How do you keep AI from hallucinating when it doesn't know the answer?

I've been building an AI document reader (DocMind — upload a PDF/DOCX, get a summary, key points, plain-language explanation, or chat with it directly).

The hardest problem wasn't extraction or summarization — it was getting the model to admit when something isn't actually in the document instead of confidently making something up. Early on, I had a small model invent an entire fake financial report for a document that was actually about fintech hardware — fluent, convincing, completely wrong.

What ended up working: switching to a stronger primary model, adding a "groundedness check" that verifies the output actually shares vocabulary with the source text before accepting it, and falling back to a retry (or a different provider entirely) if it fails that check.

Curious how others building with LLMs are handling this — are you doing something similar (grounding checks, RAG-style citations, confidence scoring), or found a different approach that works better? Launching my product Wednesday, but genuinely more interested in how other people are solving this specific problem right now.