Submit a resource

Navigation

Back to directory

FlashRank

FlashRank is an ultra-lightweight Python library that reranks search results before they are fed to an LLM, using cross-encoder and listwise LLM-based rerankers small enough to run on a CPU with no PyTorch or Transformers dependency.

Visit website (opens in a new tab)github.com
FlashRank

FlashRank is an ultra-lightweight, fast Python library that adds reranking to an existing search and retrieval pipeline, based on cross-encoders and other reranking models. It needs no PyTorch or Transformers dependency and runs on CPU, with its smallest supported model at about 4 megabytes.

It supports pairwise or pointwise cross-encoder rerankers with a 512-token limit and listwise LLM-based rerankers with an 8,192-token limit, and ships several named models of different sizes and languages, including a multilingual model covering more than 100 languages and a dedicated Arabic reranker. Because deployments such as serverless functions are often billed by memory and time per invocation, the library favors a small package size to keep cold-start times and per-call cost low.

More in RAG

See all 540

  • ai·rete·rag

    ai-rete-rag.com

    A platform that pairs a Rete rule engine with retrieval-augmented generation, so deterministic rules decide an outcome and grounded retrieval explains the decision in plain language.

  • Xtriever

    github.com

    A hybrid retrieval engine for retrieval-augmented generation, written in Rust to run on-device on a phone or a laptop, combining lexical BM25 search, dense embedding search, reciprocal-rank fusion and cross-encoder re-ranking with no server.

  • FileFlux

    github.com

    A .NET library that turns PDF, DOCX, HWP and other documents into optimized chunks for RAG systems, through a five-stage pipeline built on high-performance Rust FFI readers.

Cookies

We use necessary cookies to run the site. Measurement and marketing are optional; neither is used today. Privacy policy