PageIndex beats vector DBs
PageIndex — a hierarchical tree-index approach inspired by AlphaGo — reportedly replaces vector DBs for some RAG tasks and scored 98.7% on FinanceBench versus Perplexity’s 45%, claiming a fix for semantic-similarity failures in financial docs. That suggests alternative retrieval structures can outperform embedding-only pipelines on certain enterprise datasets. (x.com)
VectifyAI publicly released the PageIndex framework and the Mafin 2.5 agent in late February 2026, announcing the results and making the code available to developers. (marktechpost.com) The PageIndex GitHub repository shows active development with thousands of stars and an MIT license, signaling rapid community uptake and permissive reuse. (github.com) FinanceBench, the evaluation suite used in the announcement, contains 10,231 annotated questions about public-company filings and publishes a 150-question open sample that teams commonly use for reproducible testing. (arxiv.org) VectifyAI’s Mafin2.5 evaluation artifacts are public in the project repository and include evaluation scripts (eval.py) and JSON result files that show per-system outputs and the team’s human-evaluation protocol for ambiguous cases. (github.com) The PageIndex technical writeups describe building a table-of-contents-style hierarchical tree for each document and steering LLM-driven tree search across nodes rather than relying on fixed-size text chunk retrieval. (github.com) Product posts and coverage emphasize PageIndex’s “vision-native” handling of page layout, explicit node metadata for traceability, and preservation of header–cell relationships in tables as mechanisms to retain the structural context that OCR+chunking pipelines often lose. (marktechpost.com) The Mafin2.5 materials state that the pipeline has been validated across multiple base model backends (including a ChatGPT 4o configuration and a privately deployable Deepseek v3) and that the evaluation covers the full public FinanceBench set in the project repo. (github.com)
