Systems that answer
Look the answer up first, then have the model phrase it — an architecture that turns a question about the model into a question about your material.
Also called Retrieval-augmented generation · Извлечение с генерацией
The mechanism is unglamorous. The question is matched against your own passages, the best few are put in front of the model, and the model is instructed to answer from them and nothing else. What changes is where the truth lives: not in weights trained by somebody else on the internet, but in a corpus you own, can inspect, can correct in an afternoon, and can point at when somebody asks where the answer came from.
It is worth being precise about what this does and does not solve. It solves currency — your prices today, not the model's memory of a page from last year. It solves attribution — every claim traceable to a passage. It does not solve invention: a model given three passages that do not contain the answer will often produce one anyway, fluently, which is why retrieval has to be allowed to return nothing and the system has to be built to say so.
The failure mode nobody expects is the retrieval itself. Most disappointing deployments are not model failures; they are searches that returned the wrong three passages, and the model then answered the wrong question correctly. That is diagnosable — the retrieved passages are recorded — and it is where the work is.
When an answer is wrong the instinct is to change model or rewrite the prompt. Look at the retrieved passages first. If they did not contain the answer, no model would have helped and the fix is in the corpus or the indexing. Every deployment worth running records what was retrieved for each answer, and the ones that do not cannot be debugged at all.
The definitions are the easy part. Whether the figure on your dashboard was computed this way is a different question, and usually the more expensive one.