PageIndex introduces a vectorless, reasoning‑based retrieval‑augmented generation system that replaces the usual vector index with a hierarchical tree index built from document layout. In local mode the SDK’s “Flash” indexer creates a tree for each text‑based PDF without requiring an LLM to perform the structural extraction; a lightweight model (e.g., gpt‑5.6‑luna) only summarizes and refines the tree, costing about $0.001 per page (≈$1 and a few minutes for a 1,000‑page textbook) and scaling linearly from ~13 s for a 9‑page file to ~4.5 min for a 1,098‑page file. Retrieval is performed by an LLM (e.g., gpt‑5.6‑sol) that reasons over the tree, yielding traceable, block‑level citations in the cloud and page‑level citations locally. On the PageIndex‑OSS‑Benchmark (62 questions over 34 PDFs totaling 1,945 pages) the system achieved 98.7 % accuracy on FinanceBench, far surpassing vector‑based RAG. Compared with feeding the whole PDF to the model, PageIndex reduces cost by 2.1× at 52 pages and 16.6× at 420 pages, and avoids context‑window limits beyond ~800 pages. Cloud mode adds OCR, image understanding, managed storage, and corpus‑scale reasoning via the PageIndex File System.