[aprudhomme]

Summary

Adds the option to rescore quantized KNN vector search hits using the full precision vectors. This is the same idea as Elasticsearch's rescore_vector.

For hnsw_scalar_quantized fields, both the HNSW graph search and the hit scores come from the quantized vectors. The full precision float vectors are still stored in the index. A common way to recover accuracy is to oversample the quantized search, rescore those candidates with the full precision vectors, and return the top k.

Changes

KnnQuery.rescore_vector

New RescoreVector message with an oversample factor:

"knn": [
  {
    "field": "embedding_quantized",
    "k": 10,
    "num_candidates": 100,
    "query_vector": [0.18, 0.24, ...],
    "rescore_vector": { "oversample": 3.0 }
  }
]

• The vector search collects ceil(k * oversample) hits, capped at the num_candidates limit of 10000. num_candidates is raised to this value if it is lower.
• Hits are rescored with the full precision vectors, and the top k are returned with full precision similarity scores.
• oversample must be >= 1.0, or 0 (default) to disable rescoring. A value of 1.0 re-ranks the top k without collecting extra hits.
• similarityThreshold and boost are applied to the rescored score.
• Works for both top level knn queries and multi-retriever KnnRetriever, since they share the same query building path.
• Only supported for float vector fields. Byte fields are not quantized, and requests that use the option on them return an error.
• For nested vector fields, only the best matching child of each parent is rescored.

Implementation uses Lucene's RescoreTopNQuery with FullPrecisionFloatVectorSimilarityValuesSource. The new VectorRescoreQuery wrapper:

• exposes the inner query's vector total hits through WithVectorTotalHits, so vector diagnostics keep working
• preserves MatchNoDocsQuery when the filter matches no documents

ExactVectorQuery.full_precision

ExactVectorQuery scores with FloatVectorValues.scorer(), which uses the quantized vectors when the field has a quantized vector search type. The new full_precision flag switches to FloatVectorValues.rescorer(), which uses the full precision vectors. It defaults to false, so existing behavior is unchanged.

This also allows full precision rescoring of any first pass query through the existing rescorer API:

"rescorers": [
  {
    "windowSize": 100,
    "name": "vector_rescore",
    "queryRescorer": {
      "rescoreQuery": {
        "exactVectorQuery": {
          "field": "embedding_quantized",
          "query_float_vector": [0.18, 0.24, ...],
          "full_precision": true
        }
      },
      "queryWeight": 0.0,
      "rescoreQueryWeight": 1.0
    }
  }
]

Only supported for float vector fields.

Other

• Updated docs/vector_search.rst with a "Rescoring Quantized Vector Search" section. Also corrected the KnnQuery proto snippet.
• Regenerated the grpc-gateway files.

Testing

New tests in VectorFieldDefTest, using a 4 bit quantized field:

• Rescored hits have full precision scores and match the brute force top k.
• Without rescoring, scores differ from full precision.
• oversample = 1.0, normalized cosine, boost, similarity threshold, filter, filter matching no documents, and the num_candidates limit cap.
• Nested vector field rescoring.
• Validation errors for invalid oversample values and for byte fields.
• ExactVectorQuery with full_precision matches brute force full precision results. Without it, scores are quantized.

./gradlew :test --tests '*Vector*' --tests '*Knn*' --tests '*MultiRetriever*' --tests '*Retriever*'

[vim345]

Neat feature.