Hallucination
A hallucination is a confident, fabricated answer a model produces and states as fact, even though it isn't true.
Category
Industry concept
See all terms
What is a hallucination?
A hallucination is a confident, incorrect answer a model produces, stated as fact even though it isn't true. The model isn't lying in any deliberate sense, it's generating the most statistically plausible-sounding answer it can, and when it doesn't actually have the right information, that plausible-sounding answer is often wrong.
This happens because of how models are trained and graded: most evaluations reward a confident answer over an honest "I don't know," so a model that guesses scores better on average than one that admits uncertainty. The behavior that helps it score well on a test is the same behavior that produces a hallucination.
How does a hallucination differ from a grounded answer?
A hallucination comes from the model's internal weights, its best guess based on patterns learned during training. A grounded answer comes from actual retrieved text, a document, a database record, a live lookup, that the model reads before answering, so it has a concrete source to work from instead of a memorized approximation.
That's why more training alone doesn't remove hallucination, and grounding does: no amount of pretraining gives a model private, current company information, but handing it the actual document at the moment it answers does.
How does grounding reduce hallucination in practice?
In practice, grounding means retrieving relevant, current information and inserting it into the model's context before it generates an answer, whether that's a single document, a set of search results, or a broader knowledge base. The model still has to read and use that information correctly, so grounding reduces hallucination without eliminating it, a model can still misread or misapply a source it was given.
Qontext reduces the odds of that gap by keeping the source itself structured, current, and permission-aware in the context repository, so what an agent retrieves is accurate to begin with, not just present.