FlashRank

Why we included this project

First-stage retrieval usually returns a wide net of candidates, and the reranker you put after it decides how much of that noise reaches your LLM. FlashRank is a small Python library for that last step: it reranks candidates with either compact cross-encoder models or listwise LLM-based rerankers. The practical appeal is the footprint. The default model is a few megabytes and runs on CPU with no Torch or Transformers, so serverless deployments get shorter cold starts and cheaper invocations. Teams doing RAG, hybrid search, or document semantic search can improve precision without standing up heavy infrastructure. Several model options are available, including multilingual and domain-tuned ones, so you can pick between speed and accuracy based on your data and latency budget.