SimpleWikiSearch: A Clean Offline Wikipedia Environment for Agentic Search
Researchers have introduced SimpleWikiSearch, an offline Wikipedia setup designed to assess LLM-based agentic search systems in a reproducible manner. This environment tackles vague aspects in existing evaluations by offering a clear corpus construction, retrieval framework, tool specifications, and evaluation methods. It utilizes a complete English Wikipedia dump, processes and segments the corpus, creates both keyword and dense retrieval indexes, and provides a basic tool interface featuring search, open_url, and submit_answer functions. Baseline outcomes on six QA datasets utilizing open-source LLMs are presented. This initiative seeks to enhance the comparability and reproducibility of studies in agentic search.
Key facts
- SimpleWikiSearch is an offline Wikipedia environment for agentic search.
- It addresses under-specified details in current LLM-based agentic search evaluations.
- The environment includes explicit corpus construction, retrieval stack, tool contract, and evaluation protocol.
- It starts from a full English Wikipedia dump.
- It cleans and chunks the corpus.
- It builds keyword and dense retrieval indexes.
- The tool interface consists of search, open_url, and submit_answer.
- Baseline results on six QA datasets using open-source LLMs are reported.
Entities
Institutions
- arXiv