Wiring It Into LangChain

Where we left off

Lesson 15 built browse(url) as a plain function and proved it works entirely on its own, no AI involved. This lesson takes that exact same function, adds @tool, and hands it to a Gemini model, following the same pattern as Lessons 13 and 14 of the langchain course: build the tool plainly first, then connect it.

This lesson needs a GOOGLE_API_KEY in .env at the project root (see Setup in the root README); it's the first lesson in this course that makes a real call to an LLM.

The code, piece by piece

@tool
def browse(url: str) -> str:
    """Open a URL in a browser and return its visible, readable page text."""
    ...

The body of browse is unchanged from Lesson 15, character for character. @tool only adds metadata around it: a name (browse), a description (this docstring), and an argument schema (url: str). Nothing about how the function itself works had to change to make it usable by a model, that's what building it plainly first bought us.

model = ChatGoogleGenerativeAI(model="gemini-3.5-flash-lite")
model_with_tools = model.bind_tools([browse])

Exactly the same call as Lesson 14 of the langchain course, just with browse in place of calculator. bind_tools describes the tool to the model, it does not hand the model a real browser. The model can never open a page itself, it can only ask.

The three rounds, applied to a browser

Round 1: ask the question.

ai_message = model_with_tools.invoke(messages)

We ask what the first quote on a specific page is. The model has no reliable way to know the live content of that page (it wasn't necessarily part of its training data, and the page could change anyway), so instead of guessing, it should reply with a request: "call browse with url='https://quotes.toscrape.com/'."

Round 2: we run the browser, and report back.

for call in ai_message.tool_calls:
    result = browse.invoke(call["args"])
    messages.append(ToolMessage(content=result, tool_call_id=call["id"]))

This is the moment a real Chromium instance actually launches, on our machine, not the model's. The model asked, our code decided to honor that request and did the work. browse.invoke(...) here runs the exact same browse function Lesson 15 tested directly, no difference at all in how it executes.

Round 3: let the model read the page and answer.

final_response = model_with_tools.invoke(messages)

We send the full conversation, including the raw page text browse returned, back to the model. It reads through that text and picks out just the quote and its author, in a normal sentence, instead of us having to write parsing code to find "the first quote" ourselves.