How does corrective RAG act on a Correct, Ambiguous or Incorrect grade?

## The failure CRAG exists to fix A standard pipeline retrieves top-k and generates, unconditionally. If the corpus has nothing on the question, the retriever still returns k passages — the nearest neighbours of a query in a space where nothing is near — and the generator, obediently grounded, writes a confident answer on top of irrelevant text. That is the worst outcome available: wrong, fluent and apparently cited. Corrective RAG inserts a check between the two stages. A lightweight retrieval evaluator scores how well the retrieved set actually answers the question, and thresholds turn that score into one of three actions rather than a single go-ahead. ## The three actions **Correct.** The evaluator is confident relevant evidence was retrieved. CRAG does not simply forward the documents, because a relevant document is mostly not the answer — a five-page ingredient policy contains one paragraph about tree nuts. It applies a decompose-then-recompose step: split each passage into fine-grained knowledge strips, score the strips, drop the irrelevant ones, and reassemble the survivors into the context. The generator then sees dense evidence rather than a haystack. **Incorrect.** The evaluator is confident nothing retrieved is relevant. The corrective move is to *discard everything* — this is the part people skip when they describe CRAG, and it is the important part. Keeping bad context and adding good context still leaves the model reading bad context. CRAG then rewrites the question into a search-engine-shaped query and retrieves from an external source, typically web search, as a knowledge source of last resort. **Ambiguous.** Scores near the thresholds mean the evaluator itself is not sure. Rather than force a binary decision on a weak signal, CRAG combines both paths: keep the refined internal evidence and add external results. It is an explicit acknowledgement that the grader is a model too and its confidence is a distribution, not a fact. ## A worked scenario Consider a food-allergy substitution assistant over a manufacturer's product and ingredient corpus. A user asks whether a particular ready-meal is safe for someone avoiding sesame. If the corpus holds the current ingredient declaration for that product, the grader returns Correct; strip-level refinement drops the marketing copy and the storage instructions and passes the allergen table. If the product was reformulated last month and only the old spec is indexed, the retrieved passages are about the right product but do not answer the allergen question — a good grader returns Ambiguous or Incorrect, and the system consults an external source rather than reassuring the user from a stale table. And if the product is not in the corpus at all, Incorrect prevents the single worst behaviour: a confident safety claim assembled from a similar-sounding product's data. That scenario also shows the limits. External fallback is not free correctness. Web results about a food product may be user-generated, out of date or wrong, and in a safety-critical domain the right corrective action for Incorrect may be to refuse and escalate to a human rather than to search the open web. CRAG's structure is the contribution; the choice of fallback source is yours. ## Operational consequences **Thresholds are a product decision.** Where you cut the score between Correct, Ambiguous and Incorrect trades precision against coverage. Aggressive Incorrect thresholds mean more fallbacks — more latency, more cost, more exposure to external sources — but fewer confident answers from junk. In a regulated domain that trade is usually worth making. **The grader must be cheap.** It runs on every question, so a heavy grader doubles the cost of the whole system. In practice teams use a small fine-tuned scorer or a compact model, and validate its agreement with human relevance judgements before trusting it. **External text is untrusted input.** Falling back to the open web means content you do not control enters the prompt, which is a prompt-injection surface. Treat fetched text as data, never as instructions, and strip or sandbox it accordingly. **The grader is now a monitored component.** The distribution of grades is a first-class operational signal. A rising share of Incorrect grades on internal retrieval usually means the corpus has drifted out of date or that a new class of question has appeared, and it is often the earliest warning you get. ## How CRAG differs from neighbouring ideas It is easy to blur CRAG with self-reflective designs. The distinction to state in an interview: CRAG grades *retrieval*, before generation, with a dedicated evaluator, and its corrective action is to change where the knowledge comes from. Self-critique designs grade the *generated text* against the evidence. They compose well — grade the evidence, then grade the grounding — but they answer different questions, and only CRAG has a defined path for "our corpus does not know this".