How to Use PageIndex for Vectorless RAG on Complex Documents

PageIndex argues that similarity is not the same as relevance, and it offers a vectorless, reasoning based approach to retrieval that builds a hierarchical page index and uses LLM-driven tree search to find the most relevant sections. The claim is bold, it reports 98.7% on FinanceBench, and the repo is worth a close look if you work with complex documents.

PageIndex is designed for single or long-document retrieval scenarios where structure and reasoning matter more than nearest neighbor similarity.

What it is

PageIndex (VectifyAI) is a vectorless retrieval system for retrieval augmented generation workflows. Instead of embedding and chunking, it builds a hierarchical table of contents, then uses an LLM to reason over that index with a tree search strategy, much like how an expert would consult a long document. While Chroma Context-1 focuses on local agentic search with embeddings, PageIndex takes the opposite route by eliminating vectors entirely and relying on structured reasoning.

How it works

At a high level PageIndex performs retrieval in two steps:

1. Generate a tree style table of contents from the document, splitting content into meaningful page or section nodes.
2. Run a reasoning based tree search that expands promising nodes, uses the LLM to score relevance, and prunes branches to focus on the most relevant passages.

This retrieval approach is a strong alternative for developers who have tried BRAG and want a different paradigm that avoids embedding pipelines, especially for single-document scenarios with deep structural hierarchy.

# quick start
git clone https://github.com/VectifyAI/PageIndex.git
cd PageIndex
# follow README for example indexing and retrieval commands

Feature	Notes
Vectorless retrieval	No embeddings, no vector DB, no chunking
Hierarchical index	Tree structure mirrors document layout and semantics
Reasoning search	LLM guided tree search to find relevant sections
Use case	Long reports, manuals, finance docs, legal briefs

Pair PageIndex with a strong reasoning model for final answer synthesis, PageIndex focuses on relevance selection rather than answer generation.

Want to dive deeper? Here is another post you shouldnt miss: How to Use OpenWork as a Desktop Interface for DeepAgents and How to Build Interactive UIs for Agents with MCP Apps.

Pros and cons

Pros

• Avoids vector DB complexity and indexing costs
• Better alignment with true relevance for long, structured documents
• Compact index, interpretable tree search, and strong single document performance

Cons

• May not scale as well for large multi document corpora without additional orchestration
• Depends on the reasoning model to score and expand nodes accurately
• Requires work to integrate into existing RAG pipelines that expect embeddings

Try it locally

1. Clone the repo and read the examples to build a tree index from sample documents.
2. Run the supplied retrieval scripts and compare results against a vector baseline on representative tasks.

This approach shines on single long documents, but multi document setups may need additional design. Validate on your dataset before committing to a full migration away from vector search.

Project link

• https://github.com/VectifyAI/PageIndex

Here are what people are saying:

“Multi document wise this approach does not work. Single document – yes” @probatum_

If you enjoy articles about top GitHub repositories like this, don’t forget to subscribe to Technolati.com.