Plain RAG retrieves once and answers, even when what came back is irrelevant.
Corrective RAG grades the retrieved passages first and falls back to another source if they are poor.
Self-RAG goes further and lets the model decide whether to retrieve at all before it answers.
It then critiques its own draft against the passages and revises if a claim is unsupported.
Both trade extra model calls for fewer confident answers built on nothing.
A standard RAG pipeline is a straight line: embed the query, retrieve, assemble context, generate.
It has no opinion about what came back, so bad retrieval quietly becomes a bad answer.
Worse, the answer is delivered with the confident tone that grounded generation is supposed to earn.
Corrective RAG inserts a judgement step between retrieval and generation.
A grader, often a small model, scores each retrieved passage for relevance to the query.
If the passages are good the pipeline proceeds exactly as before.
If they are poor, it triggers a fallback such as a rewritten query or a web search.
An ambiguous middle case usually means keeping what is relevant and supplementing it.
Self-RAG makes the model itself responsible for those decisions during generation.
It is trained to emit special tokens deciding whether retrieval is needed for the current sentence.
For a question answerable from parametric knowledge it can skip retrieval entirely.
After drafting, it critiques its own output against the retrieved passages for support and usefulness.
Unsupported claims trigger a revision rather than being shipped.
Both approaches trade latency and model calls for a lower rate of confidently wrong answers.
The pragmatic version most teams build is a relevance grader plus one query rewrite, which captures much of the benefit.
One question on this card. Every option explains itself once you pick, right or wrong.
19 concepts in this module, then a graded test with every answer explained.