#4545th of 67 in RAG, knowledge and search

pyserini

Python toolkit for reproducible sparse and dense retrieval research

Stars
2.2k
License
Apache-2.0
Last commit
Nov 2026
Last release
Aug 2026
Language
Python
Stars, last 29 weeks+172
25 Apr 2026: 2,044 stars2 May 2026: 2,045 stars9 May 2026: 2,046 stars16 May 2026: 2,047 stars23 May 2026: 2,071 stars30 May 2026: 2,077 stars6 Jun 2026: 2,081 stars13 Jun 2026: 2,086 stars20 Jun 2026: 2,087 stars27 Jun 2026: 2,091 stars4 Jul 2026: 2,093 stars11 Jul 2026: 2,094 stars18 Jul 2026: 2,096 stars25 Jul 2026: 2,098 stars1 Aug 2026: 2,103 stars8 Aug 2026: 2,139 stars15 Aug 2026: 2,146 stars22 Aug 2026: 2,149 stars29 Aug 2026: 2,152 stars5 Sept 2026: 2,156 stars12 Sept 2026: 2,159 stars19 Sept 2026: 2,161 stars26 Sept 2026: 2,164 stars3 Oct 2026: 2,165 stars10 Oct 2026: 2,167 stars17 Oct 2026: 2,172 stars24 Oct 2026: 2,173 stars31 Oct 2026: 2,177 stars7 Nov 2026: 2,206 stars14 Nov 2026: 2,216 stars

2,044 on 25 Apr 20262,216 on 14 Nov 2026

From GitHub's daily star counts.

Checked on GitHub.

Overview

Pyserini is a Python toolkit for first-stage retrieval in IR research. It wraps Anserini (Lucene) for BM25 and learned sparse models such as uniCOIL and SPLADE, and Faiss for dense models such as DPR, Contriever and BGE, with hybrid fusion. It ships prebuilt indexes, queries, relevance judgments and evaluation scripts for MS MARCO, BEIR and other collections, and offers a REST API and an MCP server.

Who it is for: IR researchers and engineers building or benchmarking retrieval pipelines

Strengths

  • Prebuilt indexes and two-click reproductions for MS MARCO, BEIR, BRIGHT, MIRACL and more
  • BM25, learned sparse, dense and hybrid retrieval behind one Python API
  • Can index your own corpus as BM25, sparse vector or dense vector indexes
  • Ships a REST API and MCP server, plus Index Reader, Analyzer and Query Builder APIs

Weaknesses

  • Needs Java 21 and Python 3.12, with PyTorch and Faiss installed separately
  • Research-oriented first-stage retrieval; no built-in answer generation or chat UI
  • README notes dependency version mismatches as a source of install problems
  • Some prebuilt corpora are large, up to 72 GB for MS MARCO V2 doc uniCOIL

What it needs

  • GPU optional
  • Needs Java 21, Anserini (Lucene 10.5.0), Faiss, PyTorch
  • Models: BM25, uniCOIL, SPLADE, DPR, Contriever

What changed

From the release notes. No flag means the note does not say so, not that the upgrade is safe.

Also in RAG, knowledge and search

See all 67
Also in RAG, knowledge and search
RankProjectScore
1Open NotebookSelf-hosted NotebookLM alternative for research notebooks, chat, and podcasts42.1k starsMIT92 out of 100
2LightRAGGraph-based RAG framework with a REST server and WebUI40.8k starsMIT92 out of 100
3RAGFlowRAG engine with document parsing, agentic retrieval and citations Live demo ↗ (opens in a new tab)94.4k starsApache-2.090 out of 100
#4deeptutorTutoring workspace with knowledge bases, guided learning, memory and quizzes60.2k starsApache-2.090 out of 100
#5MaxKBEnterprise knowledge-base agent platform with RAG, workflows and MCP tools23.2k starsGPL-3.089 out of 100