Reranking Engine

Reranking is an optional second-stage retrieval step that reorders candidate results using a dedicated cross-encoder model. While embedding-based retrieval is fast, it operates on pre-computed vectors that may not capture fine-grained relevance. Reranking applies a more powerful model to a smaller candidate set, significantly improving precision.

How It Works

1. First stage (retrieval): Vector similarity search returns a broad set of candidates (e.g., top 50).
2. Second stage (reranking): A cross-encoder model scores each candidate against the query, producing a refined ranking.
3. Final result: The top-k reranked results are returned to the caller.

graph LR
    QUERY["Recall Query"] --> EMBED["Embedding Search<br/>Top 50 Candidates"]
    EMBED --> RERANK["Reranker<br/>Cross-Encoder Scoring"]
    RERANK --> RESULTS["Reranked Top-K"]

Why Reranking Matters

Metric	Without Reranking	With Reranking
Recall coverage	High (broad retrieval)	Same (unchanged)
Precision at top-5	Moderate	Significantly improved
Latency	Lower (~50ms)	Higher (~150ms additional)
API cost	Embedding only	Embedding + reranking

Reranking is most valuable when:

• Your memory database is large (1000+ entries).
• Queries are ambiguous or natural language.
• Precision at the top of the result list matters more than latency.

Supported Providers

Provider	Config Value	Description
Jina	PRX_RERANK_PROVIDER=jina	Jina AI reranker models
Cohere	PRX_RERANK_PROVIDER=cohere	Cohere rerank API
Pinecone	PRX_RERANK_PROVIDER=pinecone	Pinecone rerank service
Pinecone-compatible	PRX_RERANK_PROVIDER=pinecone-compatible	Custom Pinecone-compatible endpoints
None	PRX_RERANK_PROVIDER=none	Disable reranking

Configuration

PRX_RERANK_PROVIDER=cohere
PRX_RERANK_API_KEY=your_cohere_key
PRX_RERANK_MODEL=rerank-v3.5

Provider Fallback Keys

If PRX_RERANK_API_KEY is not set, the system falls back to provider-specific keys:

• Jina: JINA_API_KEY
• Cohere: COHERE_API_KEY
• Pinecone: PINECONE_API_KEY

Disabling Reranking

To run without reranking, either omit the PRX_RERANK_PROVIDER variable or set it explicitly:

PRX_RERANK_PROVIDER=none

Recall still functions using lexical matching and vector similarity without the reranking stage.

Next Steps

• Reranking Models -- Detailed model comparison
• Embedding Engine -- First-stage retrieval
• Configuration Reference -- All environment variables