[Features]Add Default Reranking to `searches` API

[earayu]

Currently, the searches API lacks a default reranking mechanism. To improve search result relevance, we should add default reranking.

Proposed Changes:

• Implement default reranking for all searches API calls.
• Introduce a parameter (e.g., enable_rerank) to control whether reranking is applied.
  • This parameter should default to True.
  • When True, the system's default reranking model will be used.

[Shushankranjan]

Hi @earayu, I’d like to work on this. Please assign it to me.

[earayu]

Thank you for your enthusiasm and quick follow-up contribution! Yes, you can definitely work on this issue.

Issue Overview

We need to implement a rerank step in the collection search functionality. Currently, our flow system has rerank capabilities, but the direct search interface (/collections/{collection_id}/searches) lacks this feature.

Implementation Details

Entry Point

@router.post("/collections/{collection_id}/searches", tags=["search"])
@audit(resource_type="search", api_name="CreateSearch")
async def create_search_view(
    request: Request,
    collection_id: str,
    data: view_models.SearchRequest,
    user: User = Depends(current_user),
) -> view_models.SearchResult:
    return await collection_service.create_search(str(user.id), collection_id, data)

Requirements

1. Add rerank step after all index operations (vector, fulltext, graph) are completed
2. Default rerank model selection: Use models with "default_for_rerank" tag. If multiple exist, select the first one
3. Fallback strategy when no rerank model is available:
   • Place graph search results first (they typically have better quality and return only 1 result)
   • Sort remaining vector and fulltext results by score in descending order

Implementation Steps

1. API Schema Updates:

   • Modify API definitions in aperag/api/ directory
   • Run make generate-models to regenerate backend models
   • Run make generate-frontend-sdk to update frontend SDK

2. Service Layer Implementation:

   • Enhance collection_service.create_search() method
   • Integrate rerank functionality from existing flow components
   • Implement fallback logic for cases without rerank models

3. MCP Interface Updates:

   • Apply corresponding changes to MCP interfaces to maintain consistency

This enhancement will significantly improve search result quality by leveraging our existing rerank infrastructure in the direct search API.

Let me know if you need any clarification on the implementation approach!