lateweave is a Rust library, with Python bindings, for composing retrieval pipelines out of engines it does not implement:

Query -> CandidateGenerator -> [Reranker] -> deterministic top-k

A Query is raw text plus named features, each stamped with the Representation (encoder identity) that produced it and materialized at most once. Stages name documents by their corpus and the ID the system of record gives them, never by an engine's internal position, so each stage reads its own indexes and whoever maintains them, in any process and with any library, never coordinates with lateweave or with the other stages. Every stage result says when the index it read last committed; a search ranks only the documents every stage holds and reports the oldest of those commits as as_of. The reranker is optional: without one, gather scores rank.

The package supplies:

• Query, Feature, Candidate, Gathered, Scored, Representation;
• the CandidateGenerator, Reranker, and MultiVectorSource contracts;
• SearchPipeline: compatibility checks, optional rerank, subset enforcement, deterministic ranking, freshness, provenance, timings;
• MaxSimReranker over one MultiVectorSource per corpus, backed by a packed CPU MaxSim kernel (SGEMM, SIMD maximum reduction, bounded token batches, worker pools);
• VectorStore, a reader of memory-mapped multi-vector stores for gatherers, such as BM25, that keep no document vectors, and VectorStoreWriter, one writer of their documented format.

All of it is implemented once, in the lateweave crate, which builds without PyO3 or NumPy. The Python package is a binding over that crate: a Python gatherer, reranker, or source is adapted to the Rust contracts, and the pipeline, MaxSim, and stores run natively.

Nothing in the package names an engine. bm25s, FastPLAID, NextPlaid and others appear only in cookbooks.

Installation

Python

pip install lateweave

Wheels are published for CPython 3.11 to 3.14 on:

Nothing needs to be installed on the host: the Linux wheels carry their own OpenBLAS. Elsewhere pip builds from the source distribution, which needs a Rust toolchain and uses the bundled pure-Rust SGEMM (see Building from source).

Rust

cargo add lateweave

The crate needs Rust 1.85 or later and, by default, no system library. On Linux, enable the openblas feature for the faster kernel if the machine has libopenblas.so.0:

[dependencies]
lateweave = { version = "0.1", features = ["openblas"] }

Layout

Cargo.toml          the `lateweave` crate (workspace root)
src/                pipeline, stages, MaxSim kernel and reranker, stores
examples/           Rust usage
bindings/python/    `lateweave-python`: the PyO3 module `lateweave._native`
python/lateweave/   the Representation dataclass, protocols, type stubs
STORE_FORMAT.md     the vector store on disk, for writers outside lateweave

Using it from Rust

use std::sync::Arc;
use lateweave::{
    Feature, MaxSimReranker, MultiVectorSource, Query, SearchPipeline, SearchRequest,
    TokenMatrix, VectorStore, DEFAULT_FEATURE,
};

let store = Arc::new(VectorStore::open("index/vectors")?);
let representation = store.representation().clone();
let gatherer = Arc::new(MyLexicalIndex::open("index/lexical")?);  // impl CandidateGenerator
let reranker = MaxSimReranker::new([store as Arc<dyn MultiVectorSource>], DEFAULT_FEATURE)?;
let pipeline = SearchPipeline::new(gatherer, Some(Arc::new(reranker)));

let query = Query::new("prescripción de una deuda tributaria").with_feature(
    DEFAULT_FEATURE,
    Feature::lazy(representation, move || {
        TokenMatrix::new(encode(&text), dimension)
    }),
);
let result = pipeline.search(&query, &SearchRequest::new(500, 100))?;
println!("reflects every write committed before {:?}", result.as_of);
for document in &result.documents {
    println!("{} {:?} {}", document.rank, document.key, document.score);
}

Stage and source errors from outside lateweave travel as Error::External. examples/stored_maxsim.rs is a complete program: cargo run --example stored_maxsim.

Building from source

cargo test                 # the Rust library and its example
pip install maturin
maturin build --release    # the Python wheel

That default build has no external library dependency: the MaxSim kernel's SGEMM comes from the bundled pure-Rust matrixmultiply, so the library links and the extension imports on any host. CONTRIBUTING.md has the full development setup.

For a deployment, take the OpenBLAS kernel instead, as the published Linux wheels do: about 1.75x faster on the shapes this kernel sees, bit-identical results. Rust dependents enable the crate's openblas feature; the wheel forwards the same feature:

maturin build --release --features pyo3/extension-module,openblas

maturin runs auditwheel's repair step, which vendors libopenblas.so.0 into lateweave.libs/ inside the wheel. So this is self-contained too: it needs OpenBLAS on the machine that builds the wheel, not on the machines that install it. Set OPENBLAS_LIB_DIR if libopenblas.so.0 lives outside the linker's default search path.

Two things to know about that command line:

• pyo3/extension-module must be repeated. --features replaces the list in [tool.maturin] instead of extending it, and without extension-module pyo3 links libpython, which auditwheel then bundles into the wheel.
• The library lateweave needs exports the plain Fortran symbol sgemm_. scipy-openblas32 on PyPI does not qualify -- it is a real threaded OpenBLAS, but every symbol carries a scipy_ prefix.

macOS needs no feature flag: Accelerate ships with the OS and is used automatically. ARCHITECTURE.md explains why Linux does not have an equivalent default.

Queries and features

from lateweave import Feature, Query, Representation

representation = Representation(
    encoder="lightonai/LateOn-Code", encoder_revision="main", dimension=128, normalized=True
)
query = Query(
    "CUDA_ERROR_ILLEGAL_ADDRESS after switching to bf16 attention",
    multi_vector=Feature(representation, provider=lambda: encode(text)),
)

provider runs the first time a stage asks for the feature, so a text-only gatherer never pays for encoding, and a gatherer and reranker share one matrix. A stage that needs the feature from another encoder is refused before anything runs.

Implementing a gatherer

from lateweave import Candidate, Gathered


class MyGatherer:
    requires = {}                      # consumes query.text only
    score_semantics = "my-gather-score"

    def __init__(self, engine, corpus):
        self.engine = engine
        self.corpus = corpus

    def gather(self, query, limit, *, subset=None):
        searcher = self.engine.searcher()          # one consistent state of the index
        accept = None if subset is None else lambda id: subset.allows(self.corpus, id)
        hits = searcher.search(query.text, limit=limit, accept=accept)
        return Gathered(
            [Candidate(self.corpus, hit.id, hit.score, rank, "my-engine") for rank, hit in enumerate(hits)],
            searcher.committed_at,                 # an aware datetime
        )

A candidate names its corpus and its document ID; how the engine finds either is its own business. as_of is when the index state the gatherer read was committed: every write committed before it is reflected in the candidates. It is the commit time, not the time the gatherer reloaded.

subset is None, or a lateweave.Subset naming, per corpus, either the documents the search may return (Subset().including(corpus, ids)) or the ones it must not (.excluding(corpus, ids)); subset.allows(corpus, id) answers for one document and subset.restriction(corpus) returns ("only" | "except", ids). A corpus it does not name contributes nothing, and an ID the index does not hold is simply not a candidate. A gatherer that cannot honour it must raise; the pipeline refuses a candidate outside the subset. A gatherer that consumes a vector feature declares it: requires = {"multi_vector": representation}. requires and score_semantics are read once, when the SearchPipeline is built.

One gatherer may search several corpora, such as two indexes whose hits it fuses, reporting the older of their commits.

Freshness

Indexes move while a service searches them, each at its own pace. The pipeline needs nothing from whoever moves them beyond what every stage reports:

• A document is ranked only when every stage holds it. A candidate the reranker's index does not hold scores None and is dropped, counted in diagnostics["dropped"]. So a delete is served as soon as any index applies it, and an insert once all of them have.
• result.as_of is the oldest commit among the indexes the search read: every write committed to all of them before it is reflected in the result. What to do with an answer older than a service tolerates, whether to fail, warn, or serve it anyway, is the service's policy:

from datetime import datetime, timedelta, timezone

result = pipeline.search(query, gather_limit=500, limit=100)
if datetime.now(timezone.utc) - result.as_of > timedelta(minutes=1):
    ...

An index that is idle still has to say it is current, or it looks older than it is: a writer commits periodically even with nothing to write, as VectorStoreWriter.commit() does.

Implementing a reranker or a source

A reranker owns whatever it needs to score. When what it needs is the token vectors of candidate documents, implement MultiVectorSource over them, one per corpus, and let MaxSimReranker do the scoring:

from lateweave import MaxSimReranker


class EngineVectors:
    """Token vectors an engine already holds; no second copy."""

    representation = representation
    score_semantics = "engine-reconstructed-full-maxsim"

    def __init__(self, engine, corpus):
        self.engine = engine
        self.corpus = corpus

    def view(self):
        return EngineView(self.engine.snapshot())   # one consistent state, with .as_of


class EngineView:
    def __init__(self, snapshot):
        self.snapshot = snapshot
        self.as_of = snapshot.committed_at

    def document_lengths(self, document_ids):
        return {item: self.snapshot.length(item) for item in document_ids if item in self.snapshot}

    def fetch(self, document_ids, *, threads=None):
        rows = [self.snapshot.vectors(item) for item in document_ids]   # float32 [tokens, D]
        return np.concatenate(rows), np.asarray([len(r) for r in rows], dtype=np.int64)


reranker = MaxSimReranker([EngineVectors(laws_engine, "laws"), EngineVectors(cases_engine, "cases")])

Each rerank takes one view of every source and reads it throughout. A document a view leaves out of document_lengths is unscored; a candidate from a corpus with no source is an error. A VectorStore passed as a source is read natively, without calling back into Python.

Or write a reranker directly:

from lateweave import Scored


class MyReranker:
    requires = {"multi_vector": representation}
    score_semantics = "my-qualified-score-semantics"

    def rerank(self, query, candidates, *, budget):
        vectors = query.feature("multi_vector", representation)
        state = self.index.snapshot()
        return Scored([state.score(vectors, c.document_id) for c in candidates], state.committed_at)

The pipeline requires exactly one score or None per candidate, in candidate order, never NaN.

Stores

A VectorStore holds the vectors of one corpus for gatherers without document vectors, and carries the Representation of its vectors, so it refuses queries from any other encoder. lateweave only reads stores; the format is plain .npy files, document IDs as JSON, and manifest.json, so any process can write one. VectorStoreWriter is the writer lateweave ships:

from lateweave import MaxSimReranker, SearchPipeline, VectorStore, VectorStoreWriter

# The indexing process.
writer = VectorStoreWriter.create("index/vectors", "laws", representation, encoding="float32")
writer.append(ids, packed_embeddings, lengths)     # replaces any document with the same ID
writer.delete(["law-17"])
writer.commit()                                    # publishes both together
writer.compact()                                   # whenever convenient: reclaims space only
writer.commit()

# The serving process.
pipeline = SearchPipeline(gatherer, MaxSimReranker([VectorStore("index/vectors")]))

Appends write a new segment and deletes a tombstone file; nothing is rewritten until a compaction, which changes no result. VectorStore.view() returns the last commit, loading only what it has not seen; views taken earlier keep reading their own commit.

See ARCHITECTURE.md for the ownership rules and cookbook/README.md for the BM25 recipe.

Contributing

Issues and pull requests are welcome; see CONTRIBUTING.md for the development setup and CHANGELOG.md for what changed in each release. Please report security issues privately, as described in SECURITY.md.

License

Licensed under the Apache License, Version 2.0. The batching and SIMD reduction in the MaxSim kernel derive from maxsim-cpu; see NOTICE.