Use this when

You want an AI to answer questions about company-specific or frequently changing information that it cannot reliably know on its own.

The pattern

QUESTION
   ↓
RETRIEVE relevant passages from your sources
   ↓
ANSWER using only those passages
   ↓
CITE which source was used

Why it works

A model's own knowledge is general and can be out of date. By retrieving current, company-specific passages first and answering from them, you keep the answer grounded in your own maintained sources. This is the core of RAG.

Answering from retrieved passages, rather than from memory, also makes the answer checkable: you can see what it was based on.

How to apply it

  1. Prepare a current, authoritative set of sources. See how to prepare documents for AI.
  2. Retrieve the most relevant passages for each question.
  3. Instruct the model to answer using those passages, and to say when they do not contain the answer.
  4. Return the source alongside the answer, following cite the source.

Good fit

  • Answering from policies, product details or documentation.
  • Knowledge that changes often enough that the model cannot keep up.
  • Any answer a person may need to verify.

Poor fit

  • Information already provided directly in the workflow, where retrieval adds needless complexity.
  • A handful of documents a person can easily select by hand.

What this does not solve

Retrieval grounds an answer in a source; it does not make the source correct or guarantee the right passage was found. Owning and preparing the sources matters as much as the retrieval step.