snowykami/meilisearch
1
0
Fork 0
mirror of https://github.com/meilisearch/meilisearch.git synced 2025-09-13 08:16:25 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'main' into indexer-edition-2024

Browse Source
This commit is contained in:
ManyTheFish
2024-09-25 07:37:32 +02:00
parent 7ad037841f 7f20c13f3f
commit 974272f2e9
94 changed files with 8510 additions and 4616 deletions
Download Patch File Download Diff File Expand all files Collapse all files

126
milli/src/update/index_documents/extract/extract_vector_points.rs
View File

@ -15,14 +15,14 @@ use serde_json::Value;
use super::helpers::{create_writer, writer_into_reader, GrenadParameters};
use crate::error::FaultSource;
use crate::index::IndexEmbeddingConfig;
use crate::prompt::Prompt;
use crate::prompt::{FieldsIdsMapWithMetadata, Prompt};
use crate::update::del_add::{DelAdd, KvReaderDelAdd, KvWriterDelAdd};
use crate::update::settings::InnerIndexSettingsDiff;
use crate::vector::error::{EmbedErrorKind, PossibleEmbeddingMistakes, UnusedVectorsDistribution};
use crate::vector::parsed_vectors::{ParsedVectorsDiff, VectorState, RESERVED_VECTORS_FIELD_NAME};
use crate::vector::settings::{EmbedderAction, ReindexAction};
use crate::vector::settings::ReindexAction;
use crate::vector::{Embedder, Embeddings};
use crate::{try_split_array_at, DocumentId, FieldId, FieldsIdsMap, Result, ThreadPoolNoAbort};
use crate::{try_split_array_at, DocumentId, FieldId, Result, ThreadPoolNoAbort};
/// The length of the elements that are always in the buffer when inserting new values.
const TRUNCATE_SIZE: usize = size_of::<DocumentId>();
@ -189,7 +189,13 @@ pub fn extract_vector_points<R: io::Read + io::Seek>(
let reindex_vectors = settings_diff.reindex_vectors();
let old_fields_ids_map = &settings_diff.old.fields_ids_map;
let old_fields_ids_map =
FieldsIdsMapWithMetadata::new(old_fields_ids_map, &settings_diff.old.searchable_fields_ids);
let new_fields_ids_map = &settings_diff.new.fields_ids_map;
let new_fields_ids_map =
FieldsIdsMapWithMetadata::new(new_fields_ids_map, &settings_diff.new.searchable_fields_ids);
// the vector field id may have changed
let old_vectors_fid = old_fields_ids_map.id(RESERVED_VECTORS_FIELD_NAME);
@ -202,65 +208,65 @@ pub fn extract_vector_points<R: io::Read + io::Seek>(
if reindex_vectors {
for (name, action) in settings_diff.embedding_config_updates.iter() {
match action {
EmbedderAction::WriteBackToDocuments(_) => continue, // already deleted
EmbedderAction::Reindex(action) => {
let Some((embedder_name, (embedder, prompt))) = configs.remove_entry(name)
else {
tracing::error!(embedder = name, "Requested embedder config not found");
continue;
};
if let Some(action) = action.reindex() {
let Some((embedder_name, (embedder, prompt, _quantized))) =
configs.remove_entry(name)
else {
tracing::error!(embedder = name, "Requested embedder config not found");
continue;
};
// (docid, _index) -> KvWriterDelAdd -> Vector
let manual_vectors_writer = create_writer(
indexer.chunk_compression_type,
indexer.chunk_compression_level,
tempfile::tempfile()?,
);
// (docid, _index) -> KvWriterDelAdd -> Vector
let manual_vectors_writer = create_writer(
indexer.chunk_compression_type,
indexer.chunk_compression_level,
tempfile::tempfile()?,
);
// (docid) -> (prompt)
let prompts_writer = create_writer(
indexer.chunk_compression_type,
indexer.chunk_compression_level,
tempfile::tempfile()?,
);
// (docid) -> (prompt)
let prompts_writer = create_writer(
indexer.chunk_compression_type,
indexer.chunk_compression_level,
tempfile::tempfile()?,
);
// (docid) -> ()
let remove_vectors_writer = create_writer(
indexer.chunk_compression_type,
indexer.chunk_compression_level,
tempfile::tempfile()?,
);
// (docid) -> ()
let remove_vectors_writer = create_writer(
indexer.chunk_compression_type,
indexer.chunk_compression_level,
tempfile::tempfile()?,
);
let action = match action {
ReindexAction::FullReindex => ExtractionAction::SettingsFullReindex,
ReindexAction::RegeneratePrompts => {
let Some((_, old_prompt)) = old_configs.get(name) else {
tracing::error!(embedder = name, "Old embedder config not found");
continue;
};
let action = match action {
ReindexAction::FullReindex => ExtractionAction::SettingsFullReindex,
ReindexAction::RegeneratePrompts => {
let Some((_, old_prompt, _quantized)) = old_configs.get(name) else {
tracing::error!(embedder = name, "Old embedder config not found");
continue;
};
ExtractionAction::SettingsRegeneratePrompts { old_prompt }
}
};
ExtractionAction::SettingsRegeneratePrompts { old_prompt }
}
};
extractors.push(EmbedderVectorExtractor {
embedder_name,
embedder,
prompt,
prompts_writer,
remove_vectors_writer,
manual_vectors_writer,
add_to_user_provided: RoaringBitmap::new(),
action,
});
}
extractors.push(EmbedderVectorExtractor {
embedder_name,
embedder,
prompt,
prompts_writer,
remove_vectors_writer,
manual_vectors_writer,
add_to_user_provided: RoaringBitmap::new(),
action,
});
} else {
continue;
}
}
} else {
// document operation
for (embedder_name, (embedder, prompt)) in configs.into_iter() {
for (embedder_name, (embedder, prompt, _quantized)) in configs.into_iter() {
// (docid, _index) -> KvWriterDelAdd -> Vector
let manual_vectors_writer = create_writer(
indexer.chunk_compression_type,
@ -376,7 +382,7 @@ pub fn extract_vector_points<R: io::Read + io::Seek>(
);
continue;
}
regenerate_prompt(obkv, prompt, new_fields_ids_map)?
regenerate_prompt(obkv, prompt, &new_fields_ids_map)?
}
},
// prompt regeneration is only triggered for existing embedders
@ -393,7 +399,7 @@ pub fn extract_vector_points<R: io::Read + io::Seek>(
regenerate_if_prompt_changed(
obkv,
(old_prompt, prompt),
(old_fields_ids_map, new_fields_ids_map),
(&old_fields_ids_map, &new_fields_ids_map),
)?
} else {
// we can simply ignore user provided vectors as they are not regenerated and are
@ -409,7 +415,7 @@ pub fn extract_vector_points<R: io::Read + io::Seek>(
prompt,
(add_to_user_provided, remove_from_user_provided),
(old, new),
(old_fields_ids_map, new_fields_ids_map),
(&old_fields_ids_map, &new_fields_ids_map),
document_id,
embedder_name,
embedder_is_manual,
@ -479,7 +485,10 @@ fn extract_vector_document_diff(
prompt: &Prompt,
(add_to_user_provided, remove_from_user_provided): (&mut RoaringBitmap, &mut RoaringBitmap),
(old, new): (VectorState, VectorState),
(old_fields_ids_map, new_fields_ids_map): (&FieldsIdsMap, &FieldsIdsMap),
(old_fields_ids_map, new_fields_ids_map): (
&FieldsIdsMapWithMetadata,
&FieldsIdsMapWithMetadata,
),
document_id: impl Fn() -> Value,
embedder_name: &str,
embedder_is_manual: bool,
@ -599,7 +608,10 @@ fn extract_vector_document_diff(
fn regenerate_if_prompt_changed(
obkv: &obkv::KvReader<FieldId>,
(old_prompt, new_prompt): (&Prompt, &Prompt),
(old_fields_ids_map, new_fields_ids_map): (&FieldsIdsMap, &FieldsIdsMap),
(old_fields_ids_map, new_fields_ids_map): (
&FieldsIdsMapWithMetadata,
&FieldsIdsMapWithMetadata,
),
) -> Result<VectorStateDelta> {
let old_prompt =
old_prompt.render(obkv, DelAdd::Deletion, old_fields_ids_map).unwrap_or(Default::default());
@ -614,7 +626,7 @@ fn regenerate_if_prompt_changed(
fn regenerate_prompt(
obkv: &obkv::KvReader<FieldId>,
prompt: &Prompt,
new_fields_ids_map: &FieldsIdsMap,
new_fields_ids_map: &FieldsIdsMapWithMetadata,
) -> Result<VectorStateDelta> {
let prompt = prompt.render(obkv, DelAdd::Addition, new_fields_ids_map)?;

42
milli/src/update/index_documents/mod.rs
View File

@ -37,7 +37,7 @@ use crate::update::index_documents::parallel::ImmutableObkvs;
use crate::update::{
IndexerConfig, UpdateIndexingStep, WordPrefixDocids, WordPrefixIntegerDocids, WordsPrefixesFst,
};
use crate::vector::EmbeddingConfigs;
use crate::vector::{ArroyWrapper, EmbeddingConfigs};
use crate::{CboRoaringBitmapCodec, Index, Object, Result};
static MERGED_DATABASE_COUNT: usize = 7;
@ -673,6 +673,24 @@ where
let number_of_documents = self.index.number_of_documents(self.wtxn)?;
let mut rng = rand::rngs::StdRng::seed_from_u64(42);
// If an embedder wasn't used in the typedchunk but must be binary quantized
// we should insert it in `dimension`
for (name, action) in settings_diff.embedding_config_updates.iter() {
if action.is_being_quantized && !dimension.contains_key(name.as_str()) {
let index = self.index.embedder_category_id.get(self.wtxn, name)?.ok_or(
InternalError::DatabaseMissingEntry {
db_name: "embedder_category_id",
key: None,
},
)?;
let first_id = crate::vector::arroy_db_range_for_embedder(index).next().unwrap();
let reader =
ArroyWrapper::new(self.index.vector_arroy, first_id, action.was_quantized);
let dim = reader.dimensions(self.wtxn)?;
dimension.insert(name.to_string(), dim);
}
}
for (embedder_name, dimension) in dimension {
let wtxn = &mut *self.wtxn;
let vector_arroy = self.index.vector_arroy;
@ -680,13 +698,23 @@ where
let embedder_index = self.index.embedder_category_id.get(wtxn, &embedder_name)?.ok_or(
InternalError::DatabaseMissingEntry { db_name: "embedder_category_id", key: None },
)?;
let embedder_config = settings_diff.embedding_config_updates.get(&embedder_name);
let was_quantized = settings_diff
.old
.embedding_configs
.get(&embedder_name)
.map_or(false, |conf| conf.2);
let is_quantizing = embedder_config.map_or(false, |action| action.is_being_quantized);
pool.install(|| {
for k in crate::vector::arroy_db_range_for_embedder(embedder_index) {
let writer = arroy::Writer::new(vector_arroy, k, dimension);
if writer.need_build(wtxn)? {
writer.build(wtxn, &mut rng, None)?;
} else if writer.is_empty(wtxn)? {
let mut writer = ArroyWrapper::new(vector_arroy, k, was_quantized);
if is_quantizing {
writer.quantize(wtxn, k, dimension)?;
}
if writer.need_build(wtxn, dimension)? {
writer.build(wtxn, &mut rng, dimension)?;
} else if writer.is_empty(wtxn, dimension)? {
break;
}
}
@ -2734,11 +2762,13 @@ mod tests {
api_key: Setting::NotSet,
dimensions: Setting::Set(3),
document_template: Setting::NotSet,
document_template_max_bytes: Setting::NotSet,
url: Setting::NotSet,
request: Setting::NotSet,
response: Setting::NotSet,
distribution: Setting::NotSet,
headers: Setting::NotSet,
binary_quantized: Setting::NotSet,
}),
);
settings.set_embedder_settings(embedders);
@ -2767,7 +2797,7 @@ mod tests {
std::sync::Arc::new(crate::vector::Embedder::new(embedder.embedder_options).unwrap());
let res = index
.search(&rtxn)
.semantic(embedder_name, embedder, Some([0.0, 1.0, 2.0].to_vec()))
.semantic(embedder_name, embedder, false, Some([0.0, 1.0, 2.0].to_vec()))
.execute()
.unwrap();
assert_eq!(res.documents_ids.len(), 3);

32
milli/src/update/index_documents/transform.rs
View File

@ -29,7 +29,8 @@ use crate::update::index_documents::GrenadParameters;
use crate::update::settings::{InnerIndexSettings, InnerIndexSettingsDiff};
use crate::update::{AvailableIds, UpdateIndexingStep};
use crate::vector::parsed_vectors::{ExplicitVectors, VectorOrArrayOfVectors};
use crate::vector::settings::{EmbedderAction, WriteBackToDocuments};
use crate::vector::settings::WriteBackToDocuments;
use crate::vector::ArroyWrapper;
use crate::{
is_faceted_by, FieldDistribution, FieldId, FieldIdMapMissingEntry, FieldsIdsMap, Index, Result,
};
@ -992,19 +993,17 @@ impl<'a, 'i> Transform<'a, 'i> {
None
};
let readers: Result<
BTreeMap<&str, (Vec<arroy::Reader<'_, arroy::distances::Angular>>, &RoaringBitmap)>,
> = settings_diff
let readers: Result<BTreeMap<&str, (Vec<ArroyWrapper>, &RoaringBitmap)>> = settings_diff
.embedding_config_updates
.iter()
.filter_map(|(name, action)| {
if let EmbedderAction::WriteBackToDocuments(WriteBackToDocuments {
embedder_id,
user_provided,
}) = action
if let Some(WriteBackToDocuments { embedder_id, user_provided }) =
action.write_back()
{
let readers: Result<Vec<_>> =
self.index.arroy_readers(wtxn, *embedder_id).collect();
let readers: Result<Vec<_>> = self
.index
.arroy_readers(wtxn, *embedder_id, action.was_quantized)
.collect();
match readers {
Ok(readers) => Some(Ok((name.as_str(), (readers, user_provided)))),
Err(error) => Some(Err(error)),
@ -1107,23 +1106,14 @@ impl<'a, 'i> Transform<'a, 'i> {
}
}
let mut writers = Vec::new();
// delete all vectors from the embedders that need removal
for (_, (readers, _)) in readers {
for reader in readers {
let dimensions = reader.dimensions();
let arroy_index = reader.index();
drop(reader);
let writer = arroy::Writer::new(self.index.vector_arroy, arroy_index, dimensions);
writers.push(writer);
let dimensions = reader.dimensions(wtxn)?;
reader.clear(wtxn, dimensions)?;
}
}
for writer in writers {
writer.clear(wtxn)?;
}
let grenad_params = GrenadParameters {
chunk_compression_type: self.indexer_settings.chunk_compression_type,
chunk_compression_level: self.indexer_settings.chunk_compression_level,

27
milli/src/update/index_documents/typed_chunk.rs
View File

@ -26,6 +26,7 @@ use crate::update::index_documents::helpers::{
as_cloneable_grenad, try_split_array_at, KeepLatestObkv,
};
use crate::update::settings::InnerIndexSettingsDiff;
use crate::vector::ArroyWrapper;
use crate::{
lat_lng_to_xyz, CboRoaringBitmapCodec, DocumentId, FieldId, GeoPoint, Index, InternalError,
Result, SerializationError, U8StrStrCodec,
@ -661,9 +662,14 @@ pub(crate) fn write_typed_chunk_into_index(
let embedder_index = index.embedder_category_id.get(wtxn, &embedder_name)?.ok_or(
InternalError::DatabaseMissingEntry { db_name: "embedder_category_id", key: None },
)?;
let binary_quantized = settings_diff
.old
.embedding_configs
.get(&embedder_name)
.map_or(false, |conf| conf.2);
// FIXME: allow customizing distance
let writers: Vec<_> = crate::vector::arroy_db_range_for_embedder(embedder_index)
.map(|k| arroy::Writer::new(index.vector_arroy, k, expected_dimension))
.map(|k| ArroyWrapper::new(index.vector_arroy, k, binary_quantized))
.collect();
// remove vectors for docids we want them removed
@ -674,7 +680,7 @@ pub(crate) fn write_typed_chunk_into_index(
for writer in &writers {
// Uses invariant: vectors are packed in the first writers.
if !writer.del_item(wtxn, docid)? {
if !writer.del_item(wtxn, expected_dimension, docid)? {
break;
}
}
@ -706,7 +712,7 @@ pub(crate) fn write_typed_chunk_into_index(
)));
}
for (embedding, writer) in embeddings.iter().zip(&writers) {
writer.add_item(wtxn, docid, embedding)?;
writer.add_item(wtxn, expected_dimension, docid, embedding)?;
}
}
@ -729,7 +735,7 @@ pub(crate) fn write_typed_chunk_into_index(
break;
};
if candidate == vector {
writer.del_item(wtxn, docid)?;
writer.del_item(wtxn, expected_dimension, docid)?;
deleted_index = Some(index);
}
}
@ -746,8 +752,13 @@ pub(crate) fn write_typed_chunk_into_index(
if let Some((last_index, vector)) = last_index_with_a_vector {
// unwrap: computed the index from the list of writers
let writer = writers.get(last_index).unwrap();
writer.del_item(wtxn, docid)?;
writers.get(deleted_index).unwrap().add_item(wtxn, docid, &vector)?;
writer.del_item(wtxn, expected_dimension, docid)?;
writers.get(deleted_index).unwrap().add_item(
wtxn,
expected_dimension,
docid,
&vector,
)?;
}
}
}
@ -757,8 +768,8 @@ pub(crate) fn write_typed_chunk_into_index(
// overflow was detected during vector extraction.
for writer in &writers {
if !writer.contains_item(wtxn, docid)? {
writer.add_item(wtxn, docid, &vector)?;
if !writer.contains_item(wtxn, expected_dimension, docid)? {
writer.add_item(wtxn, expected_dimension, docid, &vector)?;
break;
}
}