Intermittent "STOP" without tool_calls in gemini-2.5-flash using LangChain

Gemini API
, ,
1

Hi,

I am facing a critical issue where gemini-2.5-flash intermittently fails to identify/call tools and simply returns a “STOP” finish reason with zero output tokens.

The most frustrating part is the inconsistency: with the exact same query and system prompt, it works perfectly fine sometimes, but fails completely at other times.

messages = [
    SystemMessage(content=self._get_system_prompt()),
    HumanMessage(content=query),
]

# Using bind_tools to attach tools to the LLM
llm_result = self.llm_with_tools.invoke(messages)

{
  "content": "",
  "additional_kwargs": {},
  "response_metadata": {
    "prompt_feedback": {"block_reason": 0, "safety_ratings": []},
    "finish_reason": "STOP",
    "model_name": "gemini-2.5-flash",
    "safety_ratings": []
  },
  "type": "ai",
  "name": null,
  "id": "run--[REDACTED]",
  "example": false,
  "tool_calls": [],
  "invalid_tool_calls": [],
  "usage_metadata": {
    "input_tokens": 1440,
    "output_tokens": 0,
    "total_tokens": 1440
  }
}

To be honest, I am a beginner and I have been struggling with this issue for several days. It is incredibly exhausting because I cannot find any clear reason why it works one moment and fails the next. I feel completely stuck and I am desperate to understand the cause. Any guidance or help would mean so much to me. Thank you.