add Ollama client kwargs to inactivate reasoning/thinking feature using think=False

[alchemine]

Checked

• I searched existing ideas and did not find a similar one
• I added a very descriptive title
• I've clearly described the feature request and motivation for it

Feature request

Single LLM ChatModel can handle think=False argument to inactivate reasoning features.
But arguments injection at a chain for an Ollama client model is not supported.
It's needed for utilizing full functions of an Ollama model.

Motivation

Including '/no_think' in system/human prompt can inactivate reasoning, but it still generates <think> tokens.
Also, extract_reasoning argument in ChatOllama does not avoid generating <think> tokens.
If not needed, <think> tokens should not be generated.

Proposal (If applicable)

I implemented proposal code but I'm unfamiliar with PR/english and failed in testing step.. 😢

• alchemine@cd4cc9c

---

Belows are example code:

from langchain_core.runnables import Runnable
from langchain_core.messages import AIMessage
from langchain_core.prompts import ChatPromptTemplate
from langchain_ollama import ChatOllama
from langchain_core.messages.base import get_msg_title_repr


def log(prompt: Runnable, input: dict, response: AIMessage) -> None:
    for msg in prompt.invoke(input).to_messages():
        msg.pretty_print()
    response.pretty_print()
    print(get_msg_title_repr("Usage metadata"))
    print(response.usage_metadata)


prompt = ChatPromptTemplate.from_messages(
    [
        ("system", "You are a helpful assistant."),
        ("human", "{input}"),
    ]
)
llm = ChatOllama(model="qwen3:8b", temperature=0, extract_reasoning=True)
chain = prompt | llm
input = {"input": "What is 3^3?"}

response = chain.invoke(input)
log(prompt, input, response)

================================ System Message ================================

You are a helpful assistant.
================================ Human Message =================================

What is 3^3?
================================== Ai Message ==================================

<think>
Okay, the user is asking what 3^3 is. Let me think. 3 cubed means 3 multiplied by itself three times. So, 3 times 3 is 9, and then 9 times 3 is 27. Wait, let me double-check that. Yes, 3*3=9, then 9*3=27. So the answer should be 27. I should make sure there's no confusion with exponents here. Maybe the user is just testing basic math, or they might be a student learning exponents. Either way, the calculation is straightforward. No need for complicated steps. Just confirm the multiplication. Yep, 3^3 is definitely 27.
</think>

3^3 means 3 multiplied by itself three times.  
**Calculation:**  
3 × 3 = 9  
9 × 3 = 27  

**Answer:** 27
================================ Usage metadata ================================
{'input_tokens': 26, 'output_tokens': 197, 'total_tokens': 223}

response = chain.invoke(input, client_chat_kwargs={"think": False})
log(prompt, input, response)

================================ System Message ================================

You are a helpful assistant.
================================ Human Message =================================

What is 3^3?
================================== Ai Message ==================================

3^3 means 3 multiplied by itself three times.

So,  
3^3 = 3 × 3 × 3 = 27.  

The answer is **27**.
================================ Usage metadata ================================
{'input_tokens': 34, 'output_tokens': 42, 'total_tokens': 76}

response = await chain.ainvoke(input)
log(prompt, input, response)

================================ System Message ================================

You are a helpful assistant.
================================ Human Message =================================

What is 3^3?
================================== Ai Message ==================================

<think>
Okay, the user is asking what 3^3 is. Let me think. 3 cubed means 3 multiplied by itself three times. So, 3 times 3 is 9, and then 9 times 3 is 27. Wait, let me double-check that. Yes, 3*3=9, then 9*3=27. So the answer should be 27. I should make sure there's no confusion with exponents here. Maybe the user is just testing basic math, or they might be a student learning exponents. Either way, the calculation is straightforward. No need for complicated steps. Just confirm the multiplication. Yep, 3^3 is definitely 27.
</think>

3^3 means 3 multiplied by itself three times.  
**Calculation:**  
3 × 3 = 9  
9 × 3 = 27  

**Answer:** 27
================================ Usage metadata ================================
{'input_tokens': 26, 'output_tokens': 197, 'total_tokens': 223}

response = await chain.ainvoke(input, client_chat_kwargs={"think": False})
log(prompt, input, response)

================================ System Message ================================

You are a helpful assistant.
================================ Human Message =================================

What is 3^3?
================================== Ai Message ==================================

3^3 means 3 multiplied by itself three times.

So,  
3^3 = 3 × 3 × 3 = 27.  

The answer is **27**.
================================ Usage metadata ================================
{'input_tokens': 34, 'output_tokens': 42, 'total_tokens': 76}

response = chain.invoke(input, client_chat_kwargs={"thinking": True})
log(prompt, input, response)

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
Cell In[6], line 1
----> 1 response = chain.invoke(input, client_chat_kwargs={"thinking": True})
      2 log(prompt, input, response)

File /opt/miniconda3/envs/ollama-serve/lib/python3.11/site-packages/langchain_core/runnables/base.py:3048, in RunnableSequence.invoke(self, input, config, **kwargs)
   3046     input_ = context.run(step.invoke, input_, config, **kwargs)
   3047 elif isinstance(step, BaseChatModel):
-> 3048     input_ = context.run(step.invoke, input_, config, **kwargs.get("client_chat_kwargs", {}))
   3049 else:
   3050     input_ = context.run(step.invoke, input_, config)

File /opt/miniconda3/envs/ollama-serve/lib/python3.11/site-packages/langchain_core/language_models/chat_models.py:372, in BaseChatModel.invoke(self, input, config, stop, **kwargs)
    360 @override
    361 def invoke(
    362     self,
   (...)    367     **kwargs: Any,
    368 ) -> BaseMessage:
    369     config = ensure_config(config)
    370     return cast(
    371         "ChatGeneration",
--> 372         self.generate_prompt(
    373             [self._convert_input(input)],
    374             stop=stop,
    375             callbacks=config.get("callbacks"),
    376             tags=config.get("tags"),
    377             metadata=config.get("metadata"),
    378             run_name=config.get("run_name"),
    379             run_id=config.pop("run_id", None),
    380             **kwargs,
    381         ).generations[0][0],
    382     ).message

File /opt/miniconda3/envs/ollama-serve/lib/python3.11/site-packages/langchain_core/language_models/chat_models.py:957, in BaseChatModel.generate_prompt(self, prompts, stop, callbacks, **kwargs)
    948 @override
    949 def generate_prompt(
    950     self,
   (...)    954     **kwargs: Any,
    955 ) -> LLMResult:
    956     prompt_messages = [p.to_messages() for p in prompts]
--> 957     return self.generate(prompt_messages, stop=stop, callbacks=callbacks, **kwargs)

File /opt/miniconda3/envs/ollama-serve/lib/python3.11/site-packages/langchain_core/language_models/chat_models.py:776, in BaseChatModel.generate(self, messages, stop, callbacks, tags, metadata, run_name, run_id, **kwargs)
    773 for i, m in enumerate(input_messages):
    774     try:
    775         results.append(
--> 776             self._generate_with_cache(
    777                 m,
    778                 stop=stop,
    779                 run_manager=run_managers[i] if run_managers else None,
    780                 **kwargs,
    781             )
    782         )
    783     except BaseException as e:
    784         if run_managers:

File /opt/miniconda3/envs/ollama-serve/lib/python3.11/site-packages/langchain_core/language_models/chat_models.py:1022, in BaseChatModel._generate_with_cache(self, messages, stop, run_manager, **kwargs)
   1020     result = generate_from_stream(iter(chunks))
   1021 elif inspect.signature(self._generate).parameters.get("run_manager"):
-> 1022     result = self._generate(
   1023         messages, stop=stop, run_manager=run_manager, **kwargs
   1024     )
   1025 else:
   1026     result = self._generate(messages, stop=stop, **kwargs)

File /opt/miniconda3/envs/ollama-serve/lib/python3.11/site-packages/langchain_ollama/chat_models.py:741, in ChatOllama._generate(self, messages, stop, run_manager, **kwargs)
    734 def _generate(
    735     self,
    736     messages: list[BaseMessage],
   (...)    739     **kwargs: Any,
    740 ) -> ChatResult:
--> 741     final_chunk = self._chat_stream_with_aggregation(
    742         messages, stop, run_manager, verbose=self.verbose, **kwargs
    743     )
    744     generation_info = final_chunk.generation_info
    745     chat_generation = ChatGeneration(
    746         message=AIMessage(
    747             content=final_chunk.text,
   (...)    752         generation_info=generation_info,
    753     )

File /opt/miniconda3/envs/ollama-serve/lib/python3.11/site-packages/langchain_ollama/chat_models.py:678, in ChatOllama._chat_stream_with_aggregation(self, messages, stop, run_manager, verbose, **kwargs)
    669 def _chat_stream_with_aggregation(
    670     self,
    671     messages: list[BaseMessage],
   (...)    675     **kwargs: Any,
    676 ) -> ChatGenerationChunk:
    677     final_chunk = None
--> 678     for chunk in self._iterate_over_stream(messages, stop, **kwargs):
    679         if final_chunk is None:
    680             final_chunk = chunk

File /opt/miniconda3/envs/ollama-serve/lib/python3.11/site-packages/langchain_ollama/chat_models.py:763, in ChatOllama._iterate_over_stream(self, messages, stop, **kwargs)
    756 def _iterate_over_stream(
    757     self,
    758     messages: list[BaseMessage],
    759     stop: Optional[list[str]] = None,
    760     **kwargs: Any,
    761 ) -> Iterator[ChatGenerationChunk]:
    762     is_thinking = False
--> 763     for stream_resp in self._create_chat_stream(messages, stop, **kwargs):
    764         if not isinstance(stream_resp, str):
    765             if stream_resp.get("done") is True:

File /opt/miniconda3/envs/ollama-serve/lib/python3.11/site-packages/langchain_ollama/chat_models.py:665, in ChatOllama._create_chat_stream(self, messages, stop, **kwargs)
    662 chat_params = self._chat_params(messages, stop, **kwargs)
    664 if chat_params["stream"]:
--> 665     yield from self._client.chat(**chat_params)
    666 else:
    667     yield self._client.chat(**chat_params)

TypeError: Client.chat() got an unexpected keyword argument 'thinking'

[alchemine]

This feature is similar to belows, but are different in that this suggests controlling arguments in "invoke-time".

• Add Documentation for Qwen3 Extended Thinking Support in ChatOllama #31633
• Adding a parameter to OllamaLLM to disable the thinking output. #31573