DocSage: a PDF system that cites its sources

The problem

A language model handed a 90-page PDF will answer confidently and sometimes make things up. For anything that matters, a research paper, a contract, a report, you need two things a raw model doesn’t give you: answers that come from the actual document, and a way to check them.

DocSage is my take on that. You upload a PDF, ask questions in a chat, and every answer points back to the exact place it came from.

How it works

The system is a full-stack app: a Next.js front end and a FastAPI back end, with the retrieval pipeline running server-side.

1. Ingest. Text is extracted from the PDF and split into overlapping chunks.
2. Embed locally. Each chunk is embedded with sentence-transformers, on the server, with no external embedding API in the loop.
3. Store. The vectors go into ChromaDB for persistent storage; document metadata lives in SQLite.
4. Retrieve. A question is embedded the same way, and the closest chunks are pulled back as context.
5. Answer. The retrieved chunks and the question go to GPT-4o-mini, which answers only from that context and returns inline citations.
6. Stream. The answer streams back to the chat over Server-Sent Events, token by token.

Decisions worth calling out

Embeddings run locally, not through an API. Chunks are embedded with a local sentence-transformers model rather than a hosted embedding endpoint.

Answers are grounded, with citations. The model is constrained to the retrieved context and returns citations inline, so any claim can be traced to a passage in the source. This is the whole point of the project.

The hard part.

Results

Answers stay grounded in the uploaded document and carry citations for what they assert, and the pipeline persists across sessions so a document only needs indexing once.

Where I’d take it next

A few directions I’d build on: a reranking step over the retrieved chunks to lift answer quality, hybrid search (keyword plus vector) for queries where exact terms matter, and a formal evaluation loop with RAGAS to measure retrieval and faithfulness as the corpus grows.