fix: Apply rescore_oversample to V3 ES kNN retrievers#363
Merged
Conversation
Manisha4
changed the title
piket
approved these changes
May 7, 2026
V3's retriever construction was silently ignoring rescore_oversample. V2 honored it (lines 483-486, 617-620) but V3 never added rescore_vector to its kNN clauses. On quantized indices (int8_hnsw / int4_hnsw / bbq_hnsw), this meant V3 queries returned lower recall than the config promised, with no error or warning. Wire rescore_oversample into each kNN retriever the same way V2 does. Covers single-vector and multi-vector V3 queries; BM25 retrievers skip the branch since they lack a "knn" key. Existing config validation (lines 102-127) already prevents rescore on non-quantized indices, so no new validation needed. Added three unit tests in TestRetrieveOnlineDocumentsV3QueryBuilding: - rescore_vector appears in single-vector query body when configured - rescore_vector appears on every kNN retriever in multi-vector query - rescore_vector absent when rescore_oversample is None Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Manisha4
added a commit
that referenced
this pull request
May 11, 2026
* fix: Updated the elastic search to use vector_index defined in the feature view to identify vector fields (#348) * updated the elastic search to use vector_index defined in the feature view to identify vector fields * fix: formatting * Added logging and switched to use open source elastic search --------- Co-authored-by: vanitabhagwat <vbhagwat@expediagroup.com> * fix: ES integration tests (#350) * fix: ES integration tests * fix: Added fromisoformat() for converting timestamps --------- Co-authored-by: vanitabhagwat <vbhagwat@expediagroup.com> * fix:Elasticsearch online store — correctness, performance, and robustness fixes (#353) Co-authored-by: vanitabhagwat <vbhagwat@expediagroup.com> * Feature/vector store (#357) * feat: Valkey Online Write Batch Vector Search Support (#351) * Adding support for Valkey Search, adding changes to the online_write_batch functionality * Addressing PR comments * addressing linting error * fix tests * addressing PR comments * addressing PR comments * fixing linting --------- Co-authored-by: Manisha4 <Manisha4@github.com> * feat: Support Vector Search in Valkey (#354) * Adding support for Valkey Search, adding changes to the online_write_batch functionality * Addressing PR comments * addressing linting error * Adding changes to support search in valkey * fix tests * adding unit tests * reformatting files and adding checks and more tests * reformatting files and adding checks and more tests * reformatting files and adding checks and more tests * Fix linter errors: type annotations and code formatting - Add explicit type annotation for schema_fields to support both TagField and VectorField - Encode project string to bytes for consistency with other hash values - Decode doc_key bytes to string for hmget compatibility - Fix code formatting: break long lines and remove extra blank lines - Remove tests for multiple vector fields (Feast enforces one vector per feature view) - Fix config type: use 'eg-valkey' (hyphen) not 'eg_valkey' (underscore) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * addressing PR comments * addressing PR comments * fixing linting * Fix missing feature_name argument in retrieve_online_documents_v2 Add the third argument (vector_field.name) to _get_vector_index_name call to match the updated function signature. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com> * addressing comments, PR changes for some fixes and merge conflicts * fixing tests * fixing tests * fixing linting * fixing linting --------- Co-authored-by: Manisha4 <Manisha4@github.com> Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> * fix: Valkey vector search - remove unsupported SORTBY (#356) * fix: Valkey vector search - remove unsupported SORTBY and fix tag filter syntax Valkey Search KNN queries return results pre-sorted by distance, so explicit SORTBY is not supported and causes a ResponseError. This removes the .sort_by() call from the query builder. Additionally, fixes the project tag filter to use unquoted syntax with backslash escaping for special characters (e.g. hyphens, dots) instead of the quoted syntax which was returning empty results. Updates unit tests to reflect both changes: replaces three metric-specific sort order tests with a single test asserting no SORTBY is set, and updates escaping assertions to match the new backslash-escape approach. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * style: apply ruff format to eg_valkey.py and test_valkey.py Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> --------- Co-authored-by: Manisha4 <Manisha4@github.com> Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> --------- Co-authored-by: Manisha4 <Manisha4@github.com> Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> * feat: Quantization support for elastic search (#355) * feat: Add retrieve_online_documents_v3 SDK with multi-vector and hybrid fusion (#358) * feat: implement retrieve_online_documents_v3 SDK method - Multi-vector search with configurable fusion (RRF, WEIGHTED_LINEAR, VECTOR_ONLY) via the ES retriever API. Valkey gracefully degrades to single-vector KNN with warnings. - "embedding" magic key for V2→V3 migration convenience - Reserved output fields: final_score, signal_scores - include_signal_scores and distance_metric accepted as reserved params - ODFV and reserved-name collision validation - Shared signal_scores encoding via _signal_scores helper * update tests * update tests * fixing linting * docs: clarify final_score semantics in Valkey V3 docstring Correct the Valkey final_score description — Valkey's __distance__ is lower-is-better across all metrics (COSINE, L2, IP), not higher-is-better for IP. Call out the ordering inversion vs Elasticsearch so callers don't assume cross-backend score portability. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * fix: plumb include_signal_scores through V3 and align defaults to False Valkey/ES/provider/passthrough/online-store defaults were True, mismatching the SDK's False default. Align all layers on False and thread the parameter from retrieve_online_documents_v3 through the internal dispatcher, provider, and online stores so callers can opt in today and transparently pick up the explain-based per-signal path when it lands — no API change required. Tighten docstrings to describe the current best-effort behavior instead of hinting at latency tradeoffs that aren't wired yet. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * updating doc string * fix: preserve ranked row order in V3 retrieve_online_documents _retrieve_from_online_store_v3 was passing the driver's ranked rows through _get_unique_entities_from_values, which sorts and dedupes by entity-key bytes. That helper is correct for batch entity lookups but wrong here — ES/Valkey have already ordered rows by relevance, and the sort was scrambling them in the final DataFrame (e.g. doc_10 jumping ahead of doc_3 because "10" < "3" lexicographically). Replace the helper call with an identity mapping so the driver's rank order flows through untouched. No change to V1, V2, batch reads, or the helper itself; V3 output now matches the order returned by the online store. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> --------- Co-authored-by: Manisha4 <Manisha4@github.com> Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com> * fix: apply rescore_oversample to V3 ES kNN retrievers (#363) V3's retriever construction was silently ignoring rescore_oversample. V2 honored it (lines 483-486, 617-620) but V3 never added rescore_vector to its kNN clauses. On quantized indices (int8_hnsw / int4_hnsw / bbq_hnsw), this meant V3 queries returned lower recall than the config promised, with no error or warning. Wire rescore_oversample into each kNN retriever the same way V2 does. Covers single-vector and multi-vector V3 queries; BM25 retrievers skip the branch since they lack a "knn" key. Existing config validation (lines 102-127) already prevents rescore on non-quantized indices, so no new validation needed. Added three unit tests in TestRetrieveOnlineDocumentsV3QueryBuilding: - rescore_vector appears in single-vector query body when configured - rescore_vector appears on every kNN retriever in multi-vector query - rescore_vector absent when rescore_oversample is None Co-authored-by: Manisha4 <Manisha4@github.com> Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com> --------- Co-authored-by: vanitabhagwat <92561664+vanitabhagwat@users.noreply.github.com> Co-authored-by: vanitabhagwat <vbhagwat@expediagroup.com> Co-authored-by: Manisha4 <Manisha4@github.com> Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What this PR does / why we need it:
V3's retriever construction was silently ignoring rescore_oversample. V2 honored it (lines 483-486, 617-620) but V3 never added rescore_vector to its kNN clauses. On quantized indices (int8_hnsw / int4_hnsw / bbq_hnsw), this meant V3 queries returned lower recall than the config promised, with no error or warning.
Wire rescore_oversample into each kNN retriever the same way V2 does. Covers single-vector and multi-vector V3 queries; BM25 retrievers skip the branch since they lack a "knn" key.
Existing config validation (lines 102-127) already prevents rescore on non-quantized indices, so no new validation needed.
Added three unit tests in TestRetrieveOnlineDocumentsV3QueryBuilding:
Which issue(s) this PR fixes:
Misc