Show HN: PageIndex for Reasoning-Based RAG (vectifyai.notion.site)

PageIndex introduces a reasoning-based Retrieval-Augmented Generation (RAG) approach that treats long documents like a human reader: it builds a JSON “Table of Contents” (ToC) tree mapping hierarchical nodes (chapters, pages, sections) to raw content and metadata, and then iteratively reasons about where to look. Instead of relying solely on vector similarity, an LLM inspects the ToC, selects likely sections, extracts content, checks whether the information is sufficient, and repeats until it can answer. The Show HN release positions this as a practical fix for LLM context-window limits and the failure modes of traditional vector-based RAG (query/knowledge mismatch, broken chunk semantics, inability to follow in-document references, and lack of conversation-awareness). Technically, PageIndex emphasizes semantically coherent retrieval (full pages/sections), recursive JSON nodes with links to raw data, and an agent-style loop that uses chat history and structural cues (e.g., “see Appendix G”) to follow references and fetch neighboring sections when context is incomplete. For AI/ML practitioners this implies better accuracy on long-document QA, reduced hallucination, and improved multi-turn coherence; it also suggests hybrid designs where ToC-driven navigation complements vector search and where robust ToC extraction and LLM tool-calling are key integration challenges.