rerankers: Use Diverse Reranking Models Through One Python API

Overview

rerankers is a Python library that gives different reranking approaches a common interface. It addresses the integration overhead of choosing among models and services that otherwise have their own APIs and result formats.

It is a practical fit for retrieval and search developers who want to try local models, hosted reranking APIs, or different model architectures within the same application. The core package has no dependencies by default, while optional extras install dependencies for particular model families.

Key Features

• Load supported rerankers through a shared Reranker interface, with optional model-type selection.
• Rerank a query's candidate documents using a common rank() method and receive RankedResults.
• Use local cross-encoders, T5-based rankers, ColBERT, FlashRank, and LLM-based approaches.
• Connect to supported hosted reranking APIs, including Cohere, Jina, Voyage, MixedBread, Pinecone, and Isaacus.
• Preserve document IDs and metadata in results, and retrieve sorted selections with top_k().
• Use rank_async() as an asynchronous wrapper around ranking.
• Install optional dependency groups for selected backends rather than requiring all model dependencies in the core package.

Use Cases

• Retrieval engineers can add a reranking stage after initial candidate retrieval to reorder documents for a query.
• Developers evaluating reranking approaches can compare different model families behind a consistent interface.
• Application teams can switch between local inference and supported API providers while keeping much of their ranking code unchanged.
• Python developers integrating ranking into an existing pipeline can pass document text or Document objects and retain metadata with results.

Project Facts

• Language: Python
• License: Apache-2.0
• Stars: 1.6k
• Forks: 106
• Topics: none listed
• Archived: false

Getting Started

Install the core package:

pip install rerankers

Model backends may need additional dependencies. See the README for optional installs and usage examples.

Considerations

• The core installation does not install model-specific dependencies. Choose the relevant optional extra or provide the dependencies required by the backend.
• Hosted API rerankers require provider credentials. Local transformer-based models can have separate hardware and memory requirements depending on the model; the repository input does not specify fixed requirements.
• RankLLM support is documented as requiring Python 3.10 or later. The README also identifies some RankLLM model support as untested.
• The project describes itself as a beta release, so check the repository and release history for current behavior and compatibility before adopting it in production.