Add support for task-specific prompt templates for embedding models

[ww2283]

What problem does this solve?

Please check the following description and let me know if this is feasible, then I can help to implement this feature to support template prepending embedding models.

Context

Google's new https://huggingface.co/blog/embeddinggemma model introduces task-specific prompts that significantly improve embedding quality. The model requires different prompts prepended to text based on the use case:

• Queries: "task: search result | query: "
• Documents: "title: none | text: "
• Other tasks: Classification, clustering, reranking each have specific prompts

Problem

When using EmbeddingGemma via LM Studio's OpenAI-compatible endpoint (or similar servers), LEANN currently sends raw text without task-specific prompts. This prevents leveraging EmbeddingGemma's full capabilities.

Proposed Solution

Add a prompt_template parameter to provider_options in the OpenAI embedding mode:

Query time (search)

provider_options={ "prompt_template": "task: search result | query: " }

Build time (indexing documents)

provider_options={ "prompt_template": "title: none | text: " }

Implementation would prepend the template to all texts before sending to the API: prompt_template = provider_options.get("prompt_template") if prompt_template: texts = [f"{prompt_template}{text}" for text in texts]

Why LEANN Should Handle This

• Context awareness: LEANN knows whether text is a query or document
• Server agnostic: Works with any OpenAI-compatible endpoint (LM Studio, vLLM, TEI)
• Future-proof: Other embedding models may adopt similar prompt-based approaches

Benefits

• Unlock EmbeddingGemma's performance gains
• Support emerging prompt-based embedding models
• Minimal implementation complexity

Proposed solution

NA

Example usage

[yichuan-w]

Wow, that is amazing!