Repository navigation
Document answers now carry exact quotations back to their original source text: source identity, physical PDF page and character span. complete_documents and acomplete_documents validate those references for accepted root and child outputs, with explicit unknowns and bounded search.
Also included:
- Local text/PDF loading with preserved physical page boundaries and an optional
documentsextra. - An explicit long-document companion profile, with aggregate limits covering non-image PDF resources, nested Forms and font mappings.
- Standalone strict-schema JSON finals and bounded recovery from completed empty Codex CLI responses, preserving usage.
- A document CLI and structured
DocumentResult/ JSONL export.
Validation: 530 local tests passed with 94.86% branch-aware coverage. Fresh wheel and source archives passed inspection, isolated installation, worker and PDF checks. The repeated live matrix completed 48/48 attempts on a complete 157-page report and official Python documentation: all passed field and structural-citation checks, 46 passed the frozen strict evidence criteria. Both strict misses remain published with manual review notes.
Deep validation and failure cases · Initial pilot · Usage and limits
The old API with the same custom validator also rejects the controlled source corruptions. This release integrates that contract into document handling. Fixed-response repairs use supplied corrected replies; they are not a model-quality measurement. Quote matching does not prove entailment or completeness. No live matrix attempt used child calls, so the results do not establish a recursion advantage.
Install the attached wheel with optional PDF support, or use the tagged source:
pip install "recursive-llm[documents] @ git+https://github.com/grishahq/recursive-llm.git@v0.6.0"This is a GitHub release. PyPI remains v0.5.0; its separate manual publication workflow was not run.