Hello NetXMS Development Team,
 
I have a question about the AI tool invocation format of NetXMS.
 
I heard that NetXMS currently only accepts AI tool calls developed based on standard JSON format, and it cannot work properly with models that adopt XML format for tool calling.
 
The Qwen3.6 series large models use XML format to implement tool invocation. I would like to confirm whether NetXMS is indeed incompatible with XML tool calling mode right now.
 
Besides, I sincerely hope that you can add support for XML format AI tool calls in subsequent version updates. This optimization will bring great convenience for users applying Qwen3.6 and other XML-based AI models.
 
Thank you very much for your time and support!

Are you running Qwen locally? Under what tool?

The "XML tool call" output is what Qwen models emit at the raw model layer. Almost all production inference servers convert that into standard OpenAI JSON tool_calls before returning the response — so Qwen normally works fine with NetXMS if the inference server is configured to parse tool calls. For example:

- Ollama — works out of the box for Qwen models that ship with a tool-call template in their Modelfile. Point NetXMS at the Ollama provider and it will receive normal JSON tool_calls.
- vLLM — requires starting the server with --enable-auto-tool-choice --tool-call-parser hermes (or qwen3_coder for the newer parser). Then the OpenAI-compatible endpoint returns JSON tool_calls.
- SGLang / llama.cpp server — similar flags exist.

The model is deployed using vLLM with the configuration --tool-call-parser=hermes. Could you please provide guidance?

Below is a current example:
[root@ai-app-0l apps]# curl http://IP:8001/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer <your_token>" \
-d '{
  "model": "qwen3.6-35b-a3b",
  "messages": [
    {"role": "user", "content": "What's the weather like in Beijing today?"}
  ],
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "get_weather",
        "description": "Get the current weather for the specified city",
        "parameters": {
          "type": "object",
          "properties": {
            "location": {
              "type": "string",
              "description": "City name, e.g., Beijing, Shanghai"
            }
          },
          "required": ["location"]
        }
      }
    }
  ]
}'

Response:

{
  "id": "chatcmpl-a914ca7d228bc903",
  "object": "chat.completion",
  "created": 1779196243,
  "model": "qwen3.6-35b-a3b",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "\n\n<tool_call>\n<function=get_weather>\n<parameter=location>\nBeijing\n</parameter>\n</function>\n</tool_call>",
        "refusal": null,
        "annotations": null,
        "audio": null,
        "function_call": null,
        "tool_calls": [],
        "reasoning": "Here's a thinking process:\n\n1. **Analyze User Input:**\n - User asks: \"What's the weather like in Beijing today?\"\n - Key entity: Beijing\n - Request: Current weather\n\n2. **Identify Available Tools:**\n - I have one tool: 'get_weather'\n - Tool description: Get the current weather for a specified city\n - Parameters: 'location' (string, required) - City name, e.g., Beijing, Shanghai\n\n3. **Map User Input to Tool Parameters:**\n - 'location' should be \"Beijing\"\n\n4. **Execute Tool Call:**\n - Call 'get_weather' with 'location: \"Beijing\"'\n\n5. **Process Tool Response (Simulation/Expectation):**\n - Since I don't have real-time access, I'll simulate the tool call and assume it returns a reasonable weather response. However, in this environment, I will actually make the tool call and let the system handle it. I'll generate the tool call now.\n\n Tool Call:\n```json\n{\n  \"name\": \"get_weather\",\n  \"parameters\": {\n    \"location\": \"Beijing\"\n  }\n}\n```\n\n6. **Formulate Response:**\n After receiving the tool response, I will present the weather information clearly to the user in Chinese.\n - For now, I'll just make the tool call."
      },
      "logprobs": null,
      "finish_reason": "stop",
      "stop_reason": null,
      "token_ids": null
    }
  ],
  "service_tier": null,
  "system_fingerprint": null,
  "usage": {
    "prompt_tokens": 284,
    "total_tokens": 656,
    "completion_tokens": 372,
    "prompt_tokens_details": null
  },
  "prompt_logprobs": null,
  "prompt_token_ids": null,
  "ky_transfer_params": null
}

Make sure you have --enable-auto-tool-choice set as well. If won't help, try --tool-call-parser=pythonic