Extractive question answering
Also: Extractive QA
Question answering that returns a span from a real document. It can pick the wrong span — it cannot make one up.
Last updated
Extractive question answering returns a span from a real document — a phrase, a sentence, a paragraph — that best answers a question. It does not write new text; the answer is highlighted, not composed. It can fail by picking the wrong span, but it cannot make one up.
Why it matters for your research. For researchers whose claims need to be citable, extractive is the safe default. The failure mode is “no good span found”, which is honest. The failure mode of generative QA is a plausible paraphrase that may not track any real source — i.e. a hallucination.
In Archēglyph. On the 90-day roadmap. Fits cleanly inside our “we don’t hallucinate” commitment — see The citable-claim test.
Not to be confused with. RAG also retrieves real documents, but then generates an answer over them; extractive QA stops at the span.