open-book qa
**Open-Book QA** is a question-answering paradigm where the model has access to external knowledge sources—such as retrieved documents, knowledge bases, or provided context passages—during inference, analogous to an open-book examination where students can consult reference materials. The model must identify relevant information from the provided or retrieved sources and synthesize it into an accurate answer.
**Why Open-Book QA Matters in AI/ML:**
Open-Book QA is the **dominant paradigm for production QA systems** because it combines the reasoning capabilities of language models with the accuracy and updatability of external knowledge sources, dramatically reducing hallucination compared to closed-book approaches.
• **Retrieval-augmented answering** — A retriever (sparse BM25 or dense DPR) fetches relevant passages from a knowledge corpus, and a reader model (BERT, T5, or GPT-based) extracts or generates answers conditioned on the retrieved evidence, grounding responses in verifiable sources
• **Extractive vs. generative** — Extractive open-book QA selects answer spans directly from retrieved passages (high precision, limited to stated information); generative open-book QA produces free-form answers conditioned on evidence (more flexible, risk of unfaithful generation)
• **Knowledge updatability** — Unlike closed-book models where knowledge is frozen at pre-training, open-book systems update their knowledge by refreshing the document corpus—no retraining required—enabling real-time knowledge currency
• **Evidence provenance** — Open-book QA can cite source passages and provide attributions for answers, enabling users to verify correctness and building trust through transparent reasoning chains
• **Multi-document reasoning** — Advanced open-book systems retrieve and reason over multiple passages simultaneously, synthesizing information across sources to answer complex questions that no single document fully addresses
| Component | Options | Role |
|-----------|---------|------|
| Retriever | BM25, DPR, Contriever, ColBERT | Fetch relevant passages |
| Knowledge Source | Wikipedia, web, domain corpus | External information store |
| Reader/Generator | BERT, T5, GPT, LLaMA | Generate answer from evidence |
| Pipeline Type | Retrieve-then-read, RAG, RETRO | Architecture integration |
| Answer Type | Extractive span or generated text | Depends on task requirements |
| Evaluation | Exact Match (EM), F1, ROUGE | Standard QA metrics |
**Open-book QA is the foundational architecture for reliable, production-grade question-answering systems, combining neural language understanding with external knowledge retrieval to produce accurate, verifiable, and updatable answers that overcome the hallucination and knowledge-staleness limitations inherent in closed-book parametric approaches.**