Why your chatbot can get an answer wrong when the website is right

Keep the question exactly as it was asked

Save the customer’s wording, the assistant’s reply and when it happened. Keep enough of the preceding conversation to understand what words such as that or tomorrow meant. Use a redacted copy when sharing the issue with somebody who does not need the customer’s details.

Then write the answer you expected, with a link to the approved source. Be specific about the part that is wrong. An answer might identify the correct service while missing a condition that changes whether the customer can use it.

Use that same question for the first retest. Rewording it immediately can make the failure disappear without telling you whether the original problem is fixed.

Compare the page with the imported copy

Check which page is supposed to be authoritative and whether the knowledge system has a current copy of its useful text. A page can look correct in a browser while an importer has collected only a heading, an older version or the wrong section.

Look for competing sources too. An old document might still describe a withdrawn service. If two sources disagree, decide which one should be used and remove or correct the stale material.

Compare the relevant condition in both versions, rather than relying only on a successful import status. You need to see whether the information required for this answer reached the knowledge store.

Read the passage used for this reply

RAG usually gives the assistant selected passages rather than every page you supplied. Inspect those passages for the failed question. Retell, for example, provides a knowledge-retrieval view alongside relevant transcript turns.

For an invented example, imagine a collection page says Saturday collections are available only by prior appointment. A passage containing the Saturday heading but losing the appointment condition leaves the assistant with incomplete evidence.

Check whether the condition appears beside the rule in the imported text and retrieved passage. The next fix may involve how content is divided or found. Adding a stronger instruction to answer accurately will not restore text that never reached the reply.

Source: Retell: knowledge base and retrieval inspection

Check how the assistant used the information

If the full rule and its condition were present, inspect the reply itself. Did the assistant apply the condition? Did it ask whether the customer had an appointment, or turn a conditional answer into a general yes?

OpenAI’s accuracy guidance separates retrieval problems from cases where a model receives the right context and handles it incorrectly. That distinction helps choose the next change.

I want the repair tied to the evidence. It might be clearer instructions for conditional answers, a better example or a different retrieval setup. Record what changed so the next person can understand why.

Source: OpenAI: optimising LLM accuracy

Retest the exception as well as the easy question

Run the original wording again, then try nearby questions. In the invented collection example, ask about Saturday with an appointment, without one and without saying either way. Check that the assistant handles the missing detail sensibly.

Keep these questions in a small set of repeatable checks. Run them after later source or instruction changes, alongside ordinary questions that already worked.

We like to build RAG systems and dedicated brain modules around this sort of evidence. The approved sources matter, and so does being able to see which information produced an answer. Bring one failed question and its source page to the investigation; that gives us somewhere concrete to start.