What is PageIndex : Vectorless RAG?

PageIndex is an open-source framework that reimagines Retrieval-Augmented Generation (RAG) by moving away from traditional vector databases and similarity searches. Instead, it introduces a **"vectorless" and "reasoning-based"** approach, where a Large Language Model (LLM) navigates a document's structure, much like a human would use a table of contents to find precise information .

### 🔍 Why "Vectorless"?

Traditional RAG splits documents into chunks, converts them into mathematical vectors (embeddings), and retrieves chunks based on semantic similarity to your query. PageIndex argues that **similarity is not the same as relevance**, especially for complex, professional documents like financial reports or legal contracts . For example, a simple similarity search might return every page mentioning "EBITDA," but it cannot reason about which specific section contains the exact calculation or context you need .

### 🧠 How It Works: Reasoning Over Structure

PageIndex's core idea is to treat document retrieval as a navigation problem rather than a search problem . It works in two main stages:

1.  **Build a Hierarchical Index:** It processes a document (like a PDF) to create a JSON-based "tree structure," similar to a highly detailed and LLM-friendly table of contents . Each node in this tree represents a logical section (e.g., a chapter or subsection) and contains a summary, its location (page numbers), and links to its sub-sections .

    ```json

    {

      "title": "Financial Stability",

      "node_id": "0006",

      "start_index": 21,

      "end_index": 22,

      "summary": "The Federal Reserve ...",

      "nodes": [ ... ]

    }

    ```

2.  **Perform Agentic Tree Search:** When you ask a question, the LLM doesn't perform a database lookup. Instead, it acts as an agent, using the index to reason about where to look . It starts at the top level, reads section summaries, and decides which branch to "descend" into, iteratively narrowing its focus until it finds the most relevant section .

### ✨ Key Advantages and Performance

This reasoning-based method offers several significant benefits over traditional vector RAG :

| Feature | PageIndex (Reasoning-Based RAG) | Traditional Vector RAG |

| :--- | :--- | :--- |

| **Retrieval Logic** | **Reasoning & Inference:** Thinks about where the answer is likely to be (e.g., "This will be in Appendix G"). | **Similarity:** Finds text that is semantically similar to the query. |

| **Data Structure** | **Hierarchical Tree:** Preserves natural document sections (chapters, sections). | **Fixed Chunks:** Arbitrarily splits text into chunks, often breaking context. |

| **Key Capability** | **Follows References:** Can navigate internal links like "see Appendix G" to find information. | **Misses References:** Often fails to follow cross-references as they are not similar to the original query. |

| **Context Usage** | **Dynamic:** Retrieves coherent sections and can fetch more context if needed. | **Static:** Always retrieves the same top-k chunks, regardless of context. |

| **Transparency** | **High:** Provides a traceable "path" of reasoning (e.g., went to Section 4, then Appendix B). | **Low:** Retrieval is a "black box" of similarity scores. |

This approach has shown impressive results. A financial analysis system built on PageIndex, called **Mafin 2.5**, achieved **98.7% accuracy** on the FinanceBench benchmark, a well-known test for financial document Q&A .

### 🚀 Getting Started with PageIndex

You can use PageIndex in several ways:

*   **Self-Hosted (Open-Source):** You can run the framework locally. The [GitHub repository](https://github.com/VectifyAI/PageIndex) provides the code and a quickstart guide to index your own PDFs .

*   **PageIndex Chat:** A ChatGPT-style web application where you can upload and chat with long documents to experience the system firsthand .

*   **MCP Integration:** PageIndex can be integrated with AI applications like Claude Desktop or Cursor via the Model Context Protocol (MCP) .

In short, PageIndex offers a compelling alternative for complex, high-stakes document analysis, trading the speed of vector search for the accuracy and explainability of structured, reasoning-based retrieval.

references:

 https://github.com/VectifyAI/PageIndex