Search: SQ8 compression and training threshold for HNSW vector fields (Redis 8.12)

[mgravell]

Client support for RediSearch MOD-14958 (RediSearch/RediSearch#11330, merged 2026-09-22; VectorSimilarity MOD-14956/7/8/60): COMPRESSION SQ8 and TRAINING_THRESHOLD on HNSW vector fields.

What the server does

• COMPRESSION SQ8: 8-bit scalar quantization of stored vectors. HNSW only; FLOAT32/FLOAT16 with L2/IP/COSINE only.
• TRAINING_THRESHOLD n: vectors stay in the flat tier until n have arrived, then a mean is learned and they migrate into the compressed graph; searchable throughout. Default 10240, max 102400, explicit 0 = no training. Only valid with COMPRESSION.
• FT.INFO attributes gain compression and training_threshold (absent on uncompressed fields). Disk/Flex indexes reject both. Schema encoding v28.

What this PR adds

• HnswVectorField.CompressionAlgorithm and .TrainingThreshold, emitted after the graph parameters when set.
• AddHnswVectorField(..., compressionAlgorithm:, trainingThreshold:) as optional parameters appended after missingIndex, so existing positional callers compile unchanged. (Binary-breaking for compiled callers of the old signature, as any optional-parameter addition is; flagged for review.)
• VectorCompressionAlgorithm.SQ8. The enum stays shared across algorithms; the COMPRESSION token mapping is now one helper used by both HNSW and SVS-VAMANA.
• No client-side validation of type/metric/threshold combinations. The server owns the rules and its errors are clear; tests pin that they surface as RedisServerException:
  • COMPRESSION is only supported for FLOAT32 and FLOAT16 vector types
  • TRAINING_THRESHOLD is irrelevant when compression was not requested
• FT.INFO: nothing to do; InfoResult.Attributes is untyped per field and the new keys flow through.
• CI: Redis 8.12 added to both matrices (PR and nightly) via tests/dockers/.env.v8.12 pinned to redislabs/client-libs-test:8.12-m02-int. Same change as 8.12 CMS: support CELL_SIZE, fix CMS.INFO field parsing, cover negative INCRBY #562; whichever merges second is a no-op there.

new Schema().AddHnswVectorField("v", VectorType.FLOAT16, 768, VectorDistanceMetric.CosineDistance,
    compressionAlgorithm: VectorCompressionAlgorithm.SQ8, trainingThreshold: 2048);
// FT.CREATE idx SCHEMA v VECTOR HNSW 10 DIM 768 TYPE FLOAT16 DISTANCE_METRIC COSINE COMPRESSION SQ8 TRAINING_THRESHOLD 2048

Tests

• TestIndexingCreation_Hnsw_Sq8: argument rendering, with and without threshold.
• TestCreate_Hnsw_Sq8 (8.12+, standalone + cluster): create, FT.INFO round trip including the server default of 10240 and the absence of both keys on an uncompressed field, KNN below the training threshold.
• TestCreate_Hnsw_Sq8_ServerRejectsInvalidCombinations (8.12+): the two server errors above, plus an SVS-only compression value on HNSW.

Run locally against redislabs/client-libs-test:8.12-m02-int (reports redis_version:255.255.255), standalone and cluster, RESP2 and RESP3: 75/75 in the Search creation set. Library builds with 0 warnings on all TFMs. With the matrix change, the 8.12-gated tests now run in CI too.

Not done, deliberately

• No GlobalSQ8 for SVS-VAMANA (its scalar option); separate token, separate ticket if wanted.
• No typed FT.INFO surface for the new keys.