// the find
DataScienceUIBK/Rankify
🔥 Rankify: A Comprehensive Python Toolkit for Retrieval, Re-Ranking, and Retrieval-Augmented Generation 🔥. Our toolkit integrates 40 pre-retrieved benchmark datasets and supports 7+ retrieval techniques, 24+ state-of-the-art Reranking models, and multiple RAG methods.
Rankify is a Python library that wraps a large swath of the retrieval/reranking/RAG research ecosystem (BM25, DPR, ANCE, ColBERT, BGE and newer LLM-based retrievers; 24+ rerankers; multiple RAG generation methods) behind one consistent API, plus 40 pre-retrieved benchmark datasets for evaluation. It's aimed at IR/NLP researchers who want to benchmark or swap retrieval pipelines without reimplementing every paper's reference code.
The one-line Retriever(method=...)/Reranking(method=...) interface genuinely unifies dozens of otherwise incompatible third-party implementations, which is the actual hard part of this kind of project. It ships pre-retrieved benchmark datasets (BM25/DPR/ColBERT/etc. over 30+ QA datasets) on HuggingFace with a fixed JSON schema, so you can run apples-to-apples comparisons without building your own retrieval+eval harness. It has a real arXiv paper behind it and visible active development, with external contributors adding new retrievers (SFR, E5, GritLM, ReasonIR) rather than the maintainers doing it all solo.
Installation is a research-lab affair, not a pip-install-and-go library: conda env, a pinned torch 2.5.1+CUDA build, and for ColBERT specifically a manual GCC/conda-forge toolchain plus clearing torch extension caches. Running more than one reranker method in the same environment is a real dependency-conflict risk since it's stitching together vLLM, transformers, faiss, and model-specific requirements from ~24 separate upstream repos. Scope has crept well past 'retrieval toolkit' — there's a bundled Next.js demo-web app, a REST server, framework integrations, and an 'AI-powered model selection agent', all in the same package rather than split out. The README itself is a symptom: heavy badge/emoji marketing and a large commented-out HTML table left in the source, which makes it harder to find the actual API docs on first read.