Why an assistant still gets answers wrong when it has your documents

Having your file in the library doesn't guarantee a right answer. There are at least four places where things slip, and knowing which one you're looking at tells you what to fix.

Wikipedia is blunt about it: retrieval is used to prevent these errors, but it does not solve all the problems.¹

The file isn't what you think it is

Check this first, because it's the cheapest fix.

• The assistant has an old copy. You edited the price list on your computer, but the library still holds the earlier upload. Some tools may not notice that. There's more on this in the refresh and delete guide.
• Two versions sit side by side. Wikipedia describes a worst case where the model may combine details from multiple sources, producing responses that merge outdated and updated information in a misleading manner.¹ That's how a reply can mix last year's price with this year's terms.
• The file never made it in. If the file is a type your tool doesn't support, the assistant has nothing to find. The tool's documentation says what it supports.

The lookup picks the wrong passage

Anthropic says traditional RAG solutions remove context when encoding information, which often results in the system failing to retrieve the relevant information from the knowledge base.² The right information is in the library. The search just didn't find it. Two earlier pages give background: one on splitting and indexing, one on matching by meaning.

Sending more passages isn't a cure. Anthropic notes that more information can be distracting for models, so there's a limit to this.²

The model misreads what it was given

Even with the right passage in front of it, the model can go astray. Wikipedia says language models can generate misinformation even when pulling from factually correct sources, if they misinterpret the context.¹ Take a rule that says sale items can't be refunded unless they're faulty. It's easy to squash that into "no refunds."

It answers when it should say "I don't know"

AWS lists a known problem of language models: presenting false information when they do not have the answer.³ Wikipedia adds that, without specific training, models may generate answers even when they should indicate uncertainty.¹

Wikipedia defines a hallucination as a response generated by AI that contains false or misleading information presented as fact.⁴ So if your library lacks the answer, the assistant may still produce something that reads well. Check the maker's documentation for any control over what the assistant does when nothing in the library matches.

How to tell which problem you have

• Find the real answer in your own file, and note the exact sentence.
• Ask the same question in a few different wordings. If only some work, suspect the lookup.
• Look at the source links the assistant shows. If they point to an old file, suspect the library. If they point to the right file but the answer still differs, suspect the model's reading.
• If no source link points to anything relevant, suspect the answer was made up.

Reading the source links and checking answers walks through the source links. For anything that affects a customer, such as a price, a deadline or a policy, a person should confirm the answer against the original file.