[Enhancement]: Support custom reranker provider with configurable URL and model name #12119

Activity

1. JumpLink

   added

   ✨ enhancementNew feature or requestNew feature or request

   on Mar 7, 2026
2. danny-avila commented on Mar 7, 2026

   

   danny-avila

   on Mar 7, 2026

   Collaborator

   More actions

   Thanks for the request, on my list to finalize this PR: LibreChat-AI/rag-api#190

   👍React with 👍5Reacted by Pascal Garber, Jonas Flügge, Ole, Lawrence and foegra
3. mdbook commented on Jun 19, 2026

   

   mdbook

   on Jun 19, 2026

   More actions

   +1, ran into this exact thing self-hosting against a local Infinity reranker.

   two issues with the jina path right now:

   1. the model name is hardcoded to jina-reranker-v2-base-multilingual in @librechat/agents (tools/search/rerankers.ts), no way to override it
   2. there's no custom reranker type, just jina and cohere

   jinaApiUrl already lets you point somewhere else and that part works great (Infinity serves a jina-style /rerank). the problem is the hardcoded model name, so my reranker has to answer to jina-reranker-v2-base-multilingual. my current hack is aliasing the real model under that name with Infinity's --served-model-name. it works, but it's gross: the served name is now decoupled from the actual model, and anything else hitting that endpoint (a memory service, in my case) has to send the fake name too.

   rerankerType: custom + customRerankerModel would just fix it. would love to see it land. this is a UI built around local LLMs and reranking is basically the last thing still pinned to a cloud provider's model name.
4. JumpLink

   added a commit that references this issue on Aug 14, 2026

   feat: add custom reranker provider support in web search config

   ...

   This issue will close once commit 80374dd is merged into the 'main' branch.

   80374dd
5. codegraph-librechat

   added

   🗺️ Backend Infracodegraph: the taxonomy area this belongs to (classifier, confidence ≥ 0.9)codegraph: the taxonomy area this belongs to (classifier, confidence ≥ 0.9)

   on Oct 12, 2026
6. Hermsi1337 commented on Oct 20, 2026

   

   Hermsi1337

   on Oct 20, 2026

   More actions

   +1 from a self-hosted setup on mittwald AI Hosting (German hoster, OpenAI-compatible API behind LiteLLM).

   Their /v1/rerank endpoint already speaks the Jina format. I tested it with Qwen3-VL-Reranker-2B:

   POST https://llm.aihosting.mittwald.de/v1/rerank
   {"model": "Qwen3-VL-Reranker-2B", "query": "...", "documents": ["...", "...", "..."], "top_n": 3}
   → {"results": [{"index": 1, "relevance_score": 0.42, "document": {"text": "..."}}, ...]}

   So jinaApiUrl would work fine. The only thing blocking it is the model name hardcoded to jina-reranker-v2-base-multilingual in @librechat/agents (tools/search/rerankers.ts). Since I can't rename the model on a hosted service, I run reranking with rerankerType: "none" for now.

   I saw the plan to route reranking through the RAG API (rag-api reranker type in agents, rag-api #318). Will that path also let me point at an external Jina-compatible endpoint with my own model name? If not, I'm happy to help get #12121 / agents #66 rebased and tested against a real non-Jina backend.

   👍React with 👍1Reacted by Pascal Garber