Corrective RAG

A linear, one-concept-per-lesson path through Corrective RAG (CRAG), grading retrieved chunks for relevance and correcting a bad retrieval before it ever reaches generation, built from scratch with no framework, from your first relevance grade to a complete FastAPI corrective RAG service backed by ChromaDB. 26 lessons · 3 tiers.

Prerequisites: Completion of Naive RAG, or already comfortable with embeddings, cosine similarity, top-k retrieval, and why naive retrieval can confidently return the wrong chunk - Lesson 2 here is a compressed recap of that failure, not a full re-teach.

Lessons use Google's Gemini free tier (gemini-embedding-001 for embeddings, gemini-3.5-flash-lite for chat and grading), the same models as this site's other RAG courses. No Docker and no database through most of the course; the Advanced tier graduates the vector store to a local ChromaDB instance, still no server or account required.

Built on ChromaDB, the open-source vector database this course graduates to in the Advanced tier.