PageIndex: Vectorless Reasoning-based RAG Hits 98.7% on FinanceBench
victor_explore · x · 2026-11-21
VectifyAI's open-source PageIndex (40.6k stars, MIT) ditches the vector DB and chunking entirely: each document becomes a table-of-contents tree, and an LLM reasons its way down to the right section — fixing the classic vector RAG failure of retrieving similar-sounding but wrong chunks.
- 98.7% accuracy on FinanceBench, far above the 50% vector-RAG baseline
- New Page Search API finds answer-bearing pages in 1 second and sends only those to the LLM, cutting token cost by 97% while beating vector search on accuracy and speed
- Ships with cookbook, examples, and agentic APIs like collectioncontext()
More from coding & agent
- GPT-6.2 Astra and Sol Take #1 and #3 on DesignArena Game Dev Leaderboard — sandersted · 2026-11-21
- Anthropic launches 532 free Claude plugins from Stripe, Shopify, Canva and more — nikola_mr64990 · 2026-11-21
- PostHog cuts LLM costs with decision models: up to 444.6x cheaper, 193.6x faster — yawnxyz · 2026-11-21
- Dev builds GURU, a local-first personal AI framework with persistent memory and device control — iamth3morningstar · 2026-11-21
- Security researcher's 'software factory' playbook: write specs for days, add PR review late — nptacek · 2026-11-21
- Vibe coding with Claude over live voice to build and iterate a portfolio site — Different_Scholar_74 · 2026-11-21