FlashRank

What it is and what it does

FlashRank is a Python library that re-ranks search results by scoring query-passage pairs using pre-trained cross-encoder and LLM models. It sits between your retrieval stage and LLM input, improving result quality without the overhead of full transformer inference. The library emphasizes minimal footprint—the default model is ~4MB and runs on CPU—making it suitable for serverless deployments and cost-sensitive environments.

You provide a query and a list of passages; FlashRank scores and sorts them by relevance. It supports multiple model sizes and types: tiny cross-encoders for speed, larger cross-encoders for precision, T5-based listwise rerankers for out-of-domain robustness, and multilingual variants. Runtime dependencies are lightweight (tokenizers, onnxruntime, numpy, requests, tqdm), and the library handles model downloading and caching automatically.

Use it for:

• Improve LLM answer quality by re-ranking retrieval results before feeding them as context to an LLM
• Reduce inference cost in serverless deployments by using a tiny reranker model instead of a large LLM for initial filtering
• Re-rank multilingual search results using the ms-marco-MultiBERT-L-12 model for language support
• Optimize retrieval latency by tuning max_length to match your typical passage+query token count
• Fine-tune ranking for domain-specific queries using the ce-esci-MiniLM-L12-v2 model

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

FlashRank re-ranks search results using lightweight cross-encoder and LLM-based models, optimized for speed and minimal resource overhead in retrieval pipelines.

Yes. FlashRank is actively maintained, has no known vulnerabilities, and offers a practical solution for improving search-to-LLM pipelines with minimal resource cost. Install it if you're building retrieval-augmented generation systems and want to improve result quality without adding significant latency or infrastructure overhead. The permissive Apache 2.0 license poses no restrictions.

Install

flashrank on PyPI

pip install flashrank

uv add flashrank

poetry add flashrank

Installing FlashRank

Before you install

Low friction: pure Python wheel with five runtime dependencies (tokenizers, onnxruntime, numpy, requests, tqdm). Active maintenance with recent commits and strong community interest. No known vulnerabilities.

License in practice

Apache 2.0 is permissive; you can use, modify, and distribute FlashRank freely in commercial and private projects with minimal restrictions.

Quickstart

pip install flashrank

from flashrank import Ranker, RerankRequest

ranker = Ranker(max_length=128)
query = "How to speedup LLMs?"
passages = [{"id": 1, "text": "Introduce lookahead decoding..."}]
results = ranker.rank(RerankRequest(query=query, passages=passages))

Requires Python 3.6+. For LLM-based listwise rerankers, install with the [listwise] extra. Model files download on first use; cache_dir can be specified to control storage location.

Verify before relying

• Exact reranking latency and throughput benchmarks for different model sizes and passage counts
• Memory footprint during inference for each supported model variant
• Whether multilingual models work reliably for non-English queries
• Sliding window support status for rank_zephyr_7b_v1_full beyond the documented 20-passage limit

Package facts

Evidence: FlashRank-0.2.10-py3-none-any.whl

Tags

More Artificial Intelligence packages

Further reading