PageIndex kills vector DBs (claim)
PageIndex, an open-source Python tool, claims to eliminate vector DBs, chunking, and embeddings by building document trees and TOC-style retrieval—reporting 98.7% on FinanceBench and positioning itself as an alternative RAG architecture. If real at scale, it would upend assumptions about mandatory vector DB stacks for document-heavy enterprise search. (x.com)
PageIndex and the Mafin product line are published by Vectify AI, as shown on the project's official site and GitHub. (pageindex.ai). (pageindex.ai) The PageIndex GitHub repository lists about 22.1k stars, roughly 1.7k forks, and 255 commits, with recent commits as of the last week. (github.com). (github.com) The Mafin2.5 product page includes a benchmark table that reports GPT‑4o at 31% on the FinanceBench evaluation they present.. (pageindex.ai) FinanceBench is a public financial QA benchmark comprising 10,231 questions in total and a publicly available 150‑example sample that has been used in model evaluations. (patronus-ai/financebench; arXiv). (github.com) VectifyAI has published a Mafin2.5–FinanceBench repository containing evaluation scripts and JSON result files for the reported runs.. (github.com) The PageIndex codebase is released under an MIT license and the project maintains developer cookbooks, a Colab demo notebook, and API/docs for integrating the system. (github.com; docs.pageindex.ai; colab.research.google.com). (github.com) Early public discussion on Hacker News and independent write‑ups highlight the approach’s promise for long, highly structured documents while urging independent replication because the primary benchmark evaluations were produced by the same organization. (news.ycombinator.com; gpt.gekko.de). (news.ycombinator.com)
