Release v1.12.0 · jztan/pdf-mcp

What's New in v1.12.0

pdf_search hybrid mode used to return a stale, pre-fusion total_matches (and page_match_counts) alongside the post-RRF matches array, producing self-contradicting payloads like matches=[5 items], total_matches=0. Both fields are now recomputed from the fused result set. Semantic mode now includes total_matches/page_match_counts so the schema is consistent across all three modes.
pdf_search keyword mode was effectively phrase-only because _escape_fts5_query wrapped the entire query in double-quotes. Multi-word queries like "pgvector latency" returned zero matches when the words appeared on the same page but non-contiguously. Queries are now tokenised; pages must contain all tokens (implicit FTS5 AND) and BM25 still ranks by combined frequency.
pdf_search auto mode crashed with a ToolError when fastembed was installed but the embedding model could not be loaded (offline machine, HF outage, etc.). It now degrades to keyword and surfaces semantic_unavailable=true plus a semantic_unavailable_reason string.
Heuristic section detector emitted body-paragraph snippets as section titles when a line started with a heading-shaped prefix (e.g. "Section 2: This paragraph discusses ..."). Lines longer than 200 chars are now rejected as heading candidates so no spurious sections are produced.

BREAKING: pdf_info.text_coverage shape changed from list[{page, text_chars, raster_images}] to a compact dict. By default it now contains only a constant-size summary (page-count rollups + truncated OCR candidate list) so payload size stays bounded regardless of page count — a 3000-page PDF no longer ships ~6000 ints just for coverage. Pass pdf_info(path, detail=True) to opt into the per-page parallel arrays text_chars_per_page and raster_images_per_page.

Per-match low_confidence flag plus response-level all_low_confidence and confidence_threshold on pdf_search semantic-mode responses, so agents can decide whether to trust top-k semantic results below the cosine threshold.

README: updated the pdf_info description to reflect the new text_coverage shape (summary by default, per-page arrays under detail=True).
Browser demo (pages/index.html, served at pdf-mcp.jztan.com): mock response and coverage visualizations migrated to the new compact text_coverage shape; demo footer bumped to v0.3.

pip install pdf-mcp==1.12.0