* fix: prevent KeyError crash and context exhaustion in TOC processing
- Use .get() with safe defaults for all LLM response dict accesses
- Optimize extract_toc_content retry loop to grow chat_history
incrementally instead of rebuilding with full accumulated response
- Optimize toc_transformer retry loop to use chat_history instead of
re-embedding the entire raw TOC and incomplete JSON in each prompt
- Return best-effort results on max retries instead of raising
- Add 14 mock-based tests covering all fix scenarios
Closes#163
* fix: address review feedback on retry behavior and None guard
- Restore explicit Exception on max retries instead of silent warning
- Move truncation logic before the retry loop so it only runs once
on the initial incomplete response, not on every iteration
- Add explicit None guard for physical_index before passing to
convert_physical_index_to_int to prevent potential TypeError
- Update test to expect Exception on max retries
---------
Co-authored-by: Your Name <you@example.com>
litellm 1.83.7 hard-pins python-dotenv==1.0.1, which conflicts with
python-dotenv==1.2.2 in requirements.txt and makes a fresh install fail
(ResolutionImpossible under both pip 25.2 and uv). Downgrading dotenv to
1.0.1 is not an option: python-dotenv < 1.2.2 is affected by
GHSA-mf9w-mj56-hr94 (moderate).
litellm 1.84.0 relaxed its pin to python-dotenv<2.0,>=1.0.0, allowing the
patched python-dotenv==1.2.2 to remain. Full requirements.txt resolves
cleanly under both pip and uv.
Closes#286
Added functions to sanitize and secure document text against prompt injection patterns. Implemented validation for physical indices in the table of contents.
- Add Website and Book a Demo buttons; reorder and recolor to brand palette
- Replace removed/invalid shields logos (LinkedIn, envelope) and the Website
icon with inline white SVGs so all badges show an icon
- Reword header tagline (No Vector DB, No Chunking; Context-Aware Retrieval; Human-like)
- Rename top-nav Homepage -> Website
extract_nodes_from_markdown now matches `**Title**` lines as level-1
headings (alongside ATX `#` headings) and attaches the heading level
on the producer side. extract_node_text_content reads the level from
the node instead of re-running a `^#{1,6}` regex on the source line,
which was silently dropping bold-heading nodes from OCR / MinerU output.
Bold maps to level 1 even when mixed with `#` / `##` / `###` — bold-as-
heading is a courtesy heuristic for non-ATX markdown sources, and
CommonMark has no concept of bold heading depth.
Weekly scan to open PRs upgrading third-party actions used in CI.
pip dependencies remain pinned and are covered by Dependabot security
updates separately.
