Use Case
I am running Hindsight in China using SiliconFlow (https://api.siliconflow.cn/v1)
as my API provider for both embeddings and reranking. SiliconFlow works perfectly
as an embeddings provider via HINDSIGHT_API_EMBEDDINGS_OPENAI_BASE_URL, and I want
to use the same platform for reranking with BAAI/bge-reranker-v2-m3, which exposes
a Cohere-compatible /rerank endpoint.
Problem Statement
There is no working way to use SiliconFlow (or any OpenAI-compatible reranker API)
as a reranker provider. All available providers fail:
How This Feature Would Help
SiliconFlow is widely used in China as an alternative to OpenAI/Cohere.
Users who already use SiliconFlow for embeddings naturally want to use it
for reranking too — keeping all inference on one platform with one API key.
Enabling this would make Hindsight fully functional for Chinese users without
requiring a VPN or falling back to RRF, which provides lower recall quality.
Proposed Solution
Add an openai reranker provider that mirrors how the embeddings provider works:
HINDSIGHT_API_RERANKER_PROVIDER=openai
HINDSIGHT_API_RERANKER_OPENAI_BASE_URL=https://api.siliconflow.cn/v1
HINDSIGHT_API_RERANKER_OPENAI_API_KEY=sk-xxx
HINDSIGHT_API_RERANKER_OPENAI_MODEL=BAAI/bge-reranker-v2-m3
This is consistent with the existing embeddings design and would work for any
OpenAI-compatible /rerank endpoint, not just SiliconFlow.
Alternatives Considered
- cohere provider: fails with 404 due to incorrect URL path construction
- litellm provider: fails with 400 due to request body format mismatch
- rrf provider: works but no neural reranking, lower recall quality
- Hosting a local TEI reranker container: works but adds infrastructure complexity
and is overkill for users who just want a simple API-based solution
Priority
Important - affects my workflow
Additional Context
No response
Checklist
Use Case
I am running Hindsight in China using SiliconFlow (https://api.siliconflow.cn/v1)
as my API provider for both embeddings and reranking. SiliconFlow works perfectly
as an embeddings provider via HINDSIGHT_API_EMBEDDINGS_OPENAI_BASE_URL, and I want
to use the same platform for reranking with BAAI/bge-reranker-v2-m3, which exposes
a Cohere-compatible /rerank endpoint.
Problem Statement
There is no working way to use SiliconFlow (or any OpenAI-compatible reranker API)
as a reranker provider. All available providers fail:
cohere + RERANKER_COHERE_BASE_URL=https://api.siliconflow.cn/v1
→ 404: Cohere SDK does not append /rerank to the base URL
litellm + LITELLM_API_BASE=https://api.siliconflow.cn/v1
→ 400: Request reaches /v1/rerank but request body is rejected
How This Feature Would Help
SiliconFlow is widely used in China as an alternative to OpenAI/Cohere.
Users who already use SiliconFlow for embeddings naturally want to use it
for reranking too — keeping all inference on one platform with one API key.
Enabling this would make Hindsight fully functional for Chinese users without
requiring a VPN or falling back to RRF, which provides lower recall quality.
Proposed Solution
Add an openai reranker provider that mirrors how the embeddings provider works:
HINDSIGHT_API_RERANKER_PROVIDER=openai
HINDSIGHT_API_RERANKER_OPENAI_BASE_URL=https://api.siliconflow.cn/v1
HINDSIGHT_API_RERANKER_OPENAI_API_KEY=sk-xxx
HINDSIGHT_API_RERANKER_OPENAI_MODEL=BAAI/bge-reranker-v2-m3
This is consistent with the existing embeddings design and would work for any
OpenAI-compatible /rerank endpoint, not just SiliconFlow.
Alternatives Considered
and is overkill for users who just want a simple API-based solution
Priority
Important - affects my workflow
Additional Context
No response
Checklist