RAG as a Tool

Where we left off

Lesson 28 built a vector store and searched it manually, we wrote the query ourselves and called similarity_search directly. This lesson wraps that same search as a tool, the exact @tool pattern from Lesson 13, so an agent can decide for itself, per question, whether it even needs to search at all.

The RAG tool

@tool
def search_personal_notes(query: str) -> str:
    """Search the user's personal notes for information relevant to the
    query. Use this for questions about the user's projects, garden,
    recipes, hobbies, or anything that sounds like it might be
    documented in their personal notes rather than general knowledge."""
    results = vector_store.similarity_search(query, k=2)
    if not results:
        return "No relevant notes found."
    return "\n\n".join(result.page_content for result in results)

This should look familiar: it's the exact shape of every tool since Lesson 13, a function, decorated with @tool, with a docstring describing when to use it. The only thing new here is what the tool actually does internally: instead of a calculator or a word counter, this one runs Lesson 28's similarity search and hands back whatever chunks it finds. From the agent's point of view, this is just another tool, no different in kind from calculator.

Notice the docstring is doing real work

"""Search the user's personal notes for information relevant to the
query. Use this for questions about the user's projects, garden,
recipes, hobbies, or anything that sounds like it might be
documented in their personal notes rather than general knowledge."""

Recall from Lesson 13: the AI never sees your implementation, only the name, description, and arguments. This docstring is doing the actual work of telling the agent when this tool is relevant, personal, specific, documented information, as opposed to general world knowledge it already knows. A vague docstring here (like just """Search notes.""") would leave the agent guessing, and it might reach for this tool unnecessarily, or fail to reach for it when it should.

Two questions, two different outcomes

ask("How often do I practice my cello, and for how long?")
ask("What is the capital of Japan?")

The first question can only be answered by actually searching notes.txt, there's no other way the agent could know this specific detail. The second is ordinary general knowledge, the agent answers directly, the same "choosing not to use a tool" behavior from Lesson 15. Run the lesson and you'll see both outcomes: an exact, correct answer pulled from the notes for the first, and a direct answer with no tool call for the second.

What this actually demonstrates

This is RAG (Retrieval-Augmented Generation), in full, wired into an agent: the model can answer questions using information it was never trained on (your own personal notes), by retrieving just the relevant pieces instead of needing the entire document stuffed into every request. Combined with everything else in this course so far, tool calling, memory, context management, this is a genuinely production- shaped pattern: an agent that knows both general knowledge and your own private documents, and decides for itself which one a given question needs.