←back to docs

Rerankers

Explore the capabilities of the reranking algorithm in ChatBotKit, designed to enhance the accuracy and relevance of AI responses by employing a sophisticated second-level sorting mechanism for retrieved records.

Rerankers improve the quality of search results from your datasets by applying an intelligent second-pass sorting step. After a dataset store retrieves an initial set of matching records, the reranker scores and reorders them so the most relevant results reach your AI first. This feature works with Retrieval-Augmented Generation (RAG) workflows and is currently in beta.

How Reranking Works

  1. Initial retrieval - The dataset store searches your records and returns a candidate set based on the query.
  2. Reranking - The selected reranker model scores each candidate for relevance to the query and reorders them.
  3. Improved results - The AI receives the reordered list, so the best-matching content is prioritized in its response.

Available Rerankers

RerankerDescription
BGE v2 M3Lightweight model with strong multilingual capabilities and fast inference. This is the default reranker.
Sprout v0High-accuracy multilingual reranker with fast inference.
Cohere 3.5Cohere's leading reranking model, balancing performance and latency for enterprise search applications.

All rerankers process up to 20 records by default.

Setup Instructions

Reranking is configured per dataset.

  1. Navigate to your Dataset Settings in the ChatBotKit dashboard.
  2. Find the Reranker field.
  3. Select a reranker model from the dropdown.
  4. Optionally click the options icon to adjust Max Records or Region.
  5. Save your changes.

To disable reranking for a dataset, select the empty option in the reranker dropdown.

Configuration Options

After selecting a reranker, you can adjust two settings via the options icon:

  • Max Records - The maximum number of records the reranker will process. Increasing this value may improve result quality at the cost of slightly higher latency. The default is 20.
  • Region - The region where the reranker runs. All rerankers are currently available in the us region.

Best Practices

  • Start with BGE v2 M3 - It is the default and covers most use cases with fast response times.
  • Try Cohere 3.5 for enterprise or high-precision needs - It is designed for applications where accuracy outweighs latency.
  • Test with representative queries - Enable reranking, run typical user queries, and compare result quality before and after.
storagererankersearchRAG