Retrieval-Augmented Generation (RAG)

Pain Signals — You Need This Pattern When:

• The LLM confidently states facts that are wrong (hallucination)
• The LLM does not know about your proprietary data, internal docs, or recent events
• Answers need to cite specific sources for auditability or trust
• Fine-tuning is too expensive or too slow for the rate of knowledge change
• The knowledge base changes frequently — weekly, daily, or in real time
• Users ask questions that require domain-specific expertise the model lacks

Core Principle

RAG separates knowledge (what to reason about) from reasoning (how to think about it). The retriever provides relevant context; the LLM reasons over it. Neither is sufficient alone:

• LLM without retrieval → hallucination, stale knowledge
• Retrieval without LLM → keyword matches, no reasoning, no synthesis