When an AI assistant answers a question about your own documents, there are two ways it can go wrong. It can make something up, or it can be right but give you no way to check.
Answer from sources, not from memory
The common pattern is called retrieval. The system first finds the most relevant passages in your content, then asks the model to answer using only those passages. The model is a writer here, not a source of facts.
Show the sources
Attach the passages that were used to each answer, with a link back to the original page or file. Readers can verify the answer in one click, and you can see when it went wrong.
Let it say “I don’t know”
If retrieval finds nothing relevant, the right answer is to say so and, where it makes sense, point to a person. An assistant that admits a gap earns more trust than one that always sounds sure.
Things to get right
- Split documents into passages of a sensible size, so each one makes sense alone.
- Keep the source text up to date, and remove what is out of date.
- Label the assistant as AI wherever people meet it.
- Don’t use people’s content to train a shared model without their agreement.
By Foldox. All posts