vectifyai/pageindex

PageIndex is a document-indexing system for what it calls "vectorless, reasoning-based RAG." RAG stands for retrieval-augmented generation, the common technique where a large language model is given relevant pieces of a document so it can answer questions about it accurately. The usual approach chops documents into small chunks, turns each chunk into a numeric vector with an embedding model, and stores those vectors in a database, at query time it pulls the chunks whose vectors look most similar to the question. The README argues this works on shallow similarity rather than true relevance, which often misses the point on long, professional documents that need domain expertise and multi-step reasoning. The way PageIndex works is different. Instead of a vector database, it builds a hierarchical "table-of-contents" tree index of each document, sections, subsections, page references, and then has the LLM reason its way through that tree to find the parts that actually answer the question. The README describes this as a two-step pipeline: first generate the tree structure of the document, then perform retrieval as a tree search. The README highlights four core features: no vector database, no artificial chunking (sections stay as-is), human-like retrieval that mirrors how an expert flips through a document, and better explainability since each answer can cite the section and page it came from. The maintainers report state-of-the-art accuracy of 98.7% on a benchmark called FinanceBench. The repository also mentions an "agentic vectorless RAG" example built on the OpenAI Agents SDK, a "PageIndex File System" that lets the index scale across many documents, and a vision-based variant that works directly on PDF page images. You would use PageIndex when you need to ask questions over long professional documents, finance reports, legal filings, manuals, and want answers that are traceable to specific pages rather than approximated by similarity search. Deployment options listed include self-hosting from this open-source repository, a managed cloud service with stronger OCR and tree building, and an enterprise option. The code is in Python.