Retrieval-Augmented Generation (RAG)

A language model on its own only knows what it was trained on. RAG fixes that by adding a retrieval step: when a question comes in, the system first searches a vector database of your own content — help docs, product manuals, contracts, transcripts — pulls the most relevant passages, and then asks the model to answer using those passages as context. The model's general fluency stays; the facts come from your data.

RAG is the dominant pattern for enterprise AI today because it solves three real problems at once: hallucination (the model is grounded in retrieved text), freshness (you index new documents instead of retraining the model), and privacy (your data never enters the model's training set). A typical stack is an embedding model, a vector database (Pinecone, pgvector, Weaviate), a retriever, and the LLM.

RAG is not magic. Retrieval quality is the ceiling — if the right passage does not surface, the model will improvise or refuse. Most production failures trace back to chunking strategy, embedding choice, or messy source documents, not the LLM itself. Invest in evaluation harnesses that test retrieval before you tune the prompt.