<arch.design/>
← Lessons/

RAG: Retrieval-Augmented Generation

1 / 9

The LLM knowledge problem

LLMs are trained on a static snapshot of the world. Ask GPT-4 about your internal docs, last quarter's sales, or a bug filed yesterday — it cannot know. And when it doesn't know, it hallucinates.

  • CUTOFFTraining data has a fixed date — the model knows nothing after it
  • PRIVATEThe model has never seen your internal documents, codebase, or database
  • HALLUCINATIONRather than say "I don't know", LLMs confidently fabricate plausible answers
  • FINE-TUNING?Fine-tuning updates behavior, not knowledge — and costs thousands of dollars to re-run

LLMs hallucinate because they reason from training data, not from your live information.

← → navigate · S recap