8 @@ impl<'a, 'b, 'extractor> Chunks<'a, 'b, 'extractor> {
         external_docid: &'a str,
         document: D,
         settings_delta: &SD,
-        fields_ids_map: &'a RefCell<crate::GlobalFieldsIdsMap>,
+        old_fields_ids_map: &'a RefCell<GlobalFieldsIdsMap<'a>>,
+        new_fields_ids_map: &'a RefCell<GlobalFieldsIdsMap<'a>>,
         unused_vectors_distribution: &UnusedVectorsDistributionBump<'a>,
         old_is_user_provided: bool,
         full_reindex: bool,
@@ -733,10 +750,17 @@ impl<'a, 'b, 'extractor> Chunks<'a, 'b, 'extractor> {
                     old_embedder.as_ref().map(|old_embedder| &old_embedder.document_template)
                 };
 
-                let extractor =
-                    DocumentTemplateExtractor::new(document_template, doc_alloc, fields_ids_map);
+                let extractor = DocumentTemplateExtractor::new(
+                    document_template,
+                    doc_alloc,
+                    new_fields_ids_map,
+                );
                 let old_extractor = old_document_template.map(|old_document_template| {
-                    DocumentTemplateExtractor::new(old_document_template, doc_alloc, fields_ids_map)
+                    DocumentTemplateExtractor::new(
+                        old_document_template,
+                        doc_alloc,
+                        old_fields_ids_map,
+                    )
                 });
                 let metadata =
                     Metadata { docid, external_docid, extractor_id: extractor.extractor_id() };

crates/milli/src/update/new/indexer/extract.rs

@@ -372,11 +372,10 @@ where
     SD: SettingsDelta + Sync,
 {
     // Create the list of document ids to extract
-    let rtxn = indexing_context.index.read_txn()?;
-    let all_document_ids =
-        indexing_context.index.documents_ids(&rtxn)?.into_iter().collect::<Vec<_>>();
-    let primary_key =
-        primary_key_from_db(indexing_context.index, &rtxn, &indexing_context.db_fields_ids_map)?;
+    let index = indexing_context.index;
+    let rtxn = index.read_txn()?;
+    let all_document_ids = index.documents_ids(&rtxn)?.into_iter().collect::<Vec<_>>();
+    let primary_key = primary_key_from_db(index, &rtxn, &indexing_context.db_fields_ids_map)?;
     let documents = DocumentsIndentifiers::new(&all_document_ids, primary_key);
 
     let span =
@@ -391,6 +390,133 @@ where
         extractor_allocs,
     )?;
 
+    {
+        let WordDocidsCaches {
+            word_docids,
+            word_fid_docids,
+            exact_word_docids,
+            word_position_docids,
+            fid_word_count_docids,
+        } = {
+            let span = tracing::trace_span!(target: "indexing::documents::extract", "word_docids");
+            let _entered = span.enter();
+            SettingsChangeWordDocidsExtractors::run_extraction(
+                settings_delta,
+                &documents,
+                indexing_context,
+                extractor_allocs,
+                IndexingStep::ExtractingWords,
+            )?
+        };
+
+        indexing_context.progress.update_progress(IndexingStep::MergingWordCaches);
+
+        {
+            let span = tracing::trace_span!(target: "indexing::documents::merge", "word_docids");
+            let _entered = span.enter();
+            indexing_context.progress.update_progress(MergingWordCache::WordDocids);
+
+            merge_and_send_docids(
+                word_docids,
+                index.word_docids.remap_types(),
+                index,
+                extractor_sender.docids::<WordDocids>(),
+                &indexing_context.must_stop_processing,
+            )?;
+        }
+
+        {
+            let span =
+                tracing::trace_span!(target: "indexing::documents::merge", "word_fid_docids");
+            let _entered = span.enter();
+            indexing_context.progress.update_progress(MergingWordCache::WordFieldIdDocids);
+
+            merge_and_send_docids(
+                word_fid_docids,
+                index.word_fid_docids.remap_types(),
+                index,
+                extractor_sender.docids::<WordFidDocids>(),
+                &indexing_context.must_stop_processing,
+            )?;
+        }
+
+        {
+            let span =
+                tracing::trace_span!(target: "indexing::documents::merge", "exact_word_docids");
+            let _entered = span.enter();
+            indexing_context.progress.update_progress(MergingWordCache::ExactWordDocids);
+
+            merge_and_send_docids(
+                exact_word_docids,
+                index.exact_word_docids.remap_types(),
+                index,
+                extractor_sender.docids::<ExactWordDocids>(),
+                &indexing_context.must_stop_processing,
+            )?;
+        }
+
+        {
+            let span =
+                tracing::trace_span!(target: "indexing::documents::merge", "word_position_docids");
+            let _entered = span.enter();
+            indexing_context.progress.update_progress(MergingWordCache::WordPositionDocids);
+
+            merge_and_send_docids(
+                word_position_docids,
+                index.word_position_docids.remap_types(),
+                index,
+                extractor_sender.docids::<WordPositionDocids>(),
+                &indexing_context.must_stop_processing,
+            )?;
+        }
+
+        {
+            let span =
+                tracing::trace_span!(target: "indexing::documents::merge", "fid_word_count_docids");
+            let _entered = span.enter();
+            indexing_context.progress.update_progress(MergingWordCache::FieldIdWordCountDocids);
+
+            merge_and_send_docids(
+                fid_word_count_docids,
+                index.field_id_word_count_docids.remap_types(),
+                index,
+                extractor_sender.docids::<FidWordCountDocids>(),
+                &indexing_context.must_stop_processing,
+            )?;
+        }
+    }
+
+    // Run the proximity extraction only if the precision is ByWord.
+    let new_proximity_precision = settings_delta.new_proximity_precision();
+    if *new_proximity_precision == ProximityPrecision::ByWord {
+        let caches = {
+            let span = tracing::trace_span!(target: "indexing::documents::extract", "word_pair_proximity_docids");
+            let _entered = span.enter();
+
+            SettingsChangeWordPairProximityDocidsExtractors::run_extraction(
+                settings_delta,
+                &documents,
+                indexing_context,
+                extractor_allocs,
+                IndexingStep::ExtractingWordProximity,
+            )?
+        };
+
+        {
+            let span = tracing::trace_span!(target: "indexing::documents::merge", "word_pair_proximity_docids");
+            let _entered = span.enter();
+            indexing_context.progress.update_progress(IndexingStep::MergingWordProximity);
+
+            merge_and_send_docids(
+                caches,
+                index.word_pair_proximity_docids.remap_types(),
+                index,
+                extractor_sender.docids::<WordPairProximityDocids>(),
+                &indexing_context.must_stop_processing,
+            )?;
+        }
+    }
+
     'vectors: {
         if settings_delta.embedder_actions().is_empty() {
             break 'vectors;

crates/milli/src/update/new/indexer/mod.rs

@@ -1,4 +1,4 @@
-use std::collections::BTreeMap;
+use std::collections::{BTreeMap, BTreeSet};
 use std::sync::atomic::AtomicBool;
 use std::sync::{Arc, Once, RwLock};
 use std::thread::{self, Builder};
@@ -8,9 +8,11 @@ use document_changes::{DocumentChanges, IndexingContext};
 pub use document_deletion::DocumentDeletion;
 pub use document_operation::{DocumentOperation, PayloadStats};
 use hashbrown::HashMap;
-use heed::{RoTxn, RwTxn};
+use heed::types::DecodeIgnore;
+use heed::{BytesDecode, Database, RoTxn, RwTxn};
 pub use partial_dump::PartialDump;
 pub use post_processing::recompute_word_fst_from_word_docids_database;
+pub use settings_changes::settings_change_extract;
 pub use update_by_function::UpdateByFunction;
 pub use write::ChannelCongestion;
 use write::{build_vectors, update_index, write_to_db};
@@ -20,12 +22,18 @@ use super::steps::IndexingStep;
 use super::thread_local::ThreadLocal;
 use crate::documents::PrimaryKey;
 use crate::fields_ids_map::metadata::{FieldIdMapWithMetadata, MetadataBuilder};
+use crate::heed_codec::StrBEU16Codec;
 use crate::progress::{EmbedderStats, Progress};
+use crate::proximity::ProximityPrecision;
+use crate::update::new::steps::SettingsIndexerStep;
+use crate::update::new::FacetFieldIdsDelta;
 use crate::update::settings::SettingsDelta;
 use crate::update::GrenadParameters;
 use crate::vector::settings::{EmbedderAction, RemoveFragments, WriteBackToDocuments};
 use crate::vector::{Embedder, RuntimeEmbedders, VectorStore};
-use crate::{FieldsIdsMap, GlobalFieldsIdsMap, Index, InternalError, Result, ThreadPoolNoAbort};
+use crate::{
+    Error, FieldsIdsMap, GlobalFieldsIdsMap, Index, InternalError, Result, ThreadPoolNoAbort,
+};
 
 pub(crate) mod de;
 pub mod document_changes;
@@ -235,6 +243,20 @@ where
     SD: SettingsDelta + Sync,
 {
     delete_old_embedders_and_fragments(wtxn, index, settings_delta)?;
+    delete_old_fid_based_databases(wtxn, index, settings_delta, must_stop_processing, progress)?;
+
+    // Clear word_pair_proximity if byWord to byAttribute
+    let old_proximity_precision = settings_delta.old_proximity_precision();
+    let new_proximity_precision = settings_delta.new_proximity_precision();
+    if *old_proximity_precision == ProximityPrecision::ByWord
+        && *new_proximity_precision == ProximityPrecision::ByAttribute
+    {
+        index.word_pair_proximity_docids.clear(wtxn)?;
+    }
+
+    // TODO delete useless searchable databases
+    //      - Clear fid_prefix_* in the post processing
+    //      - clear the prefix + fid_prefix if setting `PrefixSearch` is enabled
 
     let mut bbbuffers = Vec::new();
     let finished_extraction = AtomicBool::new(false);
@@ -293,6 +315,8 @@ where
                 .unwrap()
             })?;
 
+        let global_fields_ids_map = GlobalFieldsIdsMap::new(&new_fields_ids_map);
+
         let new_embedders = settings_delta.new_embedders();
         let embedder_actions = settings_delta.embedder_actions();
         let index_embedder_category_ids = settings_delta.new_embedder_category_id();
@@ -327,6 +351,18 @@ where
         })
         .unwrap()?;
 
+        pool.install(|| {
+            // WARN When implementing the facets don't forget this
+            let facet_field_ids_delta = FacetFieldIdsDelta::new(0, 0);
+            post_processing::post_process(
+                indexing_context,
+                wtxn,
+                global_fields_ids_map,
+                facet_field_ids_delta,
+            )
+        })
+        .unwrap()?;
+
         indexing_context.progress.update_progress(IndexingStep::BuildingGeoJson);
         index.cellulite.build(
             wtxn,
@@ -456,6 +492,107 @@ where
     Ok(())
 }
 
+/// Deletes entries refering the provided
+/// fids from the fid-based databases.
+fn delete_old_fid_based_databases<SD, MSP>(
+    wtxn: &mut RwTxn<'_>,
+    index: &Index,
+    settings_delta: &SD,
+    must_stop_processing: &MSP,
+    progress: &Progress,
+) -> Result<()>
+where
+    SD: SettingsDelta + Sync,
+    MSP: Fn() -> bool + Sync,
+{
+    let fids_to_delete: Option<BTreeSet<_>> = {
+        let rtxn = index.read_txn()?;
+        let fields_ids_map = index.fields_ids_map(&rtxn)?;
+        let old_searchable_attributes = settings_delta.old_searchable_attributes().as_ref();
+        let new_searchable_attributes = settings_delta.new_searchable_attributes().as_ref();
+        old_searchable_attributes.zip(new_searchable_attributes).map(|(old, new)| {
+            old.iter()
+                // Ignore the field if it is not searchable anymore
+                // or if it was never referenced in any document
+                .filter_map(|name| if new.contains(name) { None } else { fields_ids_map.id(name) })
+                .collect()
+        })
+    };
+
+    let fids_to_delete = match fids_to_delete {
+        Some(fids) => fids,
+        None => return Ok(()),
+    };
+
+    progress.update_progress(SettingsIndexerStep::DeletingOldWordFidDocids);
+    delete_old_word_fid_docids(wtxn, index.word_fid_docids, must_stop_processing, &fids_to_delete)?;
+
+    progress.update_progress(SettingsIndexerStep::DeletingOldFidWordCountDocids);
+    delete_old_fid_word_count_docids(wtxn, index, must_stop_processing, &fids_to_delete)?;
+
+    progress.update_progress(SettingsIndexerStep::DeletingOldWordPrefixFidDocids);
+    delete_old_word_fid_docids(
+        wtxn,
+        index.word_prefix_fid_docids,
+        must_stop_processing,
+        &fids_to_delete,
+    )?;
+
+    Ok(())
+}
+
+fn delete_old_word_fid_docids<'txn, MSP, DC>(
+    wtxn: &mut RwTxn<'txn>,
+    database: Database<StrBEU16Codec, DC>,
+    must_stop_processing: &MSP,
+    fids_to_delete: &BTreeSet<u16>,
+) -> Result<(), Error>
+where
+    MSP: Fn() -> bool + Sync,
+    DC: BytesDecode<'txn>,
+{
+    let mut iter = database.iter_mut(wtxn)?.remap_data_type::<DecodeIgnore>();
+    while let Some(((_word, fid), ())) = iter.next().transpose()? {
+        // TODO should I call it that often?
+        if must_stop_processing() {
+            return Err(Error::InternalError(InternalError::AbortedIndexation));
+        }
+
+        if fids_to_delete.contains(&fid) {
+            // safety: We don't keep any references to the data.
+            unsafe { iter.del_current()? };
+        }
+    }
+
+    Ok(())
+}
+
+fn delete_old_fid_word_count_docids<MSP>(
+    wtxn: &mut RwTxn<'_>,
+    index: &Index,
+    must_stop_processing: &MSP,
+    fids_to_delete: &BTreeSet<u16>,
+) -> Result<(), Error>
+where
+    MSP: Fn() -> bool + Sync,
+{
+    let db = index.field_id_word_count_docids.remap_data_type::<DecodeIgnore>();
+    for &fid_to_delete in fids_to_delete {
+        if must_stop_processing() {
+            return Err(Error::InternalError(InternalError::AbortedIndexation));
+        }
+
+        let mut iter = db.prefix_iter_mut(wtxn, &(fid_to_delete, 0))?;
+        while let Some(((fid, _word_count), ())) = iter.next().transpose()? {
+            debug_assert_eq!(fid, fid_to_delete);
+            // safety: We don't keep any references to the data.
+            unsafe { iter.del_current()? };
+        }
+    }
+
+    Ok(())
+}
+
 fn indexer_memory_settings(
     current_num_threads: usize,
     grenad_parameters: GrenadParameters,

crates/milli/src/update/new/steps.rs

@@ -28,6 +28,9 @@ make_enum_progress! {
         ChangingVectorStore,
         UsingStableIndexer,
         UsingExperimentalIndexer,
+        DeletingOldWordFidDocids,
+        DeletingOldFidWordCountDocids,
+        DeletingOldWordPrefixFidDocids,
     }
 }
 

crates/milli/src/update/settings.rs

@@ -1589,33 +1589,33 @@ impl<'a, 't, 'i> Settings<'a, 't, 'i> {
 
         // only use the new indexer when only the embedder possibly changed
         if let Self {
-            searchable_fields: Setting::NotSet,
+            searchable_fields: _,
             displayed_fields: Setting::NotSet,
             filterable_fields: Setting::NotSet,
             sortable_fields: Setting::NotSet,
             criteria: Setting::NotSet,
-            stop_words: Setting::NotSet,
-            non_separator_tokens: Setting::NotSet,
-            separator_tokens: Setting::NotSet,
-            dictionary: Setting::NotSet,
+            stop_words: Setting::NotSet, // TODO (require force reindexing of searchables)
+            non_separator_tokens: Setting::NotSet, // TODO (require force reindexing of searchables)
+            separator_tokens: Setting::NotSet, // TODO (require force reindexing of searchables)
+            dictionary: Setting::NotSet, // TODO (require force reindexing of searchables)
             distinct_field: Setting::NotSet,
             synonyms: Setting::NotSet,
             primary_key: Setting::NotSet,
             authorize_typos: Setting::NotSet,
             min_word_len_two_typos: Setting::NotSet,
             min_word_len_one_typo: Setting::NotSet,
-            exact_words: Setting::NotSet,
-            exact_attributes: Setting::NotSet,
+            exact_words: Setting::NotSet, // TODO (require force reindexing of searchables)
+            exact_attributes: _,
             max_values_per_facet: Setting::NotSet,
             sort_facet_values_by: Setting::NotSet,
             pagination_max_total_hits: Setting::NotSet,
-            proximity_precision: Setting::NotSet,
+            proximity_precision: _,
             embedder_settings: _,
             search_cutoff: Setting::NotSet,
-            localized_attributes_rules: Setting::NotSet,
-            prefix_search: Setting::NotSet,
+            localized_attributes_rules: Setting::NotSet, // TODO to start with
+            prefix_search: Setting::NotSet,              // TODO continue with this
             facet_search: Setting::NotSet,
-            disable_on_numbers: Setting::NotSet,
+            disable_on_numbers: Setting::NotSet, // TODO (require force reindexing of searchables)
             chat: Setting::NotSet,
             vector_store: Setting::NotSet,
             wtxn: _,
@@ -1632,10 +1632,12 @@ impl<'a, 't, 'i> Settings<'a, 't, 'i> {
             // Update index settings
             let embedding_config_updates = self.update_embedding_configs()?;
             self.update_user_defined_searchable_attributes()?;
+            self.update_exact_attributes()?;
+            self.update_proximity_precision()?;
 
-            let mut new_inner_settings =
-                InnerIndexSettings::from_index(self.index, self.wtxn, None)?;
-            new_inner_settings.recompute_searchables(self.wtxn, self.index)?;
+            // Note that we don't need to update the searchables here,
+            // as it will be done after the settings update.
+            let new_inner_settings = InnerIndexSettings::from_index(self.index, self.wtxn, None)?;
 
             let primary_key_id = self
                 .index
@@ -2062,9 +2064,12 @@ impl InnerIndexSettings {
         let sortable_fields = index.sortable_fields(rtxn)?;
         let asc_desc_fields = index.asc_desc_fields(rtxn)?;
         let distinct_field = index.distinct_field(rtxn)?.map(|f| f.to_string());
-        let user_defined_searchable_attributes = index
-            .user_defined_searchable_fields(rtxn)?
-            .map(|fields| fields.into_iter().map(|f| f.to_string()).collect());
+        let user_defined_searchable_attributes = match index.user_defined_searchable_fields(rtxn)? {
+            Some(fields) if fields.contains(&"*") => None,
+            Some(fields) => Some(fields.into_iter().map(|f| f.to_string()).collect()),
+            None => None,
+        };
+
         let builder = MetadataBuilder::from_index(index, rtxn)?;
         let fields_ids_map = FieldIdMapWithMetadata::new(fields_ids_map, builder);
         let disabled_typos_terms = index.disabled_typos_terms(rtxn)?;
@@ -2578,8 +2583,20 @@ fn deserialize_sub_embedder(
 /// Implement this trait for the settings delta type.
 /// This is used in the new settings update flow and will allow to easily replace the old settings delta type: `InnerIndexSettingsDiff`.
 pub trait SettingsDelta {
-    fn new_embedders(&self) -> &RuntimeEmbedders;
+    fn old_fields_ids_map(&self) -> &FieldIdMapWithMetadata;
+    fn new_fields_ids_map(&self) -> &FieldIdMapWithMetadata;
+
+    fn old_searchable_attributes(&self) -> &Option<Vec<String>>;
+    fn new_searchable_attributes(&self) -> &Option<Vec<String>>;
+
+    fn old_disabled_typos_terms(&self) -> &DisabledTyposTerms;
+    fn new_disabled_typos_terms(&self) -> &DisabledTyposTerms;
+
+    fn old_proximity_precision(&self) -> &ProximityPrecision;
+    fn new_proximity_precision(&self) -> &ProximityPrecision;
+
     fn old_embedders(&self) -> &RuntimeEmbedders;
+    fn new_embedders(&self) -> &RuntimeEmbedders;
     fn new_embedder_category_id(&self) -> &HashMap<String, u8>;
     fn embedder_actions(&self) -> &BTreeMap<String, EmbedderAction>;
     fn try_for_each_fragment_diff<F, E>(
@@ -2589,7 +2606,6 @@ pub trait SettingsDelta {
     ) -> std::result::Result<(), E>
     where
         F: FnMut(FragmentDiff) -> std::result::Result<(), E>;
-    fn new_fields_ids_map(&self) -> &FieldIdMapWithMetadata;
 }
 
 pub struct FragmentDiff<'a> {
@@ -2598,26 +2614,47 @@ pub struct FragmentDiff<'a> {
 }
 
 impl SettingsDelta for InnerIndexSettingsDiff {
-    fn new_embedders(&self) -> &RuntimeEmbedders {
-        &self.new.runtime_embedders
+    fn old_fields_ids_map(&self) -> &FieldIdMapWithMetadata {
+        &self.old.fields_ids_map
+    }
+    fn new_fields_ids_map(&self) -> &FieldIdMapWithMetadata {
+        &self.new.fields_ids_map
+    }
+
+    fn old_searchable_attributes(&self) -> &Option<Vec<String>> {
+        &self.old.user_defined_searchable_attributes
+    }
+    fn new_searchable_attributes(&self) -> &Option<Vec<String>> {
+        &self.new.user_defined_searchable_attributes
+    }
+
+    fn old_disabled_typos_terms(&self) -> &DisabledTyposTerms {
+        &self.old.disabled_typos_terms
+    }
+    fn new_disabled_typos_terms(&self) -> &DisabledTyposTerms {
+        &self.new.disabled_typos_terms
+    }
+
+    fn old_proximity_precision(&self) -> &ProximityPrecision {
+        &self.old.proximity_precision
+    }
+    fn new_proximity_precision(&self) -> &ProximityPrecision {
+        &self.new.proximity_precision
     }
 
     fn old_embedders(&self) -> &RuntimeEmbedders {
         &self.old.runtime_embedders
     }
+    fn new_embedders(&self) -> &RuntimeEmbedders {
+        &self.new.runtime_embedders
+    }
 
     fn new_embedder_category_id(&self) -> &HashMap<String, u8> {
         &self.new.embedder_category_id
     }
-
     fn embedder_actions(&self) -> &BTreeMap<String, EmbedderAction> {
         &self.embedding_config_updates
     }
-
-    fn new_fields_ids_map(&self) -> &FieldIdMapWithMetadata {
-        &self.new.fields_ids_map
-    }
-
     fn try_for_each_fragment_diff<F, E>(
         &self,
         embedder_name: &str,

crates/milli/src/update/test_settings.rs

@@ -14,28 +14,21 @@ fn set_and_reset_searchable_fields() {
     let index = TempIndex::new();
 
     // First we send 3 documents with ids from 1 to 3.
-    let mut wtxn = index.write_txn().unwrap();
-
     index
-        .add_documents_using_wtxn(
-            &mut wtxn,
-            documents!([
-                { "id": 1, "name": "kevin", "age": 23 },
-                { "id": 2, "name": "kevina", "age": 21},
-                { "id": 3, "name": "benoit", "age": 34 }
-            ]),
-        )
+        .add_documents(documents!([
+            { "id": 1, "name": "kevin", "age": 23 },
+            { "id": 2, "name": "kevina", "age": 21},
+            { "id": 3, "name": "benoit", "age": 34 }
+        ]))
         .unwrap();
 
     // We change the searchable fields to be the "name" field only.
     index
-        .update_settings_using_wtxn(&mut wtxn, |settings| {
+        .update_settings(|settings| {
             settings.set_searchable_fields(vec!["name".into()]);
         })
         .unwrap();
 
-    wtxn.commit().unwrap();
-
     db_snap!(index, fields_ids_map, @r###"
     0   id               |
     1   name             |