Generation Over Extractive Retrieval

Prefer generative readers when answers need synthesis across clue-bearing passages or parametric fallback — extractive spans score zero if the string never appears.

Classic open-domain QA often extracts answer spans from retrieved docs. RAG-style generators instead emit free text conditioned on top-(k) passages. Empirically that wins when:

• Documents contain clues but not the verbatim answer — marginalization still contributes probability mass.
• The answer is in no retrieved document — generators can fall back on parametric memory (Lewis et al. report non-zero accuracy on such NQ cases; extractive systems score 0%).
• Outputs must combine facts from different docs (Jeopardy-style QGen) — favor RAG-Token so document posteriors can shift mid-sequence; non-parametric hits also guide which parametric titles or facts get drawn out.

Abstractive and verification tasks follow the same lesson: fewer hallucinations and higher factuality/specificity than generator-only baselines, and strong FEVER accuracy without supervised evidence labels when the retriever proposes its own support. Keep extractive readers when answers are guaranteed spans and latency is tight; otherwise design the harness as generative RAG under context grounding.

---

Sources

• RAG (Lewis et al.), §3–4.4