Skip to content

Reranker

Example Code: examples/knowledge/reranker

Reranker reorders retrieved candidates to improve relevance. It can work with any vector store and is injected via knowledge.WithReranker.

Supported Rerankers

TopK (simple truncation)

The most basic reranker that returns the top K results based on retrieval scores.

1
2
3
4
5
import "trpc.group/trpc-go/trpc-agent-go/knowledge/reranker/topk"

rerank := topk.New(
    topk.WithK(3), // return top 3 results (default: -1, return all)
)
Option Description Required
WithK(int) Number of results to return, default -1 (return all) No

Cohere (SaaS rerank)

Use Cohere's rerank service.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
import (
    "log"

    "trpc.group/trpc-go/trpc-agent-go/knowledge/reranker/cohere"
)

rerank, err := cohere.New(
    cohere.WithAPIKey("your-api-key"),       // required: Cohere API Key
    cohere.WithModel("rerank-english-v3.0"), // model name (default: rerank-english-v3.0)
    cohere.WithTopN(5),                      // number of results to return
    cohere.WithEndpoint("https://..."),      // optional: custom endpoint (default: https://api.cohere.ai/v1/rerank)
    cohere.WithHTTPClient(customClient),     // optional: custom HTTP client
)
if err != nil {
    log.Fatalf("Failed to create reranker: %v", err)
}
Option Description Required
WithAPIKey(string) Cohere API Key Yes
WithModel(string) Model name, default rerank-english-v3.0 No
WithTopN(int) Number of results to return No
WithEndpoint(string) Custom endpoint URL No
WithHTTPClient(*http.Client) Custom HTTP client No

Infinity / TEI (standard rerank API)

Terminology

  • Infinity: Open-source high-performance inference engine supporting multiple Reranker models
  • TEI (Text Embeddings Inference): Official Hugging Face inference engine optimized for Embeddings and Reranking

The Infinity Reranker implementation in trpc-agent-go can connect to any service compatible with the standard Rerank API, including self-hosted services using Infinity/TEI, Hugging Face Inference Endpoints, etc.

Usage

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
import (
    "log"

    "trpc.group/trpc-go/trpc-agent-go/knowledge/reranker/infinity"
)

rerank, err := infinity.New(
    infinity.WithEndpoint("http://localhost:7997/rerank"), // required: service endpoint
    infinity.WithModel("BAAI/bge-reranker-v2-m3"),         // optional: model name
    infinity.WithTopN(5),                                  // optional: number of results (default: -1, return all)
    infinity.WithAPIKey("your-api-key"),                   // optional: API key for authenticated services
    infinity.WithHTTPClient(customClient),                 // optional: custom HTTP client
)
if err != nil {
    log.Fatalf("Failed to create reranker: %v", err)
}
Option Description Required
WithEndpoint(string) Service endpoint URL Yes
WithModel(string) Model name No
WithTopN(int) Number of results to return, default -1 (return all) No
WithAPIKey(string) API key for authenticated services No
WithHTTPClient(*http.Client) Custom HTTP client No

For detailed deployment methods and examples, see the examples/knowledge/reranker/infinity/ directory.

Inject into Knowledge

1
2
3
4
5
6
import "trpc.group/trpc-go/trpc-agent-go/knowledge"

kb := knowledge.New(
    knowledge.WithReranker(rerank),
    // ... other configurations (VectorStore, Embedder, Sources, etc.)
)

Notes

  • Cohere requires a valid API key.
  • Infinity/TEI requires a reachable endpoint and a loaded model; WithModel is optional but should match the service model when set.
  • TopK has no external dependencies and suits offline or constrained environments.