RAG Node

Retrieval-augmented generation, minus the infrastructure

The LangChain course's RAG lessons built retrieval with real document loaders, splitters, and a vector store. The underlying idea is simpler than that infrastructure suggests: find the few pieces of text most relevant to a question, then ask the model to answer using only those pieces. This lesson keeps that idea, but skips the vector store, no new dependency is needed to demonstrate why retrieval matters.

A tiny local "document store"

DOCUMENTS = [
    "The office recycling bins are collected every Tuesday and Friday morning.",
    "Employees get 20 days of paid vacation per year, accrued monthly.",
    ...
]

Just a Python list of strings. No embeddings API call, no database, this is enough to prove the point: the generation node's answer quality depends entirely on what the retrieval node hands it.

The retrieval node

def retrieve(state: State) -> dict:
    question_words = set(state["question"].lower().split())
    scored = sorted(
        DOCUMENTS,
        key=lambda doc: len(question_words & set(doc.lower().split())),
        reverse=True,
    )
    return {"retrieved": scored[:2]}

Deliberately simple keyword scoring: count how many words the question and each document share, keep the top two. Real systems typically use embeddings and vector similarity for this instead of exact word overlap, but the shape of the step, "score every candidate, keep the best few," is identical either way.

The generation node

def generate(state: State) -> dict:
    context = "\n".join(f"- {doc}" for doc in state["retrieved"])
    prompt = (
        "Answer the question using ONLY the context below. If the context "
        "doesn't contain the answer, say you don't know.\n\n"
        f"Context:\n{context}\n\nQuestion: {state['question']}"
    )
    response = model.invoke(prompt)
    return {"answer": response.text}

generate never sees the full DOCUMENTS list, only whatever retrieve decided was relevant, the same separation of concerns as any two-node pipeline since Beginner Lesson 3. The instruction to answer "using ONLY the context" is what makes the retrieval step actually matter, without it, the model might just answer from its own general knowledge and mask a bad retrieval.