How PageIndex Redefines RAG: Unpacking Its Structural Advantage Over Traditional Vector Retrieval

Core Idea: Understand Structure Before Searching

PageIndex is a retrieval‑augmented generation (RAG) method that does not rely on embeddings, chunking, or vector databases. Instead, it constructs a hierarchical table of contents (TOC) from the document and uses a large language model (LLM) to reason over that structure, first locating the most relevant chapter and then generating a precise, referenced answer.

Stage One – Index Creation

1. Structure Detection

The LLM reads the source (e.g., the script of the classic film Sholay ) and detects natural boundaries such as scene titles, character introductions, act separators, and key narrative turns. The detection relies on narrative hierarchy rather than fixed‑size chunks.

Dark root node – represents the whole document

Blue – main story segments

Red – Gabbar‑related storyline

Purple – key event nodes

Gold – concrete factual events

2. Hierarchical Mapping

PageIndex builds a tree whose root is the film title. First‑level branches may include Prologue, Recruitment of Veeru and Jai, Life in Ramgarh, Gabbar’s reign of terror, and the Final battle. Each branch can contain further child nodes.

For example, the node “Gabbar's Den” is summarized as: “This section introduces Gabbar Singh, the line ‘Kitne aadmi the’, and the punishment of his henchmen.”

Each node stores:

title

nodeId

summary

child nodes

The LLM writes a concise semantic description for every node; these summaries become the retrieval signals during the query phase.

Stage Two – Query Phase

Assume a user asks: Why did Thakur lose his arms?

No full script is sent to the model, and no embeddings are generated. The LLM receives only three items: the user question, the hierarchical JSON tree, and the summary of each node.

Step 1: Structural Search

The LLM scans the tree, sees nodes such as “Thakur family massacre”, “Gabbar’s revenge”, and “Life in Ramgarh”, and reasons that the answer likely resides in sections involving Gabbar and Thakur’s injury. This is logical reasoning, not vector similarity.

Step 2: Focused Exploration

PageIndex then retrieves the original text of only those specific nodes, typically 2–3 highly relevant passages, instead of scanning the entire 50‑page script.

Step 3: Final Answer Generation

The LLM reads the retrieved snippets and produces the answer with a citation:

Thakur lost his arms because Gabbar Singh, seeking revenge for Thakur’s earlier arrest, cut them off. (nodeId: massacre-thakur-family) The retrieval process is explainable and traceable.

Differences Between PageIndex and Traditional RAG

Traditional vector‑based RAG would return any passage with semantically similar words, such as a fight scene where Jai uses his arm, or unrelated mentions of “hand”. It matches on “atmosphere” rather than narrative relevance.

PageIndex avoids this by having explicit summaries like “This section describes how Gabbar attacks Thakur’s family and cuts off his arms,” allowing the LLM to navigate directly to the factual answer instead of guessing.

Why PageIndex Works

It separates two cognitive tasks: navigation (identifying where the answer should be) and extraction (reading the identified chapter and generating the answer). This mirrors how humans read long texts—jumping straight to the relevant chapter instead of flipping through every page.

Applicable Scenarios

PageIndex excels with documents where structural hierarchy outweighs surface similarity, such as financial reports, legal contracts, policy documents, regulatory filings, academic papers, and long narrative content.

Conclusion

Traditional RAG assumes relevance equals semantic similarity; PageIndex assumes relevance equals structured reasoning. The difference may seem subtle but profoundly impacts retrieval quality for long, hierarchical documents. Rather than building a better search engine, PageIndex draws a navigation map that lets the LLM think first and then read.

by Vishal Mysore