Function Calling in LangChain

1. The Old Way: Text Parsing

In 2022, if you wanted an LLM to search Wikipedia, you prompted: "If you need to search, output JSON like {"action": "search", "query": "..."}"

Then you wrote a messy Regex parser to find that JSON in the chaotic output. It failed constantly. The LLM would forget a quote, add a comment, or hallucinate a key.

2. The New Way: Function Calling (Tool Binding)

Modern Models (GPT-4o, Claude 3.5 Sonnet) have been fine-tuned to output actions natively. They don't output "text"; they output a structured "Tool Call Object."

The Workflow:

1. Bind: You send the user query + a list of available tool schemas.
2. Decide: The Model thinks: "I cannot answer 'What is stock AAPL?' with my internal weights. I need the get_stock tool."
3. Output: The Model stops graduating text. It returns a tool_call payload: name='get_stock', args={'symbol': 'AAPL'}.
4. Execute: LangChain (the Runtime) sees this signal. It pauses the LLM, runs the Python function get_stock("AAPL"), and gets $200.
5. Resume: LangChain sends the result $200 back to the LLM.
6. Answer: The LLM now has the context. It replies: "Apple stock is currently $200."

3. Implementation

from langchain_openai import ChatOpenAI

llm = ChatOpenAI(model="gpt-4o")
llm_with_tools = llm.bind_tools([search_google, get_weather])

response = llm_with_tools.invoke("What is the weather in Paris?")
# Content=" " (Empty content!)
# Tool_Calls=[{'name': 'get_weather', 'args': {'city': 'Paris'}}]

4. Why This is Safer

It eliminates "Syntax Errors." The model is constrained to generate valid JSON that matches your Pydantic schema. You are no longer parsing natural language; you are handling an event stream.

5. Summary

Function Calling turns the LLM into a Router. It treats your API tools as extensions of its own vocabulary.

Key Intuition: "The model doesn't run the code. It just pushes the button."