Design a RAG question-answering system

## Scenario Design a **Retrieval-Augmented Generation (RAG)** system that answers user questions using an internal document corpus (e.g., product docs, policies, runbooks). The system should ground answers in the corpus and cite sources. ## Requirements ### Functional - Users submit a natural-language query and receive an answer generated by an LLM. - The answer must be grounded in retrieved documents (include citations/links/IDs). - Support document ingestion/updates (new docs, edits, deletions). - Handle multi-turn conversations (optional, but describe how you would support it). ### Non-functional - Latency target (p95): e.g., **< 3 seconds** for typical queries. - Availability: e.g., **99.9%**. - Data privacy: some documents may be access-controlled per user/team. - Quality: minimize hallucinations; provide a way to evaluate and monitor quality. ## What to cover - High-level architecture and main components. - Data ingestion and indexing pipeline (chunking, embeddings, metadata). - Retrieval strategy (top-k, filtering, reranking). - Prompting/generation strategy (context window management, citations). - Storage choices (vector DB, metadata store) and scaling approach. - Caching, monitoring, evaluation, and failure modes.