Extractive QA (the SQuAD paradigm): given a document and a question, identify the exact span of text that answers the question. Fine-tuned BERT models excel at this — they read the document, understand the question, and highlight the answer. This is fast, accurate, and verifiable (the answer is always a direct quote). But it can only answer questions whose answers appear verbatim in the document.
The dominant modern pattern: (1) user asks a question, (2) retrieve relevant documents from a knowledge base using semantic search, (3) include the retrieved documents in the LLM's context, (4) the LLM generates an answer based on the retrieved context. This combines the precision of retrieval with the fluency of generation. The key challenges are retrieval quality (finding the right documents) and faithfulness (generating answers that accurately reflect the source material).
QA accuracy is measured differently for each paradigm. Extractive QA uses exact match (EM) and F1 score against ground-truth answer spans. Generative QA is harder to evaluate automatically — multiple valid phrasings exist for any answer. RAGAS and similar frameworks evaluate RAG-based QA on faithfulness (does the answer match the source?), relevance (did you retrieve the right documents?), and answer quality. Human evaluation remains the gold standard for generative QA.