Improve MistralAiApi Jackson mapping for message content

[nicolaskrier]

Description

• Define different type of chunk instead of media content,
• Verify that new chunks allow reasoning models support,
• Introduce reasoning_effort request parameter for reasoning models,
• Introduce reference_content, reference_thinking_content and thinking_content metadata in MistralAiChatModel,
• Remove unnecessary null checks in MistralAiChatModel,
• Improve existing Javadoc,
• Polish integration tests and unit tests.

Explanations

This PR aligns the Jackson mapping with the content types defined in the Mistral AI OpenAPI Specifications. The changes ensure null safety matches the OpenAPI Specifications requirements and introduce support for new content types used by Magistral 1.2 models. New reasoning on demand feature for Mistral Small 4 and Mistral Medium 3.5 models is also added in order to cover all Mistral reasoning models.

Proposed milestone

2.0.0-M6

Related work

• Make spring-ai-mistral-ai null-safe #5354,
• Improve Mistal AI API Javadoc #5364,
• Fix MistralAiApi to handle thinking content blocks #5020,
• Provide a common ThinkingAssistantMessage class reusable for each AI vendor #5698.

Remaining tasks

• MistralAiApi.ChatCompletionMessage unit tests demonstrating content attribute serialization and deserialization,
• MistralAiApi.ChatCompletionMessage.ContentSerializer unit tests,
• MistralAiApi.ChatCompletionMessage.ContentDeserializer unit tests,
• MistralAiApi.ChatCompletionMessage.extractContent unit tests,
• MistralAiApi.ChatCompletionMessage.extractThinkingContent unit tests,
• MistralAiApi reasoning_effort parameter support for Mistral Small 4 and Mistral Medium 3.5,
• MistralAiChatOptions reasoning_effort parameter support,
• MistralAiApi integration tests demonstrating Mistral Small 4 and Mistral Medium 3.5 reasoning on demand feature,
• MistralAiApi integration test including MistralAiApi.ChatCompletionMessage.ImageUrlChunk,
• MistralAiApi integration test including MistralAiApi.ChatCompletionMessage.ReferenceChunk.

[nicolaskrier]

Hi Sébastien,

I wanted to give you a quick heads-up here: I am currently reworking the solution. I have implemented four different methods at the MistralAiApi.ChatCompletionMessage level:

• extractTextContent: currently used by MistralAiChatModel to fill the AssistantMessage content,
• extractReferenceContent: not yet used but intended to be used by MistralAiChatModel to fill a dedicated reference entry in the AssistantMessage properties,
• extractThinkingTextContent: not yet used but intended to be used by MistralAiChatModel to fill a dedicated thinking entry in the properties,
• extractThinkingReferenceContent: not yet used but intended to be used by MistralAiChatModel to fill a dedicated thinking reference entry in the properties.

I plan to finish this by the end of the long weekend. Feel free to ping me if you think this approach is not suitable.

[sdeleuze]

@nicolaskrier Is this PR ready for a new review?

[nicolaskrier]

@sdeleuze: I am currently working on a last integration test verifying that reference content metadata is properly filled and then it will be over.

[nicolaskrier]

@sdeleuze: I have slightly modified the proposed solution to use a list of integers for the reference_content and reference_thinking_content metadata.

It is ready to be reviewed now.

[nicolaskrier]

@sdeleuze: I have resolved the conflicts with the recent changes introduced in #6135 and also slightly modified MistralAiStreamFunctionCallingHelper to use the raw content of the message instead of calling the extractTextContent method.

[sdeleuze]

Merged with additional documentation and configuration property, let's polish in RC1 if needed.

@nicolaskrier Does #5020 still make sense of should we close it?

[nicolaskrier]

@sdeleuze: I think that you can close the other PR and the associated issue.