Question answering (QA)

Producing direct answers to user questions — either from the model's parametric knowledge (closed-book) or by retrieving from documents (open-book / RAG).

Question answering (QA) is the task of producing a direct answer to a question. The two main flavors: closed-book QA (the model relies entirely on what it learned during training) and open-book QA (the model is given relevant documents to consult, like in RAG). Modern AI assistants combine both — using internal knowledge for general questions and retrieval for specific or recent information. It matters because QA is the most natural interface to information and one of the highest-value AI use cases. Search engines are gradually being supplemented or replaced by QA — Perplexity, Google's AI Overviews, ChatGPT search, You.com all aim to give you an answer rather than a list of links. Inside companies, QA over internal documents (HR policies, technical documentation, legal contracts) is one of the most common AI deployments. A concrete example: "What's our company's policy on remote work?" — the system retrieves the HR handbook, finds relevant sections, and writes a synthesized answer with citations. Or "What's the population of Taipei?" — closed-book, the model just knows. Or "What did the Federal Reserve announce yesterday?" — open-book search needed. Quality depends heavily on accurate retrieval (for open-book) and faithful generation (avoid hallucination). Strong QA systems cite their sources, distinguish between high and low confidence, and refuse when they don't know. Related: RAG, search, hallucination, retrieval, perplexity (the company).