PageIndex: a vectorless RAG engine that replaces similarity search with reasoning-based tree retrieval for complex documents

What it solves

Traditional vector-based Retrieval-Augmented Generation (RAG) often relies on semantic similarity, which can fail to find truly relevant information in complex, professional documents. PageIndex addresses this by replacing similarity-based search with reasoning-based retrieval, ensuring that the system finds information based on relevance and contextual understanding rather than just similar-sounding text.

How it works

PageIndex transforms long documents into a hierarchical tree structure index (similar to a detailed table of contents) instead of breaking them into artificial chunks. It then uses LLMs to perform an agentic tree search, reasoning through the index to navigate to the most relevant sections of the document. This process mirrors how human experts read and navigate complex files, making the retrieval process traceable and explainable.

Who it’s for

It is designed for users dealing with long, complex professional documents such as financial reports, legal filings, regulatory documents, technical manuals, and academic textbooks.

Highlights

• Vectorless RAG: Eliminates the need for vector databases and chunking.
• Reasoning-Driven: Uses LLM-based tree search for context-aware retrieval.
• High Accuracy: Achieved 98.7% accuracy on the FinanceBench benchmark.
• Traceable Results: Grounded in explicit page and section references for better explainability.
• Flexible Deployment: Available as a self-hosted open-source tool, a cloud service, or an enterprise deployment.