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:

file format contents
index.rrs RRSI trigram text index (popularity-split postings)
index.rrf RRSF facet sidecar (field → category → doc-ID bitmap, with counts)
records.idx / records.bin RRSR per-doc record bytes (your encoding)
*.rrvi RRVI IVFPQ similarity index (range-fetched coarse clusters + PQ codes)

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.

Project links

Data verified by PyPI on Jul 15, 2026
Data provided by the project maintainers, verified at the time the release was uploaded to PyPI.

Key dates

PyPI data
Data sourced directly from PyPI's database.
  • Released:
Newer release available (0.42.0)

1 maintainer

PyPI data
Data sourced directly from PyPI's database.
Avatar for freeeve from gravatar.com freeeve

Credits

Author: freeeve

GitHub Statistics

Data verified by PyPI on Jul 15, 2026
The GitHub source repository was provided by the project maintainers and verified by PyPI at the time of upload. Stars, forks, and open issues/PRs are derived from that repository and have not been independently verified.

License

MIT License (MIT)

Requires

Python >=3.8

Provides Extra

embed gemma train

Tags

search full-text roaring static-site serverless

Classifiers

Report project as malware

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distributions

No source distribution files available for this release.See tutorial on generating distribution archives.

Built Distributions

If you're not sure about the file name format, learn more about wheel file names.

roaringrange-0.39.1-cp38-abi3-win_amd64.whl (479.7 kB view details)

Uploaded CPython 3.8+Windows x86-64

roaringrange-0.39.1-cp38-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (589.2 kB view details)

Uploaded CPython 3.8+manylinux: glibc 2.17+ x86-64

roaringrange-0.39.1-cp38-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl (559.5 kB view details)

Uploaded CPython 3.8+manylinux: glibc 2.17+ ARM64

roaringrange-0.39.1-cp38-abi3-macosx_11_0_arm64.whl (523.7 kB view details)

Uploaded CPython 3.8+macOS 11.0+ ARM64

roaringrange-0.39.1-cp38-abi3-macosx_10_12_x86_64.whl (541.2 kB view details)

Uploaded CPython 3.8+macOS 10.12+ x86-64

File details

Details for the file roaringrange-0.39.1-cp38-abi3-win_amd64.whl.

File metadata

  • Download URL: roaringrange-0.39.1-cp38-abi3-win_amd64.whl
  • Upload date:
  • Size: 479.7 kB
  • Tags: CPython 3.8+, Windows x86-64
  • Uploaded using Trusted Publishing? Yes
  • Uploaded via: twine/6.1.0 CPython/3.13.12

File hashes

Hashes for roaringrange-0.39.1-cp38-abi3-win_amd64.whl
Algorithm Hash digest
SHA256 1568e096ae37c57e4f210998322d754e074baf7d2e3d2b33f75472976abc84ed
MD5 cb807701e7f268a91c139dd4874906ce
BLAKE2b-256 7948f024e28fdff223557d14bb8f24fe1c8b40979b02849fa26456d2d534fff8

See more details on using hashes here.

Provenance

The following attestation bundles were made for roaringrange-0.39.1-cp38-abi3-win_amd64.whl:

Publisher: release.yml on freeeve/roaringrange

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

File details

Details for the file roaringrange-0.39.1-cp38-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.

File metadata

File hashes

Hashes for roaringrange-0.39.1-cp38-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
Algorithm Hash digest
SHA256 e6382d67b1434395871c341db8b2dfdbd49aae2596573f76e70c1fce4219a595
MD5 318b754ebb1e6b046c3edc3eec6145dd
BLAKE2b-256 f47a12707ee284cf124c7b305d633a3730b9e7dedced82fb8efbf6f621a94b14

See more details on using hashes here.

Provenance

The following attestation bundles were made for roaringrange-0.39.1-cp38-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl:

Publisher: release.yml on freeeve/roaringrange

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

File details

Details for the file roaringrange-0.39.1-cp38-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl.

File metadata

File hashes

Hashes for roaringrange-0.39.1-cp38-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl
Algorithm Hash digest
SHA256 30ffb619282046676e7f16998c3b51ac41297c9ae33a601e5e54f44ef4bf34e0
MD5 309a7c2ffa854fa5e286eadf7c6ae2d7
BLAKE2b-256 b494b2f944f5fa4489e8f35baeb8236bebeafc2836e481b2510516b032845b32

See more details on using hashes here.

Provenance

The following attestation bundles were made for roaringrange-0.39.1-cp38-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl:

Publisher: release.yml on freeeve/roaringrange

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

File details

Details for the file roaringrange-0.39.1-cp38-abi3-macosx_11_0_arm64.whl.

File metadata

File hashes

Hashes for roaringrange-0.39.1-cp38-abi3-macosx_11_0_arm64.whl
Algorithm Hash digest
SHA256 73e98c9fb3a2a469de76c12404418fb584e9a73187819d9d626919b815300803
MD5 05dedbf425ff4d40f10b3e8a1c62c641
BLAKE2b-256 91145c04db1222f4f739372f0c36ff5119b743791f77341f47389a5685f5b86a

See more details on using hashes here.

Provenance

The following attestation bundles were made for roaringrange-0.39.1-cp38-abi3-macosx_11_0_arm64.whl:

Publisher: release.yml on freeeve/roaringrange

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

File details

Details for the file roaringrange-0.39.1-cp38-abi3-macosx_10_12_x86_64.whl.

File metadata

File hashes

Hashes for roaringrange-0.39.1-cp38-abi3-macosx_10_12_x86_64.whl
Algorithm Hash digest
SHA256 4e7c62739216db2978c88deec399d037b223ee632c6d89f62bb375af753b7f23
MD5 e3b92759e74555e3f599d173c811fb98
BLAKE2b-256 90ab92b1417dc869b8bf6146e2ce1d94a7caef15f0cdc74ed5aa87d61c14f632

See more details on using hashes here.

Provenance

The following attestation bundles were made for roaringrange-0.39.1-cp38-abi3-macosx_10_12_x86_64.whl:

Publisher: release.yml on freeeve/roaringrange

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

Release history Release notifications | RSS feed

0.42.0

5 files

0.41.0

5 files

0.40.0

5 files

This release

0.39.1 This release

5 files

0.39.0

5 files

0.38.1

5 files

0.38.0

5 files

0.37.1

5 files

0.37.0

5 files

0.36.0

5 files

0.35.0

5 files

0.34.0

5 files

0.33.0

5 files

0.32.0

5 files

0.31.0

5 files

0.30.0

5 files

0.29.0

5 files

0.27.0

5 files

0.26.0

5 files

0.25.0

5 files

0.24.3

5 files

0.24.2

5 files

0.1.1

5 files

0.1.0

5 files

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page