Architecture

The Python library is the single implementation used by the CLI, benchmark harness, and local app.

1. documents.parse_document retains a canonical PDF, extracts every page with a selected OCR adapter, verifies coverage, and optionally caches the result.
2. engines.jev builds typed category and boundary questions over ordered page text. A whole packet normally fits one request. Larger packets use overlapping context windows with exactly one output owner per page.
3. classify maps a whole-document decision to the public classification schema. split creates a segment on a category change or source-document boundary, including adjacent same-category documents.
4. export copies the exact source pages into per-segment PDFs. Office inputs export their retained canonical PDF pages.
5. engines.openai provides a direct, terse structured-output baseline. Its practical splitter returns all segments in one request when supported.

No LlamaIndex Classify or Split API is called. The optional LlamaParse adapter only calls file upload and Parse.

Jev Choice confidence and winning-class probability are separate fields. Segment mean probability is only an average of page probabilities, not a calibrated joint probability. The small-LLM baseline emits no confidence value.

Results include client wall-clock stage timings and per-request usage. Network, queueing, and provider processing are part of request latency; these are not isolated GPU inference measurements. Unknown charges remain unknown.

Pages and context

Public page numbers are one-based. A DOCX page refers to the retained PDF rendering; a PPTX page refers to its rendered slide. Blank pages stay in coverage. Empty OCR text must pass a visual blank check before the page can be called blank.

Jev has separate state-plus-longest-question and complete-request budgets. Local UTF-8 size checks are approximate. A whole classification that is too large fails visibly; splitting can shrink windows, but never truncates individual page text.