VectifyAI/PageIndex

📑 PageIndex: Document Index for Vectorless, Reasoning-based RAG

PageIndex – Vector‑less, reasoning‑based Retrieval‑Augmented Generation

What it is – PageIndex is an open‑source RAG engine that replaces the usual vector‑store + chunking pipeline with a hierarchical tree index built from a document’s layout. At query time an LLM walks the tree, using its own reasoning (and the full conversation context) to locate the most relevant sections. The result is a retrieval step that is traceable, explainable, and context‑aware without any vector database.

Key ideas

Concept How PageIndex does it
Indexing Parses a PDF (or other text‑heavy document) into a tree that mirrors the document’s logical structure (chapters, sections, headings). A lightweight LLM model (index=) summarizes each node; the tree itself is derived from the layout, not from embeddings.
Retrieval When a question is asked, the chat model (chat=) reasons over the tree, selecting nodes step‑by‑step much like a human reader. The model sees the full conversation history, domain knowledge, and any external context, so relevance is driven by reasoning rather than pure similarity.
No vectors / no chunking Because the index is a deterministic tree, there is no need for a vector DB, similarity search, or fixed‑size chunks.
Explainability The path the model follows through the tree is visible, giving a clear citation (page‑level locally, line‑level in the cloud).

Quick‑start (local mode)

pip install -U pageindex

import os
from pageindex import PageIndexClient

os.environ["OPENAI_API_KEY"] = "your-openai-key"

client = PageIndexClient(
    index="gpt-5.6-luna",   # cheap model for building the tree
    chat="gpt-5.6-sol",    # model that does the reasoning search
)

# Upload a PDF and get a document identifier
doc_id = client.submit_document("report.pdf")['doc_id']

# Ask a question – the client walks the tree for you
answer = client.chat("What was the 2023 operating margin?", doc_id=doc_id)
print(answer)

The same SDK works against PageIndex Cloud by setting index="cloud" and providing a PAGEINDEX_API_KEY.

Main features

  • Tree‑based index that mirrors the document’s logical hierarchy.
  • LLM‑driven search: the model decides which node to explore next, allowing multi‑step reasoning.
  • Local and Cloud modes – run the whole pipeline on‑premises or let PageIndex’s managed service handle OCR, image understanding, and massive corpora (File System layer).
  • Cost‑effective: indexing costs ≈ $0.001 / page; query cost stays flat regardless of document length because only the visited nodes are sent to the model.
  • Traceable citations – every answer can be linked back to the exact page (local) or line (cloud).
  • Plug‑and‑play SDK with streaming, multi‑document search, and ready‑made tools for OpenAI/Claude agent frameworks.

Benchmarks & performance (as reported in the README)

  • Indexing – $0.001 / page, 13 s → 4.5 min for PDFs ranging from 9 to 1,098 pages.
  • Query cost – 2.1× cheaper than feeding the whole PDF to a model at 52 pages, 16.6× cheaper at 420 pages; beyond ~800 pages the full‑PDF approach exceeds context windows.
  • Accuracy – on the FinanceBench QA benchmark PageIndex achieved 98.7 % accuracy, far above the ~50 % typical of vector‑based RAG.
  • OSS benchmark – the companion PageIndex-OSS-Benchmark repo evaluates 62 lookup questions over 34 PDFs (1,945 pages) and shows a clear trade‑off curve between model size and per‑query cost.

Typical use cases

  • Financial reports, regulatory filings, legal contracts – where relevance depends on domain knowledge and multi‑step reasoning.
  • Technical manuals & medical literature – long, structured documents where precise citations are required.
  • Enterprise knowledge bases – especially when you need explainable retrieval without managing a vector DB.

Licensing & contribution

The repository is open‑source (the README does not list a specific license, so check the LICENSE file in the repo). Contributions are welcomed via pull requests; the project also offers a commercial Cloud service for OCR, image understanding, and large‑scale indexing.

Where to learn more

PageIndex lets you retrieve information from massive, complex PDFs the way a human would read them—without the overhead of vectors or opaque similarity scores.

Related

  • Project
    VectifyAI/pageindex-mcp

    A Model Context Protocol (MCP) server that enables LLMs to perform reasoning-based, vectorless RAG over long PDFs using hierarchical tree structures.

  • Project
    clusterzx/paperless-ai

    Paperless‑AI is an open‑source extension for the self‑hosted document manager Paperless‑ngx. It automatically classifies new documents, adds AI‑generated titles/tags, and builds a vector index so you can ask natural‑language questions across your whole archive (RAG chat). It works with any OpenAI‑compatible model—including local Ollama servers—and ships as a Docker image with a simple web UI for configuration and manual processing.

  • Project
    NirDiamant/Controllable-RAG-Agent

    A controllable autonomous RAG agent that uses a deterministic graph and adaptive planning to solve complex, multi-hop questions from custom datasets while preventing hallucinations.

  • Project
    infiniflow/ragflow

    An open-source RAG engine that combines deep document understanding and agentic capabilities to create a high-fidelity context layer for LLMs using complex unstructured data.

  • Project
    rag-web-ui/rag-web-ui

    An intelligent dialogue system that allows users to build custom Q&A services by combining their own document knowledge bases with LLMs using Retrieval-Augmented Generation.