RAG that cites itself

RAG that cites itself

Most RAG demos retrieve a chunk and hope the model stays honest. I wanted the opposite:

a pipeline where every claim carries a receipt.

The loop

query -> retrieve (hybrid dense + BM25) -> synthesise -> verify citations -> publish

• Retrieve with ColPali late-interaction for visuals and BM25 for exact terms.
• Synthesise a report with an LLM, but force inline source markers.
• Verify each marker against the retrieved set. Unverifiable claims get dropped.
• Publish the eval report, including the failures.

Why publish failures?

Because a system that hides its misses is lying to you. The eval suite in citecheck

prints what it got wrong. That is the whole point — trust is built from evidence, not vibes.