The problem is not wrong answers, it is unmarked ones
Every system that answers questions is sometimes wrong. A language model has a particular failure: it writes a wrong answer in exactly the same confident tone as a right one. Nothing on the screen tells the reader which is which.
For documentation that is expensive. A developer follows a wrong instruction and loses an afternoon. A customer is told a plan includes a feature it does not. A support agent repeats a policy that was retired last quarter. In each case the cost is not the error itself but the trust that disappears when it is discovered, and the reader’s next move is to stop asking.
What a citation is for
A citation turns a claim into something checkable. It says: this sentence comes from this passage, go and read it. The reader does not have to trust the assistant, only the documentation, which they already know how to judge.
That only works if citations are honest. A useful citation has three properties:
- It points to the passage, not just the page. “See the API reference” helps no one. The exact paragraph does.
- The passage supports the sentence. A link to a page that mentions the topic but says something different is worse than no link, because it looks like proof.
- It resolves. The link opens, to the version of the page the answer was drawn from.
What it means to abstain
Abstaining means the assistant declines to answer when the sources it can see do not support one. It is the reverse of the usual instinct, which is to always produce something. For a documentation assistant the usual instinct is the wrong one.
In practice it works like this. The assistant searches your content for passages relevant to the question. If nothing relevant comes back, or what does come back does not actually answer it, a good assistant says it could not find the answer, and does not fall back on what a model happens to remember from its training. That is the difference between an assistant that answers from your sources and one that merely talks about your topic.
Abstaining does not remove the risk of a wrong answer. It narrows it, and it makes the remainder visible, because an answer that carries citations can be checked and an answer that is withheld cannot mislead.
Declining should come with a reason
There are two quite different reasons an assistant fails to answer. Either your content does not cover the question, or the service had a problem. They call for different responses. The first is a documentation gap, and the second is an outage to retry later.
If the assistant powers something in your product, the reason matters to your software too. A support form can hand a “not covered” question to a person and retry a “temporarily unavailable” one. MITHUNAI RAG returns exactly that distinction through its API, so your application does not have to guess.
Every declined question is a to-do item
There is a second benefit that teams tend to notice after a few weeks. Each question the assistant declined is a question real people asked that your documentation did not answer. Sorted by how often they were asked, they are a ranked list of what to write next.
An assistant that always answers hides that list. This one produces it. MITHUNAI RAG shows unanswered questions and top questions in its analytics, and neither view names who asked.
How to test an assistant for both
You can check this in an afternoon, without trusting any vendor’s description.
- Ask ten things your docs answer clearly. Click every citation. Does the passage prove the sentence?
- Ask ten things your docs do not cover. Pick plausible ones, such as a feature that sounds like it should exist. Does it decline, or invent?
- Ask ten questions that need two pages. Does it combine them correctly and cite both?
- Ask something with an answer that changed. Does it use the current page or the old one?
What to expect from us, and from anyone
No assistant gets everything right, and you should be suspicious of one that claims to. What you can reasonably demand is that its answers can be checked and that it says so when it does not know.
That is what MITHUNAI RAG is built to do: it cites the passages it used and declines when your content has no answer. You can see it on the public playground with no account, read how an answer is made on the resources page, and compare the approach with others on the alternatives page. To watch it decline questions your own docs do not cover, book a demo.