Extractive QA
find answer span in given passage
question:
"When was iPhone released?"
context: [long passage]
BERT / RoBERTa + span head
predict (start_idx, end_idx)
"June 29, 2007"
verbatim span, never invents
SQuAD-style. never hallucinates.
needs passage + question together.
fails on unanswerable questions
unless SQuAD 2.0-trained.
RAG
retrieve passages, then answer
question only
retriever
BM25 + dense + rerank
top-k passages
(grounding context)
reader (LLM or extractive)
grounded answer + citations
Generative (closed-book)
answer from model weights
question
decoder-only LLM
(GPT, Claude, Llama)
parametric memory only
free-form answer
(may hallucinate)
good on common knowledge.
bad on rare / recent facts.
needs RAG for reliability.