DocJev

No, it is an independent implementation and does not call or reuse them. LlamaParse appears only as an optional cloud OCR tier for hard inputs; LiteParse is the default and extracts page text locally. The Jev call is still hosted, so local OCR does not make the pipeline offline.

On a 40-document pilot both engines classified 40 of 40 originals correctly, and Jev split 7 of 8 packets exactly against 8 of 8 for the small-LLM baseline. The next 40-document challenge targeting scans and ambiguous boundaries is a preparation protocol, not another measured result.

DOCX and PPTX both need LibreOffice, while native PDF input does not. Page numbers refer to a retained PDF rendering: DOCX pagination follows that rendering and PPTX uses one page per slide. Every successful split covers each canonical page exactly once, and exports verify the canonical PDF hash.

The docs say not to. A segment's mean category probability is an average of page scores rather than a joint probability, and the selected-category probability and provider confidence are different values with neither claimed calibrated. What you can rely on is the review margin: a boundary score within 0.1 of the threshold flags a cut or continuation.

Requires Python 3.11+ and a TypeSafe API key: clone the repo, run uv sync, then uv run docjev doctor --smoke. Export TYPESAFE_API_KEY in your shell before running a Jev decision.

Run uv sync --extra demo --extra baseline, then uv run docjev demo and open http://127.0.0.1:8765; it defaults to original public documents with editable categories and live classification and splitting, and keys stay on the local server.

TYPESAFE_API_KEY is required; LLAMA_CLOUD_API_KEY and OPENAI_API_KEY are optional for cloud OCR and the baseline. Env vars use the JEV_DOCS_ prefix, the .jev-docs/ cache keeps its name, and the CLI does not auto-load .env files.