How to disable automatic_function_calling with ChatGoogleGenerativeAI

OSS Product Help LangGraph
1 
ChatGoogleGenerativeAI(
    project=<gcp-project>
    model=<gcp-location>
    temperature=0.0,
    max_retries=2,
    include_thoughts=False,
    thinking_budget=0,
    vertexai=True,
)
2

hi @gs-awesome

If what you actually want is “force the model to not produce any tool call on this invocation”, use tool_choice="none". That maps to Gemini’s FunctionCallingConfigMode.NONE.

from langchain_google_genai import ChatGoogleGenerativeAI

llm = ChatGoogleGenerativeAI(
    model="gemini-2.5-flash",
    project="my-gcp-project",
    location="us-central1",
    vertexai=True,
    temperature=0.0,
)

# Bind the tools so the schema is there if you want it back later,
# but force this call to ignore them:
llm_no_tools = llm.bind_tools([my_tool_1, my_tool_2], tool_choice="none")
response = llm_no_tools.invoke("Summarize this document ...")
# response.tool_calls == []

or

llm_with_tools = llm.bind_tools([my_tool_1, my_tool_2])
response = llm_with_tools.invoke(
    "Summarize this document ...",
    tool_choice="none",
)
3

@pawel-twardziak

well google-genai sdk has disabling automatic function call. this is the feature I was looking.

4

if you use LangChain, it’s safer to do that in the langchain abstraction layer

5

@pawel-twardziak did you mean like this?

from langchain_google_genai import ChatGoogleGenerativeAI

class _NoAFCModel(ChatGoogleGenerativeAI):

   def _prepare_request(self, *args, **kwargs):
        request = super()._prepare_request(*args, **kwargs)
        request["config"].automatic_function_calling = AutomaticFunctionCallingConfig(disable=True)
        return request

at least this seems to be working. after some investigation I realize that langchain_google_genai doesn’t support disable automatic function call options (at least)

6

try this @gs-awesome

from google.genai.types import AutomaticFunctionCallingConfig
from langchain_google_genai import ChatGoogleGenerativeAI

llm = ChatGoogleGenerativeAI(
    model="gemini-2.5-flash",
    project="my-gcp-project",
    location="us-central1",
    vertexai=True,
)

llm_no_afc = llm.bind(
    automatic_function_calling=AutomaticFunctionCallingConfig(disable=True)
)

# Or per-invocation (same net effect):
llm.invoke(
    "...",
    automatic_function_calling=AutomaticFunctionCallingConfig(disable=True),
)
7

@pawel-twardziak thanks a lot
and how to do for structure output

from pydantic import BaseModel
class OutputSchema(BaseModel):
        person: str

llm.with_structured_output(OutputSchema, method="json_schema", include_raw=True)
8

I think with method="json_schema" there is no config.tools at all.
So for method="json_schema", disabling AFC has no behavioral effect imho.

OR

from google.genai.types import AutomaticFunctionCallingConfig
from langchain_google_genai import ChatGoogleGenerativeAI
from pydantic import BaseModel


class OutputSchema(BaseModel):
    person: str


llm = ChatGoogleGenerativeAI(
    model="gemini-2.5-flash",
    project="my-gcp-project",
    location="us-central1",
    vertexai=True,
)

llm_json = llm.bind(
    response_mime_type="application/json",
    response_json_schema=OutputSchema.model_json_schema(),
    automatic_function_calling=AutomaticFunctionCallingConfig(disable=True),
)

raw = llm_json.invoke("Tell me about Alice.")
parsed = OutputSchema.model_validate_json(raw.content)

OR

your workaround does the job (it’s valid by the source code)

9

@pawel-twardziak your solution also works too :+1: , and method=”json_schema” do exist. I have been mostly using with_structured_output because it is much more reliable to get the structure data from model, specifically for multi-turn conversation with model.

1 Like
10

Alright @gs-awesome :slight_smile:
Have you managed to sort out your problem? Anything else I can help with? :slight_smile: