Tool call not actually running my tool function in Flask WhatsApp chatbot (function never fires)

[ADVISE-AI]

I’m building an AI-powered WhatsApp chatbot using flask, langchain, langgraph, postgres database(not vector store), and whatsapp's cloud api.

The workflow I’m aiming for:

1. User asks for a sample invite card/video.
2. LangChain routes request → tool call → tool queries Postgres for media_id.
3. Tool sends the media to the user via WhatsApp Cloud API.

The tool is defined with LangChain’s decorator like this:

from langchain_core.tools import tools

@tool
def RespondWithMedia(media_type: str, media_description, caption: str = "") -> dict:
    """
    Send the user WhatsApp videos.
    Args:
       
 media_type: 'wedding', 'anniversary', or 'birthday'.
        media_description: 
            - if media_type = 'wedding' → '2d sample' OR '3d sample with caricature'
            - if anniversary/birthday → '2d with caricature'
        caption: optional caption for the video
    """
    print("RespondWithMedia Tool Running!") # <-- This line is for debugging which never fires suggesting that tool is never executed

    id_list = []
    with conn.cursor() as curr:
        sql_query = """SELECT media_id FROM sample_media_library 
                       WHERE media_type = %s AND media_description = %s"""
        curr.execute(sql_query, (media_type.lower(), media_description.lower()))
        result = curr.fetchall()

    for row in result:
        id_list.append(row[0])

    user_ph = RunnableConfig.get_current().get("configurable", {}).get("thread_id")
    print("Sending to Phone: ", user_ph)
    url = f"https://graph.facebook.com/v23.0/{WHATSAPP_PHONE_NUMBER_ID}/messages"
    headers = {
        "Authorization": f"Bearer {WHATSAPP_ACCESS_TOKEN}",
        "Content-Type": "application/json",
    }

    responses = []
    for media_id in id_list:
        data = {
            "messaging_product": "whatsapp",
            "recipient_type": "individual",
            "to": f"{user_ph}",
            "type": "video",
            "video": {"id": media_id, "caption": caption},
        }
        try:
            resp = requests.post(url, headers=headers, json=data)
            responses.append({"media_id": media_id, "status": resp.status_code})
        except requests.RequestException as e:
            responses.append({"media_id": media_id, "status": "failed", "error": str(e)})

    return {"results": responses}

prompt_template = ChatPromptTemplate.from_messages([
    ("system", "{system_message}"), 
    MessagesPlaceholder("messages")
])

tool_node = ToolNode([RespondWithMedia])

llm = init_chat_model("openai:gpt-4.1")
llm_model = llm.bind_tools([RespondWithMedia])
agent = prompt_template | llm_model

with open("sys_prompt.txt") as f:
    SYSTEM_PROMPT = f.read()

def should_continue(state: State):
    messages = state["messages"]
    last = messages[-1]
    if hasattr(last, "tool_calls") and last.tool_calls:
        return "tool"
    return END

def chatbot(state: State):
    ai_resp = agent.invoke({
        "system_message": SYSTEM_PROMPT, 
        "messages": state["messages"]
    })
    print(ai_resp)

    return {"messages": [ai_resp]}

graph_builder = StateGraph(State)
graph_builder.add_node("chatbot", chatbot)
graph_builder.add_node("tool", tool_node)

graph_builder.add_edge(START, "chatbot")
graph_builder.add_conditional_edges("chatbot", should_continue, ["tool", END])
graph_builder.add_edge("tool", "chatbot")

graph = graph_builder.compile(checkpointer = checkpointer)

def stream_graph_updates(user_ph: int, user_input: str) -> dict:
    for event in graph.stream({"messages":[{"role":"user", "content": user_input}]}, config = {"configurable": {"thread_id": user_ph}}):
        for value in event.values():
            print(value['messages'][-1])        
            ai_resp = value['messages'][-1].content
            metadata = value['messages'][-1].usage_metadata
            
            resp = {
                "content": ai_resp,
                "metadata": metadata
            }
            return resp

Values like user_ph and user_input are passed into the stream_graph_updates() from the payload received on the flask webhook.

When I ask the bot on whatsapp to send me sample files for wedding, It does nothing, but on the backend, the logs suggest that the bot does generate correct arguments for the tool,

  "content": "",
  "additional_kwargs": {
    "tool_calls": [
      {
        "id": "call_vBwifQlbYwRLxjth8aRiK0iP",
        "function": {
          "arguments": "{\"media_type\":\"wedding\",\"media_description\":\"2d sample\"}",
          "name": "RespondWithMedia"
        },
        "type": "function"
      }
    ],
    "refusal": null
  },
  "response_metadata": {
    "token_usage": {
      "completion_tokens": 25,
      "prompt_tokens": 1669,
      "total_tokens": 1694,
      "completion_tokens_details": {
        "accepted_prediction_tokens": 0,
        "audio_tokens": 0,
        "reasoning_tokens": 0,
        "rejected_prediction_tokens": 0
      },
      "prompt_tokens_details": {
        "audio_tokens": 0,
        "cached_tokens": 1536
      }
    },
    "model_name": "gpt-4.1-2025-04-14",
    "system_fingerprint": "fp_daf5fcc80a",
    "id": "chatcmpl-C6iXgfwC8S8HOL0Iw8aEAqSZKP4gq",
    "service_tier": "default",
    "finish_reason": "tool_calls",
    "logprobs": null
  },
  "id": "run--72263376-0ea2-4d4c-a6f5-5a02d2be8b46-0",
  "tool_calls": [
    {
      "name": "RespondWithMedia",
      "args": {
        "media_type": "wedding",
        "media_description": "2d sample"
      },
      "id": "call_vBwifQlbYwRLxjth8aRiK0iP",
      "type": "tool_call"
    }
  ],
  "usage_metadata": {
    "input_tokens": 1669,
    "output_tokens": 25,
    "total_tokens": 1694,
    "input_token_details": {
      "audio": 0,
      "cache_read": 1536
    },
    "output_token_details": {
      "audio": 0,
      "reasoning": 0
    }
  }
)

The arguments perfectly match with the values in the database so I don't suspect that it fails.