langchain-openai: support custom fallback content for tool-only messages

[louisgthier]

Some APIs (notably Gemini, when accessed via the OpenAI-compatible endpoint) require every message to have a non-null, non-empty string as content. However, messages that contain only tool calls (e.g., ToolMessage) often have content=None, leading to a 400 Bad Request with:

Error code: 400 - [{‘error’: {‘code’: 400, ‘message’: ‘Expected string or list of content parts, got: null’, ‘status’: ‘INVALID_ARGUMENT’}}]

This PR adds support for specifying a fallback string (e.g. " " or "N/A") to populate the content field of tool-only messages, ensuring compatibility with stricter API implementations.

Issue:
N/A — discovered while integrating LangChain with Gemini API.

Dependencies:
None

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

Name	Status	Preview	Comments	Updated (UTC)
langchain	✅ Ready (Inspect)	Visit Preview	💬 Add feedback	Apr 13, 2025 2:07am

[ccurme]

Tool calling is supported for Gemini models via the langchain-google-genai and langchain-google-vertexai packages, depending on how you are authenticating. See docs here. Will these work for your use-case?

Many providers offer a chat-completions API such that you can use the OpenAI SDK to interact with them. But they often feature small differences like this. If the dedicated Google packages will work for you, that's preferable to attempting to support more providers via ChatOpenAI.

[louisgthier]

I gave it a try, and the issue seems to be present in both langchain-google-genai and langchain-openai. Tool calling generally works fine, except in cases where the LLM returns only tool calls without any content—that’s when it breaks.

Since I frequently switch between models, and most providers support the OpenAI SDK, I prefer sticking with that—it works out of the box and avoids these edge cases. I get that keeping things simple is important, and adding parameters everywhere isn’t ideal.

Maybe when the provider is detected to be Google, we could just make sure the content field is never left empty?

It's not a huge issue anyway—I ran into it using the built-in LangGraph React Agent, and I worked around it by adjusting the messages using a "prompt" pre-hook.

[ccurme]

Thanks @louisgthier. Are you able to provide a minimal reproducible example? This would help us figure out the best solution.

[louisgthier]

Thanks @louisgthier. Are you able to provide a minimal reproducible example? This would help us figure out the best solution.

from langchain_core.tools import tool
from langchain_core.messages import HumanMessage, SystemMessage
from langchain_google_genai import ChatGoogleGenerativeAI
from langgraph.prebuilt import create_react_agent
from langgraph.checkpoint.memory import MemorySaver
from typing import List, Any
import os

# 1. Define a simple tool
def add_numbers(a: int, b: int) -> int:
    """Add two numbers and return the result."""
    return a + b
add_numbers = tool(add_numbers)

# 2. Set up the Gemini model
model_name = "gemini-2.0-flash"
api_key = os.getenv("GEMINI_API_KEY")

llm = ChatGoogleGenerativeAI(
    model=model_name,
    google_api_key=api_key,
)
llm = llm.bind_tools([add_numbers])

# 3. Define a minimal state class
class State(dict):
    messages: List[Any]
    remaining_steps: int

# 4. Create the agent
graph = create_react_agent(
    llm,
    [add_numbers],
    state_schema=State,
    checkpointer=MemorySaver()
)

def main():
    # 5. Prepare the initial state
    system = SystemMessage(content="You are a helpful assistant. Use tools if needed.")
    user = HumanMessage(content="Please use the add_numbers tool to add 2 and 3.")
    state = State(messages=[system, user], remaining_steps=5)  # 5 is arbitrary, just needs to be present

    # 6. Run the agent
    result = graph.invoke(state, config={"configurable": {"thread_id": "test-thread"}})
    print("Final messages:")
    for msg in result['messages']:
        print(f"{type(msg).__name__}: {getattr(msg, 'content', msg)}")

if __name__ == "__main__":
    main()

When running this simple script, I get:

langchain_google_genai.chat_models.ChatGoogleGenerativeAIError: Invalid argument provided to Gemini: 400 * GenerateContentRequest.contents: contents is not specified

[louisgthier]

Please note I have the same result with ChatOpenAI

[ccurme]

Hi @louisgthier,

Thanks very much for the minimal example. I've simplified it a bit further using documented usage patterns, and it appears to run:

from langchain_core.messages import HumanMessage, SystemMessage
from langchain_google_genai import ChatGoogleGenerativeAI
from langgraph.prebuilt import create_react_agent
from langgraph.checkpoint.memory import MemorySaver


def add_numbers(a: int, b: int) -> int:
    """Add two numbers and return the result."""
    return a + b

llm = ChatGoogleGenerativeAI(model="gemini-2.0-flash")

graph = create_react_agent(
    llm,
    [add_numbers],
    checkpointer=MemorySaver(),
    prompt="You are a helpful assistant. Use tools if needed.",
)

config = {"configurable": {"thread_id": "abc123"}}
for step in graph.stream(
    {"messages": [HumanMessage(content="Please use the add_numbers tool to add 2 and 3.")]},
    config,
    stream_mode="values",
):
    step["messages"][-1].pretty_print()

• If this doesn't work but captures your intent, can you share what versions of packages you are running? The output of python -m langchain_core.sys_info might suffice.
• If this doesn't capture your intent, could you update it or clarify?