Feature Request: Support multimodal documents in /v1/rerank endpoint for VL reranker models #25921

Prerequisites

• I am running the latest code. Mention the version if possible as well.
• I carefully followed the README.md.
• I searched using keywords relevant to my issue to make sure that I am creating a new issue that is not already open (or closed).
• I reviewed the Discussions, and have a new and useful enhancement to share.

Feature Description

The /v1/rerank endpoint currently only accepts documents as a flat string array. When serving a vision-language reranker model such as Qwen3-VL-Reranker-2B, there is no way to pass image content through this endpoint.

I tried two workarounds:

Passing an image URL as a plain string — the model tokenizes the URL text literally and never fetches the image.
Passing a base64 data-URI as a plain string — the model tokenizes the base64 characters as text, producing meaningless scores.
Server log when sending object-format documents:

Wrong type supplied for parameter 'documents'. Expected 'array',
using default value: [json.exception.type_error.302]
type must be string, but is object

What I expected: the ability to send structured multimodal documents, similar to what vLLM and the Alibaba DashScope API already support. For example:

{
  "model": "Qwen3-VL-Reranker-2B",
  "query": "mountain range patrol",
  "documents": [
    {"text": "A continuous mountain range landscape"},
    {"image": "https://example.com/mountain.jpg"},
    {"text": "caption text", "image": "https://example.com/photo.jpg"}
  ]
}

The llama-server already supports multimodal input in /v1/chat/completions via the mtmd/clip pipeline. The rerank endpoint could reuse the same image preprocessing path to encode image tokens before feeding them into the cross-encoder scoring logic.

Motivation

Vision-language reranker models (Qwen3-VL-Reranker, Jina-Reranker-M0, etc.) are designed to jointly rank text and image candidates against a query. They are a key component in multimodal RAG pipelines — for example, retrieving the most relevant figure or page image for a text query, or ranking candidate images by semantic relevance.

Without multimodal document support in /v1/rerank, users who deploy these GGUF models via llama-server cannot use the model's core capability. The only alternatives today are:

Switch to vLLM (heavier dependency, different deployment stack)
Write a custom Python wrapper around Transformers (loses the efficiency of llama.cpp quantized inference)
Adding this to llama-server would make it a complete serving solution for VL reranker models, consistent with its existing multimodal chat support.

Possible Implementation

A few thoughts:

Extend the /v1/rerank request schema to accept documents as either string[] (backward compatible) or an array of objects: {"text": "..."} for text-only, {"image": "url-or-data-uri"} for image-only, {"text": "...", "image": "..."} for combined.

For image documents, reuse the existing mtmd image preprocessing (resize, normalize, patch embedding) that /v1/chat/completions already uses, then prepend the resulting image tokens to the document token sequence before cross-encoder scoring.

The query field could similarly accept a string (text) or an object {"image": "..."} for image-as-query use cases.

Reference implementations: vLLM score API accepts {"text": ..., "image": ...} objects; DashScope rerank API uses {"image": "url-or-base64-data-uri"}; Qwen3-VL-Reranker model card shows the Transformers example.