Reranking Models
PRX-Memory supports multiple reranking providers through the prx-memory-rerank crate. Each provider implements the same adapter trait, allowing seamless switching.
Jina AI
Jina offers cross-encoder reranking models with multilingual support.
bash
PRX_RERANK_PROVIDER=jina
PRX_RERANK_API_KEY=your_jina_key
PRX_RERANK_MODEL=jina-reranker-v2-base-multilingual| Model | Notes |
|---|---|
jina-reranker-v2-base-multilingual | Multilingual cross-encoder |
jina-reranker-v1-base-en | English-optimized |
INFO
Jina reranking can use the same API key as Jina embedding. Set JINA_API_KEY once to cover both.
Cohere
Cohere provides high-quality reranking through their Rerank API.
bash
PRX_RERANK_PROVIDER=cohere
PRX_RERANK_API_KEY=your_cohere_key
PRX_RERANK_MODEL=rerank-v3.5| Model | Notes |
|---|---|
rerank-v3.5 | Latest model, best quality |
rerank-english-v3.0 | English-optimized |
rerank-multilingual-v3.0 | Multilingual support |
Pinecone
Pinecone offers reranking as part of their inference API.
bash
PRX_RERANK_PROVIDER=pinecone
PRX_RERANK_API_KEY=your_pinecone_key
PRX_RERANK_MODEL=bge-reranker-v2-m3For custom Pinecone-compatible endpoints:
bash
PRX_RERANK_PROVIDER=pinecone-compatible
PRX_RERANK_API_KEY=your_key
PRX_RERANK_ENDPOINT=https://your-endpoint.example.com
PRX_RERANK_API_VERSION=2025-01Choosing a Reranker
| Priority | Recommended Provider | Model |
|---|---|---|
| Best quality | Cohere | rerank-v3.5 |
| Multilingual | Jina | jina-reranker-v2-base-multilingual |
| Integrated with Pinecone | Pinecone | bge-reranker-v2-m3 |
| No reranking needed | -- | PRX_RERANK_PROVIDER=none |
Combining Embedding and Reranking
A common high-quality configuration pairs Jina embeddings with Cohere reranking:
bash
# Embedding
PRX_EMBED_PROVIDER=jina
PRX_EMBED_API_KEY=your_jina_key
PRX_EMBED_MODEL=jina-embeddings-v3
# Reranking
PRX_RERANK_PROVIDER=cohere
PRX_RERANK_API_KEY=your_cohere_key
PRX_RERANK_MODEL=rerank-v3.5This setup leverages Jina's fast multilingual embeddings for broad retrieval and Cohere's high-precision reranker for final ordering.
Next Steps
- Embedding Models -- First-stage embedding model options
- Configuration Reference -- All environment variables