How We Cut LLM Hallucinations Down in Legal

In DocuChat, we ran into a classic AI-era problem: the model confidently cites a law that… never existed. It looks polished and convincing, but in legal, that’s no longer a bug — it’s a mini disaster.

What we did: we added ground truth verification. Now every citation to a legal act is not accepted at face value, but checked against the database. If the act isn’t there, the LLM is hallucinating again, and we block it at the input stage.

Long documents were a separate kind of fun: 100+ pages, chunking + retrieval, and the model loses the big picture along the way. Pretty much like us after the fifth call of the day. So we balanced chunk size and retrieval quality to avoid losing either context or accuracy.

Against the hype around AI agents and copilots, this is table stakes: in sensitive domains, there’s no getting around verification anymore.