Visual RAG for technical PDFs
Search and Q&A over product catalogues and technical PDFs with complex layouts: drawings, tables and dimension callouts. It finds the right page by how it looks, answers from the page image with a vision model and cites the page.
How it works
- Searches page images and the PDF text together; adding the text search finds the right page more often than image search alone.
- Seven search modes were compared on test questions with known answer pages. A paid query-rewriting step did not help, so it is off by default.
- Web UI and API: choose a search mode per question, and every answer cites its page.
Thin layer over PixelRAG (StarTrail): the search modes, combined image and text search, testing and UI are mine.
- Python
- FAISS
- BM25 + RRF
- PixelRAG
- Gemini
- Claude