RAG Architecture

Retrieval-augmented generation (RAG) is the most important architecture pattern in applied LLM engineering. It solves the model’s two biggest limitations at once: it doesn’t know your private data, and it doesn’t know anything after its training cutoff. RAG fixes both by fetching relevant information at request time and putting it in the prompt.

Why RAG instead of fine-tuning

Fine-tuning teaches behavior; RAG supplies knowledge. RAG data is always current, every answer can cite its source, access control is enforceable, and there’s no retraining when facts change. For “the model needs to know X,” RAG is almost always the right tool.

In this section

The RAG Pipeline The two phases — indexing and retrieval-plus-generation — and a working baseline implementation.

Chunking & Retrieval Chunking strategies, hybrid search, query transformation, and reranking — the levers of retrieval quality.

Advanced RAG & Evaluation Beyond naive RAG — advanced patterns, RAG evaluation metrics, and a failure-mode playbook.

What you’ll be able to do

Build a RAG system from scratch, diagnose why a RAG system gives bad answers, and apply the right fix — because “RAG isn’t working” almost always has a specific, locatable cause.

Prerequisites

Vector Databases and LLM Engineering.