PageIndex

PageIndex is an open-source RAG framework developed by Vectify AI that replaces vector similarity search with LLM-driven tree search over structured document indexes. It is available as a self-hosted Python package, a cloud chat platform, and an MCP/API service for developers and enterprises. The project is authored by a team of AI researchers from UCL and Oxford with backgrounds at Anthropic and UiPath.

What It Is

PageIndex is a vectorless, reasoning-based retrieval-augmented generation (RAG) system. Instead of embedding documents into vector space and retrieving by cosine similarity, it builds a hierarchical "table of contents" tree index from a PDF or Markdown document, then uses an LLM to reason over that tree — simulating how a human expert would navigate a complex document. The core insight is that similarity ≠ relevance: professional documents require multi-step reasoning to find the right section, not approximate nearest-neighbor lookup.

Retrieval works in two steps:

• Index generation: the document is parsed into a semantic tree of titled nodes, each with a page range and LLM-generated summary.
• Tree search: at query time, an LLM reasons over the tree to identify and retrieve the most relevant nodes, incorporating full conversation history and domain context.

Architecture: No Vectors, No Chunks

Traditional RAG pipelines split documents into fixed-size chunks, embed them, and store them in a vector database. PageIndex discards all three of those steps. Documents are organized into natural sections that mirror the document's own structure. Retrieval is traceable — every answer cites the specific page and section from which it was drawn, making results interpretable rather than opaque. The open-source package supports standard PDF parsing and Markdown files; the cloud service adds enhanced OCR and a more robust tree-building pipeline for complex PDFs.

Key architectural properties:

• No vector database dependency
• No chunking — sections follow document structure
• Context-aware retrieval that incorporates conversation history
• Page and section references for full traceability
• Multi-LLM support via LiteLLM (OpenAI, and other providers)

Deployment Options

PageIndex offers three deployment paths:

• Self-hosted: run the open-source Python package locally with standard PDF parsing; install via pip and point at any PDF or Markdown file.
• Cloud service: production-grade pipeline with enhanced OCR, accessible via the PageIndex Chat platform, MCP integration, or REST API.
• Enterprise: private or on-premises deployment; contact the team for details.

The self-hosted path requires an LLM API key (e.g., OpenAI) and a few CLI commands. The cloud service is accessible immediately through the chat interface without any setup.

Performance Signal: FinanceBench

The PageIndex team reports that Mafin 2.5 — a reasoning-based RAG system for financial document analysis powered by PageIndex — achieved 98.7% accuracy on the FinanceBench benchmark, which tests question answering over SEC filings and earnings disclosures. The team attributes this result to PageIndex's hierarchical indexing and reasoning-driven retrieval, which they claim significantly outperforms traditional vector-based RAG on this benchmark. Full benchmark results are published in the VectifyAI/Mafin2.5-FinanceBench GitHub repository.

Update: Agentic Vectorless RAG and PageIndex File System

Recent updates to the project include two notable additions. The PageIndex File System extends the tree index to corpus-level search, allowing PageIndex to reason over millions of documents rather than a single file by adding a file-level tree layer above individual document trees. The Agentic Vectorless RAG example demonstrates an end-to-end agentic pipeline using the OpenAI Agents SDK with self-hosted PageIndex, providing a minimal but complete reference implementation. The project is cited as: Mingtian Zhang, Yu Tang and PageIndex Team, "PageIndex: Next-Generation Vectorless, Reasoning-based RAG," PageIndex Blog, Sep 2025.

Who It Is For

PageIndex targets developers and enterprises working with long, complex professional documents — financial reports, regulatory filings, legal manuals, academic textbooks, and technical documentation that exceeds LLM context windows. The chat platform serves non-technical users who need verifiable, source-grounded answers from uploaded documents. The MCP and API interfaces serve developers integrating document intelligence into their own applications or agent pipelines.