server : support typed content (vision/audio/video) input for /v1/embeddings endpoint

[timothywang21]

Overview

Right now, multimodal support for embeddings is possible but requires janky workarounds like #18665 (comment)

This PR consolidates these workarounds by implementing native multimodal support as well as introducing an OpenAI-style API for the /v1/embeddings endpoint. This new OAI-style API is the defacto standard that is used by Openrouter and other providers for multimodal embeddings. The previous legacy API is still supported.

Legacy API (still supported):

{
  "input": {
    "prompt_string": "Describe this image<__media__>",
    "multimodal_data": ["iVBORw0KGgoAAAANSUhEUg..."]
  }
}

OAI-compatible API (new), multiple content arrays in one request:

{
  "input": [
    {
      "content": [
        { "type": "image_url", "image_url": { "url": "data:image/jpeg;base64,/9j/4AAQSkZJRg..." } },
        { "type": "text", "text": "Describe this image" }
      ]
    },
    {
      "content": [
        { "type": "text", "text": "This is a second content array you can pass in one request" }
      ]
    }
  ]
}

Incorrect (bare content array - rejected:

{
  "input": [
    { "type": "image_url", "image_url": { "url": "data:image/jpeg;base64,/9j/4AAQSkZJRg..." } },
    { "type": "text", "text": "Describe this image" }
  ]
}

Code Changes

1. Implemented tokenize_oai_content_array, which accepts the OpenAI-style wrapped content array format for multimodal embedding requests. Each {"content": [...]} object is one input that produces one embedding; text parts are concatenated and image_url parts are decoded via handle_media then spliced with process_mtmd_prompt.

• NOTE: The legacy formats (plain string, token arrays, mixed arrays, and the {prompt_string, multimodal_data} object) continue to work unchanged via tokenize_input_prompts. Bare content arrays (the unwrapped shape) are rejected with a migration message.

2. Disabled KV prefix reuse for stateless embedding/rerank tasks so that repeated inputs do not incorrectly share cached KV across requests. Each new request is done from scratch without reusing previous context from other requests. This was done for logical correctness as well as resolving a really pesky bug that I believe was caused by stale pointers from the multimodal inputs when it gets reused by another embedding/rerank request.

3. Updated /v1/embeddings README.

Additional information

This is the second PR in support of adding full multimodal support for embedding and reranker endpoints so that models like the Qwen3-VL embedding and rerank can be fully utilized. See #28876

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: Assisted-by: Opencode Qwen3.8 27B but all code was verified by me

[ngxson]

thanks. this is something I planned from #25093 (comment)

I will take over this PR for now, please don't push to it

[timothywang21]

Got it. I have the multimodal rerank PR ready to go when this is finished.