Glossary
What Is RAG (Retrieval-Augmented Generation)
RAG is a model-side mechanism in which a large model retrieves external documents before generating an answer.
- Author
- Xingge
- Published
- 2026-11-09
- Updated
- 2026-11-09
- Level
- L2 Reproducible
- Version
- 1.0
Term details
- Name
- What Is RAG (Retrieval-Augmented Generation)
- Type
- Glossary
- ID
- rag-en
What Is RAG (Retrieval-Augmented Generation)
RAG (Retrieval-Augmented Generation) is a mechanism in which a large model first retrieves external documents and then generates an answer based on the retrieved results. RAG is a model-side process, not a content optimization method, but it explains why crawlability and extractability directly affect citations.
Definition
This site's working definition: RAG is a class of methods that insert an external retrieval step before generation and use the retrieved passages as the basis for generation, usually including four steps: retrieval, ranking, stitching, and generation.
For content teams, only two stages can be influenced: whether the page enters the retrieval results, and whether the retrieved passage is self-contained. How the model stitches and generates is not controlled by the content side.
Boundaries: what it is easily confused with
| Easily confused concept | Key difference |
|---|---|
| GEO | GEO is a content-side method; RAG is a model-side mechanism |
| Fine-tuning | Fine-tuning changes model parameters; RAG does not |
| Caching | Caching reuses results and does not involve retrieval quality |
Common misconceptions
- Treating RAG as a writing format. RAG requires no specific format, only that passages can be understood.
- Assuming that being retrieved means being cited. A retrieval hit is only a precondition; whether the answer uses the passage is decided by the model.
- Ignoring paragraph self-containment. Passages that lose meaning after splitting are hard to stitch into an answer.
Basis and verification
This definition is this site's working definition, marked L2. The mechanism description covers only the publicly observable general flow and makes no capability judgment about specific models.
Review cycle: 180 days.
Related sections: glossary index, method index, metric index, editorial policy.