PageIndex is an open-source, vectorless RAG framework with 39,000+ stars on GitHub that retrieves information from long documents by reasoning over a hierarchical tree index instead of searching vector embeddings.
Vector-based RAG retrieves by semantic similarity, but similarity is not relevance: on long professional documents it returns passages that look similar while missing the ones that actually answer the question. Inspired by AlphaGo, PageIndex replaces the vector index with a tree-structure index built from each document’s own layout, then lets an LLM reason its way through the tree the way a human expert turns to the right section of a report. The result is traceable, explainable retrieval with explicit references instead of opaque “vibe retrieval.” PageIndex reached a state-of-the-art 98.7% accuracy on the FinanceBench financial document QA benchmark, and it is built for financial reports, legal documents, regulatory filings, technical manuals, medical literature and academic textbooks.
PageIndex is free and open source under the MIT license for self-hosting; in local mode you only pay for your own LLM API usage. PageIndex Cloud is an optional managed service that handles indexing and storage for you — check the official site for its current plans.
Developers and teams building question-answering or analysis over long, complex documents — finance, legal, compliance and research workflows where retrieval accuracy and auditability matter more than raw similarity search.