tors-documents

The tors.documents payload wheel: document-format extraction to GitHub-Flavored Markdown or plain text, GIL-free. PDF, the office and text formats (doc/docx, xls/xlsx, ppt/pptx, rtf, odt/ods/odp, epub, csv/tsv), and HTML, converted on the engine measured best for each family (pdf_oxide for PDF, anydoc for the office/text families, html-to-markdown-rs for HTML, office_oxide as the caller-selectable backend="oxide" lane). The whole read + sniff + convert pass runs under one py.detach, so the GIL stays free for the rest of your program while a large document converts.

Install

pip install tors[documents]

or this wheel directly:

pip install tors-documents

The base tors wheel ships no engines at all; this payload is what pip install tors[documents] pulls in. The split is the lazy-import design too: import tors never touches this wheel, and import tors.documents peaks under 19 MB in a fresh process despite all four engines compiled into the one payload .so (demand paging). The measured import-cost numbers are in the tors.documents section of the full API reference (linked below).

Import path

import tors.documents

The base wheel re-exports this payload as tors.documents, with an install hint pointing at the extra when this wheel is absent, so that spelling is the surface to target. tors_documents is the distribution name, not the API.

Version lockstep with tors

This wheel carries the same version number as the base tors wheel, and the two release together (one release PR bumps both manifests; see the repository's release-please-config.json). The tors[documents] extra resolves this wheel; the matching versions install side by side.

Documentation

• Documents: what the surface is, the engine table, and why the split into two wheels
• The full API reference: to_markdown/to_text/sniff/pdf_extract/pdf_page_count/pdf_classify/pdf_link_uris, the Backend/Format/PageKind enums, PdfClassification, NeedsOcrError, and tors.documents.aio