Corrective RAG (CRAG) is a retrieval-augmented generation method that inserts a lightweight retrieval evaluator between the retriever and the generator, grades how relevant the retrieved documents are, and then corrects the evidence by refining it, replacing it with web search results, or combining both before the language model writes an answer. In the Retrieval & knowledge part of artificial intelligence, it addresses a weakness of standard RAG: the generator usually trusts whatever the retriever returns, even when those passages are off-topic or misleading.
The approach was introduced in 2024 by Yan and colleagues in the paper on corrective retrieval augmented generation. Its central idea is simple but consequential: retrieval quality should be checked, not assumed, and the pipeline should take a different path when confidence in the retrieved context is low.
How corrective RAG evaluates and repairs retrieval
Corrective RAG works by scoring each retrieved document against the query and turning the scores into one of three corrective actions. The evaluator in the original work is a fine-tuned T5-large model, which is small compared with the generator and therefore adds modest latency.
The process runs in four steps. First, a conventional retriever returns candidate passages. Second, the evaluator assigns each passage a relevance score. Third, two confidence thresholds split the outcome into Correct, Incorrect or Ambiguous. Fourth, the selected knowledge goes to the generator, which can be any instruction-following model without retraining.
The following table summarizes the three actions and what each one does with the evidence.
| Action | Trigger condition | What happens to the evidence |
|---|---|---|
| Correct | At least one document scores above the upper threshold | Documents are refined and irrelevant strips are removed |
| Incorrect | All documents score below the lower threshold | Retrieved documents are discarded and web search is used |
| Ambiguous | Scores fall between the two thresholds | Refined internal documents and web results are combined |
Knowledge refinement and web search fallback
When retrieval is judged correct, CRAG applies a decompose-then-recompose algorithm. Each document is split into short knowledge strips of a few sentences, the evaluator scores every strip, and only the relevant strips are concatenated into the final context. This removes noise inside relevant documents, which matters because a passage can be on topic overall while most of its sentences add nothing.
When retrieval is judged incorrect, the method rewrites the query into search keywords and calls a web search API. Results from the open web are then refined in the same strip-based way. The original paper prefers authoritative pages, such as encyclopedic sources, to limit the risk of importing low-quality web content.
Each component has its own role and its own cost, as the next table shows.
| Component | Function | Main cost or risk |
|---|---|---|
| Retrieval evaluator | Scores query and document relevance | Extra inference step and training data |
| Confidence thresholds | Map scores to an action | Tuning per domain and corpus |
| Strip decomposition | Splits documents into small units | May cut context that spans strips |
| Query rewriter | Turns questions into search keywords | Can drift from user intent |
| Web search connector | Supplies external knowledge | Latency, cost and source reliability |
Corrective RAG compared with related retrieval methods
Corrective RAG differs from neighboring methods mainly in where the quality check sits and whether the generator must be trained. Standard RAG has no check at all. Reranking reorders candidates but still passes the top results forward, even when every candidate is poor. Self-RAG trains the generator to emit reflection tokens that judge retrieval and its own output.
Because CRAG is plug-and-play, it can be wrapped around an existing RAG system or even around Self-RAG. The comparison below captures the practical differences.
| Approach | Quality check | Fallback when retrieval fails | Generator retraining |
|---|---|---|---|
| Standard RAG | None | None | Not required |
| Reranking | Relative ordering of candidates | None | Not required |
| Self-RAG | Reflection tokens from the generator | Answer without retrieval | Required |
| Corrective RAG | Separate retrieval evaluator | Web search with query rewriting | Not required |
| Agentic RAG | Agent reasoning over tool results | Agent chooses new tools or queries | Usually not required |
In practice, CRAG is often implemented as a graph-based workflow in orchestration frameworks, where the grading step is a node that routes to either generation or search. Such implementations frequently replace the trained T5 evaluator with an LLM grader prompt, which is easier to deploy but slower and less consistent.
When to use corrective RAG
Corrective RAG is most useful when the internal corpus is incomplete and wrong answers are costly. It helps least when the knowledge base is curated, closed and well covered, because the evaluator then rarely changes the outcome while still adding latency.
The decision table below helps match the method to typical deployment situations.
| Situation | Recommended choice | Reason |
|---|---|---|
| Open-domain questions with sparse internal data | Full CRAG with web fallback | External search fills coverage gaps |
| Regulated or confidential knowledge base | CRAG without web search | Grading helps, but data must stay internal |
| Strict latency budget under one second | Reranking only | Evaluator and search add round trips |
| Long noisy documents such as manuals | Strip refinement only | Filtering irrelevant text improves grounding |
When web search is disabled, the Incorrect branch should trigger a clear abstention or an escalation to a human rather than an unsupported answer. This preserves the main benefit of CRAG, which is knowing when the evidence is weak.
Limitations and common mistakes
The method is only as reliable as its evaluator. A grader that is miscalibrated for a new domain will discard good documents or accept poor ones, and the errors propagate silently. Thresholds tuned on benchmark data such as PopQA or PubHealth rarely transfer without adjustment.
- Web search is not a truth source: it introduces outdated, biased or adversarial pages unless sources are filtered.
- Latency grows with every branch: grading, rewriting and searching can add several hundred milliseconds or more.
- Strip filtering can break reasoning: multi-hop questions may need sentences the evaluator scores as irrelevant in isolation.
- Data leakage risk: sending rewritten queries to a public search API can expose confidential intent.
A frequent mistake is to skip evaluation of the grader itself. Teams should measure its precision and recall on labeled query and document pairs, alongside end-to-end RAG evaluation, before trusting its routing decisions.
In everyday language
A normal question-answering system looks things up and then answers using whatever it found, even if what it found is useless. Corrective RAG adds a checker in the middle. Before the answer is written, the checker reads the found material and decides whether it is good, bad or unclear.
If the material is good, the system trims away the irrelevant parts. If it is bad, the system throws it out and searches the web instead. If it is unclear, it uses both. Think of a careful research assistant who checks the sources before writing the report, instead of copying the first pages that came up.
Seen in the real world
A university library runs a question-answering assistant over its catalog, course guides and internal policy pages. Students often ask about recent events or general topics that the internal collection does not cover, and the original RAG assistant answered those questions with confident but unrelated passages.
The development team adds a corrective layer. A small grading model scores the top five retrieved passages; when all of them fall below the lower threshold, the system rewrites the question into keywords and queries a filtered web search limited to encyclopedic and academic domains. For policy questions, web search is disabled and low scores trigger a referral to library staff.
After a month of review, staff report that off-topic answers dropped noticeably, while average response time rose by roughly half a second, a trade-off the library considered acceptable.