RAG

A technique that enhances LLM responses by retrieving relevant information from external knowledge bases. Combines the power of search with generative AI for grounded, factual outputs.

Retrieval-augmented generation, RAG, is the architecture that gives a language model the facts it did not memorize: at query time, relevant documents are retrieved, usually by vector similarity, and injected into the model's context, grounding the answer in your data instead of the model's training corpus.

Why it matters to an indie builder: it is the difference between a chatbot that hallucinates and a product that answers from your documentation, contracts or catalog. RAG is also the cheapest path to an AI feature, no fine-tuning, no GPU bill, just chunking, embeddings and retrieval, all available as managed services.

A realistic pipeline: documents split into overlapping chunks, embedded into a vector store, retrieved by top-k similarity with a reranker, then assembled into a prompt with citations back to sources. Evaluation on a held-out question set tells you whether chunk size or retrieval depth needs tuning.

Trade-offs: retrieval quality dominates answer quality, and naive RAG fails on multi-hop questions and tables. Compare with long-context models that simply read everything, and with fine-tuning when the goal is style or behavior rather than knowledge.