What is the core mechanism that distinguishes Self-RAG (Asai et al. 2023) and Corrective-RAG (CRAG, 2024) from a standard single-shot retrieve then generate pipeline?
Option C is correct. Both Self-RAG and CRAG add a quality check step that standard single-shot RAG does not have.
Standard RAG runs one retrieval, takes whatever chunks come back, and asks the generator to write an answer. If the retrieved chunks are irrelevant, the answer will be wrong or hallucinated.
Self-RAG, from Asai et al. 2023, fine-tunes the generator to emit special reflection tokens during decoding. [Retrieve] decides whether to retrieve at all. [Relevant] scores each retrieved chunk. [Supported] checks whether the generated answer is grounded in the chunks. A critique token scores the final answer.
CRAG (Corrective RAG, 2024) is lighter. It runs a small evaluator model over the retrieved chunks and gets a relevance score. On low scores, it triggers a corrective action: a web search via Google or Brave, a rewritten query, or a refusal.
The shared idea is the same: do not blindly use whatever came back. Inspect retrieval quality and act on it.
**Option C** captures the shared mechanism. Both methods add a self-reflective check on retrieval quality and take a corrective action when quality is low.
Standard RAG retrieves once and generates. The assumption is that retrieval always returns relevant chunks. That assumption fails on out of distribution questions, ambiguous queries, or stale corpora, and the failure mode is silent hallucination.
Self-RAG fine-tunes the generator to emit four reflection tokens during decoding. `[Retrieve]` decides whether to retrieve. `[Relevant]` scores chunk relevance. `[Supported]` checks whether the answer follows. A critique token scores the final answer's utility. The generator can skip retrieval, choose between candidate chunks, and self-flag low-confidence outputs, all within one autoregressive pass.
CRAG is the lighter weight cousin. A small T5-based evaluator scores retrieval relevance as Correct, Ambiguous, or Incorrect. On Incorrect, the system falls back to a web search (Google, Brave) or rewrites the query. On Ambiguous, it combines the original chunks with web results.
The other options are wrong: (A) neither requires a larger generator. (B) parallel retrieval is multi-query / HyDE, a different technique. (D) is invented; embedding-based retrieval is still the substrate.
Quick reference
TL;DR
Both add a self-reflective quality check on retrieval and act on it: Self-RAG via reflection tokens, CRAG via a lightweight evaluator. Standard RAG just trusts whatever came back.
Imagine you ask a friend to look something up in a library. Standard RAG is a friend who grabs the first books off the shelf and reads from them no matter what. Self-RAG and CRAG are friends who pause and ask themselves: are these books actually about the question? If not, they go back, try a different shelf, or search online. Self-RAG was trained to spend extra effort emitting little check mark words as it works, like 'do I even need a book here?' and 'does the book actually answer this?' CRAG is simpler: it has a small helper that scores how relevant the books look, and if the score is low, it switches strategies. Both share one big idea: do not blindly trust the first results.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
Corpus is curated, narrow, and high-quality (e.g. internal API docs) Retrieval is reliable by construction; adding a gate burns cost without catching failures that do not happen.
Latency budget is sub-second and adding an evaluator call is not feasible CRAG's evaluator adds 50-150 ms; Self-RAG's reflection tokens lengthen decoding. Both are visible in P95 latency.
Web search is not allowed (data residency, privacy, compliance) CRAG's corrective web fallback is its strongest action; without it, the method degrades to a relevance filter only.
You cannot fine-tune the generator (closed weight model, no GPUs) Self-RAG specifically requires reflection token fine-tuning. Use CRAG or tool use loop patterns instead.
Quick scan · 2–4 min
Everything important, quickly.
8 min: explain the shared insight, walk Self-RAG's four reflection tokens, walk CRAG's three-grade evaluator, contrast with multi-query, and close with the 2026 tool-use reproduction.
Real products, models, and research that use this idea.
1Self-RAG paper by Asai et al. ships a 7B and 13B fine-tuned LLaMA-2 model that emits the four reflection tokens during decoding.
2CRAG paper by Yan et al. uses a T5-large evaluator and falls back to Google Search via the Serper API for the corrective web step.
3LangChain's adaptive rag template implements a CRAG-style gate using a grader then search router.
4LlamaIndex ships a Corrective RAG query engine that wraps any retriever with a relevance score then decide gate.
5Anthropic Claude 4.7 contextual retrieval combined with tool-use achieves Self-RAG-like gating without fine-tuning: the model decides when to call the retrieve_documents tool.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QHow would you implement Self-RAG-style gating on top of a closed weight frontier model like Claude Opus 4.7 or GPT-5.5?
Sign in or go Pro to see the approach for this follow-up.
QCRAG's evaluator is one extra inference call per query. How would you train it, and what is the latency budget?
Sign in or go Pro to see the approach for this follow-up.
QCompare the failure modes of Self-RAG and CRAG when the corpus and the question are both out of distribution.
Sign in or go Pro to see the approach for this follow-up.
QHow does Adaptive RAG combine Self-RAG / CRAG patterns with multi-hop decomposition?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Conflating these with multi-query retrieval. Multi-query expands the question into paraphrases and runs them in parallel; Self-RAG and CRAG inspect the results and react to quality.