Day 17 — Corrective RAG (CRAG)

Traditional RAG Blind Spot

🔧

Blindly Trusting Retrieved Documents

Traditional RAG has a fundamental flaw — it always trusts whatever documents the retriever returns. If retrieval returns wrong or irrelevant documents, a wrong answer is guaranteed. There is no mechanism to catch or correct this failure.

⚠️ In business applications, confidently wrong answers have serious consequences. Corrective RAG eliminates this risk with a self-correction mechanism.

What CRAG Does

🏗️

Evaluate, Then Decide What to Do

CRAG adds a Retrieval Evaluator that judges whether retrieved documents are actually relevant before using them. Based on that judgment, it takes one of three different actions.

✓Correct — Documents are relevant — proceed with normal RAG pipeline

→Ambiguous — Partially relevant — combine retrieved docs with web search results

✕Incorrect — Documents are irrelevant — discard them and use web search instead

Knowledge Refinement

🔨

Cleaning What Is Retrieved

Even when documents pass the relevance check, they may contain mixed content. Knowledge Refinement cleans this up in three steps.

DecomposeBreak the document into individual sentences

FilterScore each sentence for relevance to the query — keep only relevant ones

RecomposeJoin the relevant sentences back into clean focused context

Web Search Integration

🌐

Fallback to Live Web Search

When retrieved documents are insufficient, CRAG queries the web using services like Tavily. The query is first rewritten into a form optimized for search engines — producing better results than the raw user question.

💡 Query rewriting example: "latest AI news" → "recent AI model releases and developments in the last 30 days" — far more effective for search engines.

CRAG in LangGraph

📊

The Graph Architecture

·retrieve — Fetch documents from vector store

·evaluate — Score each document for relevance (0 to 1)

·refine — Extract only relevant sentences from documents

·web_search — Query web when documents are incorrect or ambiguous

·generate — Produce final answer from clean verified context

✦

Corrective RAG transforms traditional RAG from a trust-everything system into a self-correcting one. The retrieval evaluator judges relevance, knowledge refinement cleans the context, and web search provides a reliable fallback. The result is a system that never blindly trusts bad retrieval — and almost never produces hallucinated answers.

← Day 16 Day 17 / 20 Day 18 →