Module quant 

Training-free vector quantization. Training-free vector quantization.

Structs

Bit4Factors

Per-row scalars required by the four-bit unbiased dot estimator.

Bit4Query

Query-side signed-byte representation prepared once and reused per row.

Int8Query

Query-side signed-byte representation prepared once and reused per row.

Int8Vec

Borrowed signed-byte row and its per-vector affine reconstruction factors.

RescoreHit

One exact hit returned from the rescore stage.

RescorePool

Describes either dense coarse scores or a graph traversal’s retained sparse pool.

RescoreResult

Exact top-k hits, candidate count, and byte accounting for both stages.

SearchByteCounts

Stored-data bytes touched by one two-stage query.

Enums

QuantError

Typed failure from a training-free quantizer or estimator.

QuantScheme

Persisted quantization scheme identifier.

RescoreError

Typed failure from the two-stage rescore helper.

RescoreMetric

Full-precision ranking applied to the retained coarse pool.

Functions

dequantize_bit4

Reconstructs the norm-scaled selected four-bit code direction.

dequantize_int8

Reconstructs one affine signed-byte row into caller-owned storage.

dot_int8_query

Estimates a dot product against one affine Int8 row.

est_dot_bit4

Estimates a dot product directly from one packed four-bit row.

est_dot_bit4_batch

Estimates dot products for contiguous packed four-bit rows in one dispatch.

prepare_bit4_query

Prepares one query for repeated four-bit row scoring.

prepare_int8_query

Prepares one full-precision query for repeated affine Int8 row scoring.

quantize_bit4

Quantizes one row to four-bit Extended-RaBitQ codes.

quantize_int8

Quantizes one finite row to signed bytes using a per-vector affine map.

rescore_top_k

Selects a coarse frontier and exactly rescores its best k rows.