What generative QA does
Generative question answering produces a free form answer in the model's own words rather than copying a span. This lets it synthesize across sentences and rephrase, but it can also state things that are not true.
Closed book versus open book
- Closed book answers from parameters alone, relying on knowledge absorbed during training.
- Open book retrieves documents first and conditions the answer on them, which is retrieval augmented generation.
Open book grounding reduces hallucination and lets the system cite sources.
The retrieval augmented pipeline
- A retriever finds passages relevant to the question.
- A generator reads the question plus passages and writes the answer.
- Quality depends on both stages, since irrelevant retrieval poisons the answer.
Faithfulness and evaluation
- The answer should be grounded, meaning supported by the retrieved evidence.
- Automatic overlap metrics are weak here, so human or model based judgments of correctness and attribution matter.
- Returning citations lets a reader verify each claim.
Key idea
Generative QA writes free form answers, and grounding it on retrieved documents in a retrieve then generate pipeline reduces hallucination and enables citations the reader can verify.