Building a reliable RAG chatbot

An assistant that answers correctly nine times out of ten is not 90 % reliable. It is unusable — because nobody knows which of the ten answers is wrong, and one is enough to quote a price that no longer exists or a deadline you cannot meet.

The problem is almost never the model

When a RAG chatbot gets something wrong, the model takes the blame. In practice it answered correctly from what it was given: last year's price list, two versions of the same document, a scanned PDF from which it read one word in three. A model has no way of knowing a file is out of date. There is no bad answer from a good source.

So before a line of code is written, a non-technical question has to be settled: for each subject, which document is authoritative? That question often takes longer than the build, and rightly so — it decides the outcome.

Three rules that change everything

• One source per subject. Give it two documents that contradict each other and the assistant will decide for you — wrongly, half the time.
• The right not to know. “I don't have that information, here's who to contact” is worth far more than an invented answer. It goes in the instructions, and it gets tested.
• Cite the source. The user checks in one click, and you immediately see which document is causing trouble.

What happens after launch

A useful assistant changes your documents more than it changes your support desk. Log every question that went unanswered: after two weeks you hold the exact list of what your customers are looking for and you have written down nowhere. That is often the real deliverable — and nobody put it in the brief.