tors

Fast, GIL-free text and document operations for Python, backed by Rust: normalization, Unicode segmentation, diffing, fuzzy and phonetic matching, multi-pattern search and redaction, chunking, and lightweight retrieval (TF-IDF, BM25, SimHash, Merkle integrity). An optional extra adds cross-format document extraction (PDF, Office, RTF/ODF/EPUB, CSV, HTML to markdown or plain text). In the spirit of orjson for JSON or polars for dataframes.

Full documentation lives at azx-pbc-oss.github.io/tors and in docs/ in this repo: the API reference, Documents, Performance, Async use, Design and scope, Dependencies and licensing, and three worked recipes.

Why

Python's re module and str methods never release the GIL, no matter how large the input is: a multi-megabyte text transform runs as one long GIL-held call that stalls every other thread and the asyncio event loop for its whole duration. tors does the same kind of transform as a single native Rust pass, wrapped in py.detach (PyO3's GIL-release call) for the entire computation, so the GIL is free for the rest of your program while it runs.

That covers the functions with a stdlib equivalent (normalize's pipeline mirrors unicodedata.normalize + a few re.sub calls; quote/unquote mirror urllib.parse; b64_encode_bytes/b64_decode mirror base64; the one-shot hashing family mirrors hashlib/hmac; the random-generation family mirrors secrets/uuid). Where the stdlib has no equivalent at all (Unicode text segmentation: grapheme clusters, word and sentence boundaries; leftmost-longest multi-pattern search; edit-distance and phonetic matching; content-defined chunking; SimHash near-duplicate detection; Merkle tree integrity; encoding detection), tors supplies it over maintained Rust crates, GIL-released the same way, so async services and threaded pipelines don't pay a blocking tax for text work.

GIL-release is not the same as free-threaded-build support: tors wheels target the standard CPython ABI (cp310-abi3), which a free-threaded 3.13t interpreter cannot load — free-threaded support starts at 3.14t (source build; PyO3's floor), and GIL-release is what makes that floor irrelevant for the standard build: the GIL is already free for the rest of your program on every CPython tors installs into.

Install

pip install tors

Building from source (a Rust toolchain and maturin):

pip install maturin
maturin develop --release

The document-extraction surface is a second wheel behind an extra: pip install "tors[documents]". From a checkout, uv sync --locked --extra documents builds and installs the payload wheel from this tree. See Documents.

The underlying Rust crate is also on crates.io, published separately as tors-core (the plain tors name belongs to an unrelated, dormant crate). cargo add tors-core, then use tors::... in code: [lib] name in Cargo.toml keeps the importable crate name tors regardless of the published package name.

Pyodide / WebAssembly

The Rust cores are OS-free, so tors also builds for Pyodide as a PEP 783 pyemscripten wheel: the abi3-py310 story carries over unchanged, so one wheel covers every Pyodide Python >= 3.10. The WASM workflow builds it on every push (artifact only; publishing to PyPI's Emscripten platform is not wired up). To build one locally:

rustup target add wasm32-unknown-emscripten
uvx --python 3.14 --from pyodide-build pyodide build . -o dist

The driving interpreter matters: Python 3.14 targets Pyodide 314.x / pyemscripten_2026_0 with a stable Rust toolchain; driving from 3.13 pins a Rust nightly older than this crate's MSRV. SIMD-dependent dependencies (base64, simdutf8, memchr, aho-corasick) compile their scalar fallbacks for wasm; the extension has been smoke-tested (import + representative calls across every core family) in Pyodide under node.

Quickstart

import tors

tors.normalize("line one  \n\n\n\nline two\r\n")
# 'line one\n\nline two'

tors.finalize("line one  \n\n\n\nline two\r\n")
# ('line one\n\nline two', 'e986ba083c7c1a9361143d2d8ccd8477d1d5eeef8b94b67c6ad4693f8f7b942a')

tors.diff_opcodes_lines("l1\nl2\nl3\n", "l1\nX\nl3\nl4\n")
# [("equal", 0, 1, 0, 1), ("replace", 1, 2, 1, 2), ("equal", 2, 3, 2, 3),
#  ("insert", 3, 3, 3, 4)]

thread = (
    "Ana: kickoff at nine.\n"
    "Ben: We briefed the U.S. team on the numbers. They asked for a follow-up.\n"
    "Ana: done."
)
chunks = tors.chunk_hierarchical(thread, 40, ["\n", None])
# [(0, 21), (22, 59), (59, 95), (96, 106)]

[thread[s:e] for s, e in chunks]
# ['Ana: kickoff at nine.',
#  'Ben: We briefed the U.S. team on the ',
#  'numbers. They asked for a follow-up.',
#  'Ana: done.']

corpus = [
    "the quick brown fox jumps over the lazy dog",
    "a lazy cat sleeps all day",
    "the fox and the dog are friends",
]
tors.bm25_rank("quick fox", corpus)
# [(0, 1.3162195220480066), (2, 0.4798180901812613), (1, 0.0)]

All of these run against the built extension; the output above is what they actually return.

What's inside

107 functions plus two small helper classes and six pinned constants (five charset alphabets and the key-family tuple), grouped by what they do; the documents extra adds seven document-extraction functions and its own helper types. Full signatures, argument contracts, and edge cases are in the API reference.

Highlights

• GIL release on every call. 12 MiB of prose through tors.finalize in a background thread holds the event loop's worst heartbeat gap to 10-14 ms; the pure-Python equivalent holds it for 92-108 ms. (The one exception: utf8_byte_len deliberately does not detach — its whole body is the borrow, and a detach bracketing no work starves a co-resident loop's heartbeat; see API reference.) Full measured tables: Performance.
• Async where it matters. tors.aio wraps exactly the large-input functions in asyncio.to_thread, unconditionally, with no size-based branch; everything else keeps one sync spelling. Details: Async use.
• Stateless by design. No pipeline objects, no caches, no handles to manage; the build-once CompiledLemmaDict and CompiledPatterns handles are the measured exceptions, and the scope cuts (no regex engine, no bundled lemma data, no search index) are decisions, not oversights. Details: Design and scope.
• Permissive-only dependency tree, machine-checked by cargo deny on every push. The full table and license accounting: Dependencies and licensing.

Contributing

See CONTRIBUTING.md for setup, the pre-PR checklist (tests, clippy, fmt, ruff, the license gate), fuzzing, release process, and commit-message conventions. For security vulnerabilities, see SECURITY.md instead of filing a public issue.