Ch 9 Readers for Question-AnsweringApply
The component that turns retrieved passages into an actual answer.
Core concepts
- Extractive readers. A BERT-style model predicts the start/end span of the answer within a passage — the answer is copied verbatim from the source.
- Fine-tuning on SQuAD-style data. Question + context → answer span.
- Extractive vs. generative. Extractive readers are grounded and traceable but limited to spans present in text; generative readers (LLMs) synthesize fluent answers but can hallucinate.
- Confidence & no-answer. Good readers can say “the answer isn’t here”.
What you must master
- Explain extractive span prediction (start/end logits) Level 1
- Contrast extractive vs. generative readers and their failure modes Level 2
- Run a retriever→reader QA pipeline end-to-end Level 2
- Choose extractive vs. generative for a compliance/traceability requirement Level 2
Architect’s lens
The extractive-vs-generative choice is a governance decision, not just a quality one. Regulated domains (legal, medical, finance) often demand traceable, source-grounded answers — where extractive readers or heavily-cited generative RAG win over free-form LLM output. Know how to bound hallucination risk.