PageIndex

PageIndex is a retrieval engine for long, professional documents — SEC filings, prospectuses, regulatory submissions, technical manuals — built by Vectify AI. It replaces the usual vector pipeline with a hierarchical tree index and an LLM that walks that tree, so there is no chunking, no embedding model and no vector database anywhere in the stack.

Why vectorless

Vector search retrieves by semantic similarity, and on dense professional documents similarity is a poor proxy for relevance: it misses passages that are relevant but phrased differently, and returns passages that read alike but answer nothing. PageIndex takes the approach the project compares to AlphaGo — build a structured search space, then reason through it. Indexing generates a tree that mirrors the document's own table of contents; retrieval is an agentic traversal of that tree, the way an analyst turns to the section they know will hold the answer.

The practical payoff is traceability. Because retrieval lands on a real node of a real document rather than an anonymous chunk, every answer comes back with the page it came from, and the path the model took is inspectable.

What you get

• Tree indexing, two speeds. PageIndex Flash extracts structure heuristically from the PDF's own layout and calls a model only for node summaries, turning a document around in seconds; the full LLM-built index trades time for structure on documents with no usable layout information.
• PageIndex File System extends the same idea above the document level, tree-indexing an entire corpus so the engine can reason across millions of files rather than one at a time.
• MCP server and REST API, so Claude, Cursor or any agent framework can query an indexed corpus as a tool.
• Python and JavaScript SDKs. pip install -U pageindex ships a local mode that indexes, retrieves and chats entirely on your machine against your own LLM key, or points the same client at PageIndex Cloud.
• PageIndex Chat (chat.pageindex.ai), a hosted document-analysis agent for people who want the engine without writing code.

Track record and cost

The project reports 98.7% accuracy on FinanceBench, the financial-document QA benchmark, against roughly 50% for vector-based RAG — the evaluation and its results are published in a companion repository. The maintainers' own measurements put indexing at roughly $0.0011 per page, with nine test PDFs spanning 9 to 1,098 pages indexing in 13 seconds to 4.5 minutes.

Licence and pricing

The core engine is open source under the MIT licence. PageIndex Cloud starts with a free trial of 200 credits and 200 active pages; Standard is $30/month for 1,000 monthly credits and 10,000 active pages; Pro is $50/month for 2,000 credits and 50,000 pages; Max is $100/month for 6,000 credits and 500,000 pages, with multiple workspaces and priority support. Indexing costs one credit per page as a one-time charge, top-up credits are $0.01 each and never expire, and bringing your own LLM key is unmetered. Enterprise adds private deployment, SLAs and dedicated support.