PageIndex launches tree-based reasoning retrieval, with FinanceBench testing outperforming traditional RAG

AIMPACT News, April 26 (UTC+8), traditional RAG systems rely on vector similarity search, which is limited in long professional documents because the correct answer is often not in the most semantically similar paragraph. PageIndex introduces a new retrieval method by constructing a hierarchical directory-style tree index to replace traditional chunk embedding methods, using LLMs to reason and locate answers on the tree structure without vectors or embeddings, enabling cross-domain queries. Developers have built an index based on PageIndex for the Transformer paper “Attention Is All You Need,” and when performing cross-domain queries, GPT-5.4 can identify relevant sections solely based on node titles and summaries. This approach significantly improves retrieval accuracy in benchmarks like FinanceBench and is suitable for fields requiring precision and deep understanding, such as financial reports, research papers, and legal texts. Access at: dash.pageindex.ai.