Vector RAG quietly assumes semantic similarity is a good stand-in for relevance. On a 200-page 10-K or a dense regulatory filing, that assumption breaks down — the passage that sounds closest to your question is often not the one that answers it. PageIndex takes the opposite bet: drop embeddings entirely and let an LLM reason its way through a document the way a person flips to the right chapter.
What Sets It Apart
- A table-of-contents tree, not a pile of chunks. Each document is parsed into a hierarchical index of sections and subsections, so retrieval becomes agentic tree search — the model walks the structure instead of scoring isolated fragments.
- No embeddings, no vector database, no chunk size to tune. That removes a whole class of failure modes (bad chunk boundaries, stale indexes) and an entire piece of infrastructure to operate.
- Answers you can audit. Because retrieval lands on explicit sections and page numbers, every result traces back to where it came from — which matters when the document is a contract or a financial report.
- Relevance over similarity, with receipts. It reports 98.7% accuracy on FinanceBench, the kind of long-document financial QA where pure similarity search tends to struggle.
Great Fit If...
You work with long, structured professional documents — SEC filings, legal contracts, technical manuals, textbooks — where being able to explain why a passage was retrieved matters as much as the answer itself. Look elsewhere if your corpus is millions of short, unstructured snippets, or you need sub-second lookups over a huge collection: agentic tree search trades raw latency and per-query cost for reasoning depth, and shines most when each document has real internal structure to exploit.
