[Day 41/100] RAG Fundamentals: Why Your LLM Needs Retrieval, Not a Bigger Context Window

Long context and fine-tuning both look like shortcuts. Why retrieval wins for company documents, how to build a LangChain RAG baseline that cites its sources, and the eight questions it will get wrong.

AI Generated Image

Day 41 of 100 · 100 Days of Learning Agentic AI

Picture the agent you finished in Phase 4. It remembers that Priya is vegetarian, keeps her memories away from every other user, and forgets on schedule. Now Priya, who works at a company called Kestrel Labs, asks it something ordinary: “Can I carry over my unused PTO, and is there a limit?”

The agent has never seen Kestrel’s handbook. It has seen thousands of other companies’ policies, though, so it does what language models do with a gap: it fills it. “Up to five days” might be right by coincidence. “Up to ten days” would sound exactly as confident. Priya plans a December trip around the answer, and nobody finds out it was invented until her days expire.

This post opens Phase 5, fifteen days on retrieval-augmented generation (RAG): why retrieval beats the two obvious alternatives, the four stages of a RAG pipeline, a complete LangChain…