Package (Required)
Checked other resources
Example Code (Python)
## Reproduction (no API key required)
from types import SimpleNamespace
from langchain_google_genai.chat_models import _parse_response_candidate
fc = SimpleNamespace(name="echo", args={"text": "안녕하세요"})
part = SimpleNamespace(
text=None, thought=False, thought_signature=None,
executable_code=None, code_execution_result=None,
inline_data=None, function_call=fc,
)
candidate = SimpleNamespace(content=SimpleNamespace(parts=[part]))
msg = _parse_response_candidate(candidate, streaming=False, model_name="gemini-2.5-flash")
assert msg.tool_calls[0]["args"] == {"text": "안녕하세요"} # OK
assert msg.additional_kwargs["function_call"]["arguments"] == '{"text": "안녕하세요"}' # FAILS
# Actual: '{"text": "\\uc548\\ub155\\ud558\\uc138\\uc694"}'Error Message and Stack Trace (if applicable)
Description
Repository / Version
- Repo: langchain-ai/langchain-google
- Package: langchain-google-genai
- Installed version (where reproduced): 4.2.2
Buggy location
File: libs/genai/langchain_google_genai/chat_models.py
# libs/genai/langchain_google_genai/chat_models.py
if part.function_call:
function_call = {"name": part.function_call.name}
# dump to match other function calling llm for now
# Convert function call args to dict first, then fix integer-like floats
args_dict = dict(part.function_call.args) if part.function_call.args else {}
function_call_args_dict = _convert_integer_like_floats(args_dict)
function_call["arguments"] = json.dumps( # ← bug
{k: function_call_args_dict[k] for k in function_call_args_dict}
)
additional_kwargs["function_call"] = function_calljson.dumps()is called without ensure_ascii=False. Python's default (True) escapes all non-ASCII characters to \uXXXX. The resulting escaped string lands in AIMessage.additional_kwargs["function_call"]["arguments"].
Reproduction (minimal)
from langchain_google_genai import ChatGoogleGenerativeAI
from langchain_core.tools import tool
@tool
def echo(text: str) -> str:
"""Echo the given text."""
return text
llm = ChatGoogleGenerativeAI(model="gemini-2.5-flash").bind_tools([echo])
msg = llm.invoke("Call the echo tool with the text: 안녕하세요")
print("tool_calls:", msg.tool_calls)
#→ tool_calls: [{'name': 'echo', 'args': {'text': '안녕하세요'}, ...}] ✅ correct (decoded back)
print("additional_kwargs:", msg.additional_kwargs)
#→ {'function_call': {'name': 'echo',
# 'arguments': '{"text": "\\uc548\\ub155\\ud558\\uc138\\uc694"}'}} ❌ escapedSame input via langchain_openai.ChatOpenAI produces arguments with raw UTF-8.
Why it matters
- Persistence layers downstream (DBs with JSON columns, file logs) store the escape sequences verbatim, making CJK / emoji / accented strings unreadable when inspected directly.
- Data duplication inconsistency: the same arguments are exposed twice on the same message — once correctly in tool_calls[i].args (clean dict, because parse_tool_calls round-trips through json.loads) and
once incorrectly in additional_kwargs.function_call.arguments (escaped string). Consumers see different content depending on which field they read.
- Cross-provider inconsistency: langchain_openai already passes ensure_ascii=False in the analogous spot (langchain_openai/chat_models/base.py:3323):
"arguments": json.dumps(tool_call["args"], ensure_ascii=False),
- and langchain_core consistently passes ensure_ascii=False in all 8 of its json.dumps call sites that touch message content. This makes the google_genai line an outlier.
Proposed fix (one-line)
function_call["arguments"] = json.dumps(
{k: function_call_args_dict[k] for k in function_call_args_dict},
ensure_ascii=False,
)
Package (Required)
Checked other resources
Example Code (Python)
Error Message and Stack Trace (if applicable)
Description
Repository / Version
Buggy location
File: libs/genai/langchain_google_genai/chat_models.py
json.dumps()is called withoutensure_ascii=False. Python's default (True) escapes all non-ASCII characters to \uXXXX. The resulting escaped string lands inAIMessage.additional_kwargs["function_call"]["arguments"].Reproduction (minimal)
Same input via langchain_openai.ChatOpenAI produces arguments with raw UTF-8.
Why it matters
once incorrectly in additional_kwargs.function_call.arguments (escaped string). Consumers see different content depending on which field they read.
"arguments": json.dumps(tool_call["args"], ensure_ascii=False),
Proposed fix (one-line)