PageIndex: Next Generation Vectorless, Reasoning-Based RAG (pageindex.ai)

PageIndex is a new, “vectorless” Retrieval-Augmented Generation (RAG) framework that replaces static embedding-based retrieval with an LLM-driven, reasoning-first approach to navigate long, structured documents. Instead of relying on vector similarity and fixed chunks, PageIndex provides a JSON-style Table of Contents (ToC) tree that the model reads, traverses, and uses to iteratively decide which sections to fetch and when it has enough evidence to answer a question. This matters because traditional vector RAG degrades with long contexts, mishandles intent-vs-content mismatch, fragments meaning with hard chunking, ignores chat history, and misses in-document references—limitations that undermine accuracy for legal, financial, or technical texts. Technically, PageIndex stores a hierarchical node tree (node_id, name, description, metadata, sub_nodes) mapped to raw content; the index is fed into the LLM’s context as an “in-context” index the model can reference and reason over. Retrieval follows a loop: read ToC → select section → extract content → assess sufficiency → iterate or answer. The model can fetch coherent sections (full pages/chapters), follow cross-references (e.g., “see Appendix G”), pull neighboring nodes if context is incomplete, and incorporate multi-turn chat history to refine searches. By shifting from similarity matching to reasoning-guided navigation, PageIndex promises higher precision, fewer hallucinations, and better multi-step document understanding without vector DBs.