Document Q&A systems are one of the highest-value enterprise AI applications. Full stack: 1) Document ingestion: LlamaParse (cloud) or Unstructured (self-hosted) for intelligent PDF parsing preserving tables and structure. Split into semantic chunks with overlapping context. 2) Embedding and indexing: OpenAI text-embedding-3-small for embeddings, pgvector or Qdrant for storage. Include document metadata (filename, page, section) for citations. 3) Query processing: expand user query with hypothetical answer (HyDE) or similar questions, retrieve top-10 chunks, rerank with Cohere Rerank, select top-5. 4) Answer generation: pass chunks + query to GPT-4o with instruction to cite sources by [doc, page]. Parse citations from response. 5) Access control: row-level security ensuring users only access permitted documents. Implement document-level permissions. 6) UI: source highlighting showing which document sections were used, confidence indicators, follow-up question suggestions. Performance: target <3s latency for queries. Optimize with: caching common queries, pre-computation of frequent document summaries, async embedding generation. Scale: for >100K documents, partition by topic or department for focused retrieval. Evaluation: human annotation of 200 questions, measure retrieval precision and answer accuracy quarterly.