[FEATURE] Support enable_thinking parameter for Qwen models (qwen3-4b, qwen3-8b, etc.)

[jackiemoo]

Problem Description

When using Qwen models (specifically qwen3-4b, qwen3-8b, qwen3-14b, qwen3-32b) through ContextGem, we encounter the following error:

LLMAPIError( contextgem.internal.exceptions.LLMAPIError: Exception occurred while calling LLM API (after 3 retries) - Original error: litellm.BadRequestError: OpenAIException - parameter.enable_thinking must be set to false for non-streaming calls LiteLLM Retried: 3 times

This error occurs because:

1. Qwen3 series models support thinking mode (reasoning capabilities)
2. In non-streaming calls, these models require enable_thinking=false to be explicitly set
3. ContextGem currently doesn't handle this parameter automatically

Expected Behavior

ContextGem should automatically handle the enable_thinking parameter for Qwen models, similar to how it handles other model-specific parameters.

Current Workaround

Currently, we need to use Monkey Patching to modify the _build_request_config method:

def apply_qwen_thinking_patch():
    from contextgem import DocumentLLM
    
    original_build_request_config = DocumentLLM._build_request_config
    
    def patched_build_request_config(self):
        request_config = original_build_request_config(self)
        
        if not self._supports_reasoning:
            request_config["enable_thinking"] = False
        
        return request_config
    
    DocumentLLM._build_request_config = patched_build_request_config

Proposed Solution

Add support for enable_thinking parameter in the _build_request_config method:

# In _build_request_config method, before return request_config:
if not self._supports_reasoning:
    # For models that support thinking but are used in non-reasoning mode
    # (like Qwen3 series), disable thinking mode
    request_config["enable_thinking"] = False

Environment

• ContextGem version: [latest]
• Python version: 3.11+
• Model: dashscope/qwen3-4b, dashscope/qwen3-8b, etc.
• API: DashScope (Alibaba Cloud)

Additional Context

This issue affects users who want to use Qwen3 series models with ContextGem. The models work fine when enable_thinking=false is set, but ContextGem doesn't handle this automatically, requiring manual workarounds.

Related

• This is similar to how ContextGem handles other model-specific parameters
• The fix should be backward compatible and not affect other models
• The solution should work for all Qwen3 series models (4b, 8b, 14b, 32b)

Priority

This is a blocking issue for users wanting to use Qwen3 models with ContextGem. A simple fix would greatly improve the user experience.

[SergiiShcherbak]

@jackiemoo Thanks for reporting this issue.

I added tests for Qwen3 models in #80 , and such tests pass completely without any qwen3-related changes or enable_thinking parameter handling:

• Tested "ollama/qwen3:14b" and "ollama_chat/qwen3:14b"
• Tested StringConcept extraction using llm.extract_concepts_from_document(...) method, with reasoning explicitly enabled and disabled (by setting llm._supports_reasoning param to True/False respectively)
• Tested llm.chat(...) with and without chat session object.

Since there are no issues with using Qwen3 models via ollama, it looke like the issue you are experiencing is provider-specific, i.e. related to Dashscope API via LiteLLM.

Could you try using LiteLLM's completion(...) directly with "dashscope/qwen3-..." with streaming disabled and see if you encounter the same error? If you do, then LiteLLM does not handle the enable_thinking parameter properly for Dashscope API. In this case, this issue might be better handled at the LiteLLM level since it's provider-specific, i.e. opening an issue with LiteLLM is advisable.

I don't have access to Dashcrope API, therefore I can't test "dashscope/qwen3-..." models live.