AI chatbots & RAG

A language model without your material answers from what it was trained on, and it sounds just as sure whether it is right or not. RAG, retrieval-augmented generation, means we first search for the documents that look relevant to the question and then let the model answer with them in front of it. That gives the model something to stand on. It does not make the answer right by itself, so both the search and the answers need evaluating.

What you need in place

• Material that is up to date. If the assistant answers from last year’s price list, the mistake is yours, and the customer cannot see it was an old file. Decide where the material is maintained and how it reaches the index.
• Permissions kept at document level. The assistant must never answer with something the user would not have been allowed to read. That belongs in the search, so documents the user lacks access to are never fetched, not in an instruction asking the model to keep quiet.
• Source references in the answers, so whoever asked can open the document and check that it really says so.
• A list of real questions to evaluate against, taken from support, with answers you know are right.

Why are the answers bad?

For two reasons that need keeping apart. Either the search fetched the wrong documents, and the model had nothing right to answer from. Or it fetched the right documents and the model still got it wrong, or added something that wasn’t there. The evaluation has to show which, because the fixes differ: the first is fixed in search, the second in how the model is instructed and checked.

So we start in the search system and measure that it fetches the right things before the language model goes on top. We also set limits on what the assistant may answer, and a “don’t know” mode for when the material isn’t enough. A missing answer is something the user can move on from. A wrong answer that sounds confident is not.