PageIndex is a retrieval-augmented generation (RAG) engine that replaces vector search with LLM reasoning over a document's structure. Its premise is that similarity is not the same as relevance: on long professional documents, similarity search can miss relevant passages and return similar but irrelevant ones. PageIndex instead builds a hierarchical tree index of each document, like a table of contents, and has an LLM navigate that tree to find the right sections, without a vector database or chunking.
Retrieval happens in two steps. Indexing extracts the tree from the document layout and uses a basic model to summarize and refine it, at about $0.001 per page in local mode. Retrieval then has a stronger chat model search the tree, and answers trace back to explicit page references. It targets financial reports, legal and regulatory filings, technical manuals, medical literature, and textbooks.
| Local (open source) | Cloud | |
|---|---|---|
| Handles | Text-based PDFs | Text-based, scanned, and image-rich documents |
| Indexing and storage | On your machine | Managed by PageIndex |
| Citations | Page-level | Block-level |
| OCR, metadata, folders, MCP server | No | Yes |
PageIndex reports 98.7% accuracy on the FinanceBench financial document QA benchmark, against 50% for vector RAG. Compared with passing the whole PDF to the model on every question, native PDF input cost 2.1 times more at 52 pages and 16.6 times more at 420 pages, and an 805-page document no longer fit the context window. The same client works in local mode with your own LLM key or against PageIndex Cloud, and its tools can be added to the OpenAI Agents SDK, the Claude Agent SDK, or other agent frameworks.
Features
- Tree index: builds a hierarchical, table-of-contents style index per document instead of embeddings and chunks
- Reasoning-based retrieval: an LLM searches the tree, using conversation history and domain context rather than a query embedding alone
- Traceable answers: retrieved content cites explicit page references locally and block-level references in the cloud
- Local mode: index, retrieve, and chat entirely on your machine with your own LLM key
- PageIndex Flash: fast tree generation for text-based PDFs, the default local indexing method
- Split model roles: a basic model for indexing and a stronger model for tree search
- Agent framework tools: PageIndex tools plug into the OpenAI Agents SDK, the Claude Agent SDK, and other frameworks
- PageIndex Cloud: managed parsing, OCR, image understanding, storage, metadata, folders, and an MCP server
- PageIndex File System: a cloud-only file-level tree layer for reasoning over a whole document corpus
