Achieving Vectorless Accuracy with Proxy-Pointer RAG

The recent launch of PageIndex marks a broader shift in AI architecture towards "Vectorless RAG" or "Reasoning-Based Retrieval." Instead of the standard method of splitting documents into random chunks and searching through mathematical similarity, PageIndex builds a "Smart Table of Contents" — a hierarchical tree that allows language models to navigate documents like a human expert.

Numerous blogs, including one from Microsoft, outline the working principles of this technology, achieving 98.7% accuracy on a financial benchmark while emphasizing that Vectorless RAG is best suited for deep-dive queries on complex structured or semi-structured documents, such as financial statements. This is due to the fact that PageIndex’s tree-based approach cannot practically scale to multi-document scenarios.

The primary reason is that building the hierarchical tree index is a slow and costly process that requires an LLM. In contrast, creating a vector index is fast and inexpensive, and the retrieval step uses an LLM only once during the response synthesis.

PageIndex demonstrates its accuracy due to three architectural advantages: structural navigation rather than pattern matching, contiguous context extraction, and the ability to work with complete sections, which avoids the need for chunking.

In this article, I will walk through a real use case on a large, complex document to build Proxy-Pointer RAG — an ingestion and retrieval pipeline that achieves high accuracy while maintaining low latency and cost, making it scalable across enterprise databases.