What Is Abstention in AI Answers?
When an AI assistant declines to answer because its sources don't cover the question, and says so plainly instead of writing a confident guess.
An assistant that answers every question will sometimes answer questions its sources never covered. Abstention is the opposite: when retrieval finds nothing that addresses the question, the assistant says it doesn't know instead of writing a fluent guess. It is the main defence against hallucination in a customer-facing assistant.
Most systems ask for abstention in the prompt ("if the context doesn't answer, say you don't know"). Models follow that most of the time, not every time, because they are still handed the nearest pages, which share the subject without answering the question. Enforced abstention takes the decision out of the prompt: if retrieval judges that nothing answers the question, the model never sees those pages.
Too much abstention is also a failure, so it depends on a good relevance judgment, usually from a reranker. Done well, every abstention is useful data: it marks a knowledge gap, a real question your content can't answer yet.
How Kelu uses this
Every Kelu answer gets a coverage verdict after reranking: answerable, partial or not covered. On a not-covered verdict, an abstention gate withholds both the excerpts and the citations, so the model has nothing to guess from and tells the reader the question isn't covered. A partial answer says plainly what is missing. The gate runs on every answering surface, and abstained questions show up in the Coverage Gaps report.
Related terms
Hallucination
When a language model generates fluent, confident content that is factually wrong or entirely invented.
Grounding
Tying an AI answer to specific source content, so each claim comes from retrieved passages rather than the model's memory and can be checked against them.
Reranking
A second pass in which a stronger model re-scores the top search results against the question, so the passages that actually answer it rise to the top.
Knowledge Gap
A question real users ask that your content can't answer, or answers badly. Tracked over time, knowledge gaps become a ranked list of docs to write next.
Citations
Links from an AI answer back to the specific source documents used to generate it, allowing users to verify every claim.
From definition to deployment
Connect a knowledge source and see cited AI answers on your own content in minutes — free, no credit card required.
Free to start · No credit card required