RAG in Production

What RAG Is (and the Problem It Solves)

Start here, even if RAG is a new acronym. What 'Retrieval-Augmented Generation' actually means in plain English, the problem it fixes, and a playground where you can watch retrieval happen.

When to Use RAG — and Which Projects Should

RAG is not always the answer. The clear decision guide: RAG vs long context vs fine-tuning, the project types that genuinely need it, and the ones that don't.

From Demo to System: The Production Architecture

A prototype is a straight line; production is two pipelines and a feedback loop. The full architecture at a glance, before we build each piece.

Ingestion & Chunking — Where Quality Is Won

Retrieval can only find what ingestion prepared well. Structure-aware parsing, chunking (with an interactive sizer), metadata, and keeping the index fresh.

Embeddings & the Vector Store

How text becomes searchable geometry, how to choose an embedding model without regret, and what a production vector store must do.

Retrieval That Actually Works

Naïve top-k is the floor. The stacking techniques that win most of your quality: hybrid search, re-ranking, and query transformation.

Generation & Grounding — Making Answers Trustworthy

Good retrieval is wasted if generation ignores it or invents around it. Citations, honest refusals, context ordering, and defending against injection through documents.

Evaluation — You Can't Improve What You Don't Measure

The step that turns RAG from prompt-roulette into engineering: retrieval vs generation metrics, the golden set, LLM-as-judge, and gating changes in CI.

The Strange Scenarios (Where Naïve RAG Breaks)

The edge cases that quietly break RAG in production — stale and contradictory data, access control, PII, out-of-scope questions, duplicates, and adversarial documents — and how to handle each.

Operating It: Cost, Observability & the Frontier

A live RAG system is something you operate, not just ship — tracing, caching, cost control, the feedback loop — plus where RAG is heading next.