Cited PDF Chat
Chat with any PDF, scanned ones too. Every sentence of the answer links to the exact passage, highlighted on the page.
- 23/23 correct on the built-in test set, digital and scanned
- EN + 中文 questions and OCR in both languages
- 0 made-up quotes shown: unverified citations are flagged
The problem
PDF chatbots answer confidently and cite loosely. When a citation points at the wrong page, or at a sentence that doesn’t exist, nobody can trust the answer. That matters most for the reports and contracts people actually ask about.
What we built
A retrieval-augmented chatbot where every sentence of the answer carries its source. Each quote is checked word for word against the PDF’s own text before it is shown, and the matching passage is highlighted on the page. Quotes that can’t be verified are flagged, never displayed as fact.
- OCR with word positions, so scanned and photographed PDFs get the same highlights as digital ones
- Hybrid keyword and semantic search across the document
- A numbers table that lists every figure with its label, context and page, exportable to CSV
- An honest “not in this document” answer for out-of-scope questions
Result
The app ships with its own accuracy test. On the included question set it scores 23 out of 23 on both the digital PDF and a scanned copy of it. It’s live, so you can try it with your own file.