FlashRank

// Overview

FlashRank is a small Python library that adds a reranking step to an existing search or retrieval pipeline. After a first-stage retriever returns a candidate list, FlashRank reorders those passages by how well they match the query, so the most relevant ones move to the top before you pass them to an LLM.

It is aimed at developers building RAG systems and search backends who want better ranking without heavy dependencies. It does not require Torch or Transformers and runs on CPU, and the default model is about 4MB, which keeps install size and cold-start times low, including on serverless platforms like AWS Lambda.

Within the rerankers and hybrid search category, FlashRank focuses on the final reranking leg of retrieval. It supports pairwise/pointwise cross-encoders (max 512 tokens) and listwise LLM-based rerankers (max 8192 tokens), letting you trade model size for ranking quality.

// What it does

• No Torch or Transformers dependency; runs on CPU
• Default model is around 4MB, with a range of larger models for higher precision
• Supports both pairwise/pointwise cross-encoders and listwise LLM-based rerankers
• Includes a multilingual model (ms-marco-MultiBERT-L-12) covering 100+ languages and a dedicated Arabic reranker
• Small package size keeps cold starts and redeploys quick on serverless setups
• Configurable max_length to match passage size and control latency

// Getting started

Install FlashRank, create a Ranker, then build a RerankRequest from your query and retrieved passages and call rerank.

pip install flashrank

pip install flashrank[listwise]

from flashrank import Ranker, RerankRequest

ranker = Ranker(max_length=128)

query = "How to speedup LLMs?"
passages = [
    {"id": 1, "text": "Introduce lookahead decoding...", "meta": {"additional": "info1"}},
    {"id": 2, "text": "LLM inference efficiency...", "meta": {"additional": "info2"}},
]

rerankrequest = RerankRequest(query=query, passages=passages)
results = ranker.rerank(rerankrequest)
print(results)

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Reordering top-k results from a vector or keyword search before sending context to an LLM in a RAG pipeline
• Adding a low-latency reranking step to a serverless API (e.g. AWS Lambda) where package size and cold starts matter
• Improving result ordering on CPU-only or resource-constrained hosts without installing Torch
• Reranking multilingual search results using the 100+ language model or the dedicated Arabic reranker

// How FlashRank compares

FlashRank alongside other open-source rerank, search & hybrid tools AI/TLDR tracks, ranked by GitHub stars.