pdf-mcp

pip install pdf-mcp

Search is hybrid, combining BM25 keyword hits with semantic hits through reciprocal rank fusion, and it can run across a whole folder of documents with results attributed per file. Reading is paginated afterwards, so only the pages that matched are spent. The detail worth noting is content trust: text rendered invisibly, set at sub-point sizes, filled white on white or positioned off the page gets flagged rather than ingested as if a human had vetted it, which is the exact vector for prompt injection hidden in a document. Nothing is stripped, the flag is left for the agent to weigh. Charts are read from vector plot geometry as coordinate tables, Japanese vertical text is reordered geometrically, and results are cached in SQLite across restarts.