RAG (2020): Retrieval-Augmented Generation

RAG retrieves relevant documents for LLMs. Introduced in the 2020 paper "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks" by Patrick Lewis and colleagues, submitted to arXiv on May 22, 2020.

Before RAG, language models could only answer from training data. RAG added a retrieval step: search a document index, fetch relevant passages, and pass them to the model at inference time. That single change made LLMs usable for enterprise Q&A, support tools, and domain-specific products.

def rag_answer(query: str, retriever, llm):
    # Step 1: Retrieve relevant documents
    documents = retriever.search(query, top_k=5)

    # Step 2: Build context from retrieved passages
    context = "\n".join(doc.content for doc in documents)

    # Step 3: Generate answer grounded in context
    prompt = f"Context:\n{context}\n\nQuestion: {query}\nAnswer:"
    return llm.generate(prompt)

The retrieve-then-generate pattern remains the backbone of production AI today. Vector search, BM25, reranking, and chunking strategies all exist to make this retrieval step more accurate.

Every layer that came after RAG assumes retrieval exists. LangChain wraps it in chains. LangGraph uses it inside agent nodes. GraphRAG extends what gets retrieved. But the core idea from 2020 never went away: ground the LLM in real documents before it answers.