Added support for the "think" for Ollama

[sunyuhan1998]

Fixes #3383
As mentioned in issue: #3383 , Ollama added support for "think" in its latest 0.9.0 version:
https://github.com/ollama/ollama/releases
https://github.com/ollama/ollama/blob/main/docs/api.md#generate-a-chat-completion.

This PR implements support for that attribute and includes the following key changes:

1. Added the think field to Ollama's ChatRequest
2. Added the thinking field to Ollama's Message
3. Added the think property to OllamaOptions, allowing users to specify whether to enable or disable thinking

Actually, there is currently another issue:

As stated in Ollama's API documentation here, during requests to Ollama, the message field supports sending the model's own reasoning (thoughts) back to it. However, AssistantMessage does not currently support transmitting this field, which means the model will not be aware of its previous thoughts.

Therefore, perhaps we need to add a specialized Message implementation for Ollama, such as OllamaAssistantMessage. I'm not sure whether this would be considered a significant change.

[sunyuhan1998]

@tzolov @ilayaperumalg @markpollack Could you please help review this PR? Thank you.

[markpollack]

Yes, we will review. Thanks

[markpollack]

It seems like 'think' is not a part of the options map in ollama, but a 'top level' field in the request object.
In https://ollama.com/blog/thinking there is the example

curl http://localhost:11434/api/chat -d '{
  "model": "deepseek-r1",
  "messages": [
    {
      "role": "user",
      "content": "how many r in the word strawberry?"
    }
  ],
  "think": true,
  "stream": false
}'

and the golang type supporting this feature also shows the same structure.

https://github.com/ollama/ollama/blob/45f56355d557b7130c7c07bbd6e1b634a758d946/api/types.go#L91

So it shouldn't be added to the options map.

[markpollack]

The comment in the code

and Ollama will return
	 * the thought process within the `content` field of the response, wrapped in `<thinking>` tags.

seems to contradict what was documented on the ollama web site that shows the 'think' response as a separate field from 'content', and not nested inside the 'content' field.

[markpollack]

Maybe we want to also expose it in OllamaOptions as that right now would be the only way to pass in this feature flag when making calls via ChatModel or ChatClient. I think if the feature of enabling thinking mode is implemented from a ChatModel or ChatClient level the right solution will present itself. Can you improve this PR to handle this scenario please?

[sunyuhan1998]

Maybe we want to also expose it in OllamaOptions as that right now would be the only way to pass in this feature flag when making calls via ChatModel or ChatClient. I think if the feature of enabling thinking mode is implemented from a ChatModel or ChatClient level the right solution will present itself. Can you improve this PR to handle this scenario please?

Of course, no problem. I will continue improving this PR and update my progress here in a timely manner.

Before proceeding, I want to confirm whether I’ve correctly understood your point: Are you suggesting that we should not add the think flag in OllamaOptions, but instead make adjustments at the OllamaChatModel level? Or do you mean that we should implement support for think at the ChatModel or ChatClient level? If it's the latter, I think we would also need to adjust the implementations of different ChatModels to support this option (of course, depending on whether the underlying model actually supports it).

[sunyuhan1998]

seems to contradict what was documented on the ollama web site that shows the 'think' response as a separate field from 'content', and not nested inside the 'content' field.

Firstly, the content in the comment is actually a summary I derived through practical testing.

I think this is a form of backward compatibility Ollama implemented for users still employing the old parameter-passing method. Before Ollama supported the "think" flag, if we made a request to a model that supports thinking like this:

curl http://localhost:11434/api/chat -d '{
  "model": "qwen3:4b",
  "messages": [
    { "role": "user", "content": "why is the sky blue?" }
  ]
}'

Ollama would enable thinking by default and return the thought process wrapped in <thinking> tags within the content field of the response. The response would look something like this:

{"model":"qwen3:4b","created_at":"2025-06-12T12:17:08.385341Z","message":{"role":"assistant","content":"\u003cthink\u003e"},"done":false}
{"model":"qwen3:4b","created_at":"2025-06-12T12:17:08.402756Z","message":{"role":"assistant","content":"\n"},"done":false}
{"model":"qwen3:4b","created_at":"2025-06-12T12:17:08.420851Z","message":{"role":"assistant","content":"Okay"},"done":false}
{"model":"qwen3:4b","created_at":"2025-06-12T12:17:08.439825Z","message":{"role":"assistant","content":","},"done":false}
{"model":"qwen3:4b","created_at":"2025-06-12T12:17:08.457618Z","message":{"role":"assistant","content":" the"},"done":false}
{"model":"qwen3:4b","created_at":"2025-06-12T12:17:08.474711Z","message":{"role":"assistant","content":" user"},"done":false}
{"model":"qwen3:4b","created_at":"2025-06-12T12:17:08.491833Z","message":{"role":"assistant","content":" is"},"done":false}
{"model":"qwen3:4b","created_at":"2025-06-12T12:17:08.509124Z","message":{"role":"assistant","content":" asking"},"done":false}

This remains the case in the latest version of Ollama: if the think flag is not specified, the behavior of Ollama remains unchanged from before. This is what I intended to express in the comment:

If this value is not specified, it defaults to null, and Ollama will return the thought process within the content field of the response, wrapped in <thinking> tags.

Only when we specify the think flag will Ollama return the thought process in the thinking field of the response:

curl http://localhost:11434/api/chat -d '{
  "model": "qwen3:4b",
  "messages": [
    { "role": "user", "content": "why is the sky blue?" }
  ],
  "think": true
}'

Response:

{"model":"qwen3:4b","created_at":"2025-06-12T12:22:48.135211Z","message":{"role":"assistant","content":"","thinking":"Okay"},"done":false}
{"model":"qwen3:4b","created_at":"2025-06-12T12:22:48.152511Z","message":{"role":"assistant","content":"","thinking":","},"done":false}
{"model":"qwen3:4b","created_at":"2025-06-12T12:22:48.169911Z","message":{"role":"assistant","content":"","thinking":" the"},"done":false}
{"model":"qwen3:4b","created_at":"2025-06-12T12:22:48.187023Z","message":{"role":"assistant","content":"","thinking":" user"},"done":false}
{"model":"qwen3:4b","created_at":"2025-06-12T12:22:48.204039Z","message":{"role":"assistant","content":"","thinking":" is"},"done":false}
{"model":"qwen3:4b","created_at":"2025-06-12T12:22:48.221233Z","message":{"role":"assistant","content":"","thinking":" asking"},"done":false}

I think it's the same for Spring AI. We should maintain compatibility with users who are using older versions, meaning that if the "think" flag is not specified, the returned format should remain unchanged.

[markpollack]

We should propagate the 'thinking' response back into the ChatGenerationMetadata so that it is accessible in the response when using OllamaChatModel. This would be inside the internalCall method of OllamaChatModel and add add a test for that in OllamaChatModelIT

[sunyuhan1998]

No problem, I will add the implementation for this part.