Retrieval as a Tool

Where we left off

Lesson 3's get_current_temperature tool was a placeholder, a FunctionDeclaration describing a function that never actually ran. This lesson swaps it for a real one: search_notes(), everything naive_rag's retrieve() already does, embed the query, rank fixture notes by cosine similarity, keep the top k, wrapped so it can be declared as a tool and genuinely executed once the model asks for it.

The code, piece by piece

def search_notes(query: str, store: list[dict], k: int = 2) -> str:
    ...
    return "\n\n---\n\n".join(f"[{r['source']}]\n{r['text']}" for r in top_k)

Same retrieval math every earlier course used, cosine similarity over embedded fixture notes, with one change: it returns a plain string instead of a list of dicts. Tool results eventually get handed back to Gemini as text (Lesson 5), so search_notes() does that formatting itself, tagging each passage with its source file, rather than leaving formatting to whatever calls it.

SEARCH_NOTES_DECLARATION = types.FunctionDeclaration(
    name="search_notes",
    description=(
        "Search a personal notes collection (...) for passages relevant "
        "to a query. Use this only for questions about these specific "
        "personal topics, not for general knowledge questions."
    ),
    ...
)

The description here is doing real work, it's the only information the model has about when to reach for this tool. Naming the five topics covered and explicitly steering away from general-knowledge questions is what gives the model a fighting chance at Lesson 6's job: correctly declining to call this tool when it isn't needed.

result = search_notes(call.args["query"], store)

This lesson still does the "actually run the function" step by hand, one if branch, one call. That's deliberately not yet a general loop, seeing it done manually once, for one tool, makes Lesson 5's loop version legible instead of magic.

Why the model's chosen query matters

Watch what call.args["query"] actually contains when you run this: it's often not a verbatim copy of the original question. Gemini frequently rewrites it into something closer to what the notes themselves would use ("sourdough starter feeding schedule" instead of "How often does Clarence need feeding"). That's the model doing its own lightweight query optimization, for free, as a side effect of deciding what to search for, something the fixed pipeline's retrieve(query, ...) never had the chance to do because query was always just the original question, verbatim.