Imagine a new employee answering customer questions. Without a handbook, they rely on memory. With RAG, they first look up the relevant pages in the approved handbook, then write an answer using those pages.
The four steps
- Prepare: collect trusted documents and split them into useful passages.
- Find: turn a question into a search and retrieve likely relevant passages.
- Answer: give the question plus retrieved passages to the language model.
- Check: show citations or source snippets so a person can verify the answer.
When RAG helps
RAG is useful when answers must depend on a specific, changing body of information: product manuals, company policies, course notes or a catalogue. It can be updated by changing the source collection rather than retraining the whole model.
What RAG does not solve
| Failure | Why it happens | Better practice |
|---|---|---|
| Wrong document retrieved | Search found similar words, not the right meaning | Improve chunking, metadata and queries |
| Correct source, wrong answer | The model misread or combined passages badly | Require quotes and verify them |
| Missing current information | The source collection was not updated | Track document dates and owners |
| Private data leakage | Permissions were ignored during retrieval | Enforce access before search, not after |
A no-code mental test
Before building anything, take ten realistic questions and identify the exact source passage each answer needs. If humans cannot find a clear source, RAG will not magically create one. Better documents often improve the system more than a more complex model.
RAG in one lineSearch trusted material first; generate an answer second; keep the source visible for checking.