Crate nabled_embeddings 

Lightweight, ndarray-native, Arrow-zero-copy compute and rerank layer for embedding vectors.

Bring vectors from any model, compute exactly, deploy anywhere. nabled-embeddings composes existing nabled-linalg vector kernels and nabled-ml PCA into the post-embedding numerics a retrieval pipeline needs: normalize, score, rerank, exact brute-force kNN, and PCA compression.

What this is (and is not)

This is the exact rerank / compute layer that sits next to a vector store, not a vector database. It is deliberately:

• bring-your-own-vectors: it accepts Array2<T> / ArrayView2<T> (f32/f64 via NabledReal). Any encoder’s dense float vectors plug in unchanged because the crate depends only on shape and dtype, never the model.
• numerics-only: no model inference, no tokenizers, no ONNX/candle.
• Arrow-free in the core: Arrow/Lance ingress lives at the nabled::arrow facade and in pynabled, never here.

Non-goals: it is not an ANN/vector-search engine (no HNSW/IVF index), not storage, and not a FAISS/HNSW replacement. Use an ANN index for recall, then rerank its candidates here.

Two rules the math cannot enforce

1. Query and corpus must come from the same model and share the same dim.
2. You pick the Metric to match how the model was trained: Metric::Cosine (the default), Metric::Dot for maximum-inner-product (MIPS) models, or Metric::L2 where applicable. Dot on un-normalized vectors favors larger-norm rows by design.

Feature flags

1. blas: forwards BLAS acceleration to nabled-linalg/nabled-ml.
2. lapack-provider: enables provider-dispatched PCA paths.
3. openblas-system/openblas-static/netlib-system/netlib-static/magma-system: provider backends forwarded to the lower crates.

Quick example

use nabled_embeddings::{Metric, rerank};
use ndarray::arr2;

let corpus = arr2(&[[1.0_f64, 0.0], [0.0, 1.0], [0.9, 0.1]]);
let query = arr2(&[[1.0_f64, 0.0]]);
let top = rerank(&query.row(0), &corpus.view(), 2, Metric::Cosine)?;
assert_eq!(top[0].index, 0);

Re-exports

pub use cache::CorpusWorkspace;

pub use compress::PcaModel;

pub use compress::compress;

pub use compress::compress_into;

pub use compress::compress_view;

pub use compress::fit_pca;

pub use error::EmbeddingError;

pub use knn::brute_force_knn;

pub use metrics::mean_reciprocal_rank;

pub use metrics::ndcg_at_k;

pub use metrics::recall_at_k;

pub use metrics::reciprocal_rank;

pub use mmr::mmr;

pub use normalize::normalize_rows;

pub use normalize::normalize_rows_into;

pub use normalize::normalize_rows_view;

pub use quantize::QuantizedMatrix;

pub use quantize::dequantize;

pub use quantize::quantize_rows;

pub use rerank::batch_rerank_with_ids;

pub use rerank::rerank;

pub use rerank::rerank_with_ids;

pub use similarity::Metric;

pub use similarity::query_corpus_scores;

pub use similarity::query_corpus_scores_into;

pub use similarity::query_corpus_scores_view;

pub use topk::Neighbor;

pub use topk::NeighborWithId;

pub use topk::attach_ids;

pub use topk::top_k;

Modules

cache

Reusable corpus workspace for repeated queries against a static corpus.

compress

PCA-based dimensionality reduction for embedding matrices.

error

Error taxonomy for embedding retrieval compute.

knn

Exact brute-force k-nearest-neighbors over a full corpus.

metrics

Offline retrieval-quality metrics over ranked id lists.

mmr

Maximal Marginal Relevance (MMR) re-ranking for result diversification.

normalize

Row-wise L2 normalization for embedding matrices.

quantize

int8 row quantization for embedding matrices (encode/decode + dequantized scoring).

rerank

Exact single-query rerank: score candidates, then select the best k.

similarity

Query-versus-corpus scoring across the three common retrieval metrics.

topk

Direction-aware partial top-k selection over a score vector.