| license: mit | |
| task_categories: | |
| - text-retrieval | |
| tags: | |
| - agentic-search | |
| - bm25 | |
| # Pi-Serini | |
| This repository contains data and artifacts associated with the paper [Rethinking Agentic Search with Pi-Serini: Is Lexical Retrieval Sufficient?](https://huggingface.co/papers/2605.10848). | |
| Pi-Serini is a reusable, benchmark-driven search-agent workspace designed for index-driven BM25 retrieval, agentic search, and benchmark-aware evaluation. It revisits whether a lexical retriever (BM25) suffices when paired with capable frontier Large Language Models (LLMs) in an agentic loop. | |
| - **Project Page:** [https://ricky42613.github.io/piserini](https://ricky42613.github.io/piserini) | |
| - **GitHub Repository:** [https://github.com/justram/pi-serini](https://github.com/justram/pi-serini) | |
| ## Supported Benchmarks | |
| The Pi-Serini framework supports several benchmarks for evaluation: | |
| - `browsecomp-plus`: Default packaged benchmark used to evaluate deep research capabilities. | |
| - `msmarco-v1-passage`: Index-driven MS MARCO v1 passage benchmark. | |
| - `benchmark-template`: A tiny local end-to-end demo benchmark for validation. | |
| ## Citation | |
| ```bibtex | |
| @misc{hsu2026rethinkingagenticsearchpiserini, | |
| title = {Rethinking Agentic Search with Pi-Serini: Is Lexical Retrieval Sufficient?}, | |
| author = {Tz-Huan Hsu and Jheng-Hong Yang and Jimmy Lin}, | |
| year = {2026}, | |
| eprint = {2605.10848}, | |
| archivePrefix = {arXiv}, | |
| primaryClass = {cs.IR}, | |
| url = {https://arxiv.org/abs/2605.10848} | |
| } | |
| ``` |