roaringrange (Python)

Build static, range-fetchable search datasets from Python, then search millions of records in the browser with no backend. These bindings wrap the core Rust build module, so the files they emit are byte-identical to the Go and Rust builders and are read by the same WASM reader. Two index types: a trigram text index (Builder) and a similarity / vector index (VectorBuilder).

What it produces

Builder.build(out_dir) writes the four files the text reader serves over HTTP Range; VectorBuilder.build(path) writes one .rrvi similarity index:

Upload them to S3/CloudFront and point the WASM reader at the URLs.

Install

Prebuilt abi3 wheels (one wheel for CPython 3.8+) are published to PyPI:

pip install roaringrange

CI builds and tests the extension on CPython 3.12, 3.13, and 3.14.

From source (dev)

cd python
maturin develop --release      # builds + installs into the active venv
# or: maturin build --release   # produces a wheel in target/wheels/

Requires a Rust toolchain and pip install maturin.

Usage

import roaringrange as rr, json

b = rr.Builder(gram_size=3)
for row in rows:                              # rows from a DataFrame, DB, JSONL, …
    b.add(
        rank=row["citations"],                # higher rank = listed first (doc-ID order)
        text=f'{row["title"]} {row["abstract"]}',   # tokenized into trigram keys
        record=json.dumps({"t": row["title"], "y": row["year"]}).encode(),
        facets={"year": [str(row["year"])], "type": [row["type"]]},  # field → categories
    )

stats = b.build("out/")        # writes out/index.rrs, index.rrf, records.idx, records.bin
print(stats)                   # BuildStats(docs=..., ngrams=..., fields=...)

rr.tokenize(text, gram_size=3) returns the n-gram keys a string maps to — useful for understanding why a query does or doesn't match.

Vector / similarity search

VectorBuilder trains an IVFPQ index over your embeddings and writes a single .rrvi file that the WASM reader range-fetches like the text index. Use the same doc_id as the text index so a vector hit maps to the same record (and can hybridize with trigram search). Vectors are L2-normalized for the default "ip" (cosine) metric.

import roaringrange as rr

vb = rr.VectorBuilder(dim=256, nlist=4096, m=32, metric="ip")  # m must divide dim
for doc_id, embedding in enumerate(embeddings):     # embeddings: any float sequences
    vb.add(doc_id, embedding.tolist())              # numpy row → list of floats
# or in one call: vb.add_many([(i, e.tolist()) for i, e in enumerate(embeddings)])

stats = vb.build("out/vectors.rrvi")
print(stats)   # VectorBuildStats(vectors=..., dim=256, nlist=..., m=32, nbits=8)

Parameters: nlist coarse clusters (≈ 4·√N, clamped to the vector count), m PQ subquantizers (must divide dim), nbits (1–8) → 2^nbits codes per subspace, metric "ip"/"cosine" or "l2". Training is deterministic (seed, kmeans_iters). One .rrvi per embedding model — each model is a different vector space. See ../VECTORS.md for the byte layout.

This pure-Rust trainer suits small/medium corpora and tests; at very large scale train with FAISS and export the same RRVI layout (the reader is identical).

Scale: train with FAISS, export to RRVI

For large corpora, train OPQ,IVF,PQ with FAISS and export the trained parts — no retraining in Rust. python/scripts/faiss_to_rrvi.py does this end to end (install the extra: pip install 'roaringrange[train]' for numpy + faiss-cpu):

from faiss_to_rrvi import export_to_rrvi
stats = export_to_rrvi(vectors, doc_ids, "vectors.rrvi", nlist=4096, m=32, metric="ip")

Under the hood it calls the low-level roaringrange.write_rrvi_from_faiss(...), which takes the FAISS arrays (OPQ rotation, coarse centroids, PQ codebooks, per-vector cluster + 8-bit codes) as little-endian byte buffers — so the wheel needs no numpy dependency. The export is verified against the Rust reader (recall@10 ≈ 0.9995 vs FAISS's own search on the same index).

Embedding text (mode 2: model2vec, no backend)

python/scripts/model2vec_embed.py embeds text with a model2vec static model (minishlab/potion-retrieval-32M, 512-d, mean-pooled token vectors — no transformer, fast on CPU) and builds a .rrvi. Install the extra: pip install 'roaringrange[embed]'.

from model2vec_embed import build_rrvi_from_texts
stats, _ = build_rrvi_from_texts(titles, doc_ids, "vectors.rrvi", nlist=256, m=32)

It's "mode 2" because the same model2vec recipe can run in the browser at query time, so similarity search needs no backend at all. The query embedding must use the identical model + pooling as the corpus, or the spaces won't match.

Notes

• Ranking is baked in. Doc IDs are assigned in descending rank, so the top-K of any query is free at read time (no query-time scoring). Pick a good rank signal (citations, holdings, popularity, …).
• Records are opaque. record= is raw bytes; the format never dictates your schema. Decode them however you like on the client.
• In-memory build. This builds the whole index in RAM — ideal for up to many millions of records. For corpora whose index exceeds memory, the core crate's chunked path (build::chunk) is the route; exposing it here is a follow-up.

MIT — see ../LICENSE.