PageIndex Hits 31,709 Stars. No Vectors, No Embeddings, 98.7% Accuracy on FinanceBench.

What if your RAG pipeline doesn't need a vector database at all?

VectifyAI's PageIndex eliminates vector databases entirely from document retrieval. Instead of chunking documents, generating embeddings, and running approximate nearest-neighbor search, it builds a hierarchical Table of Contents tree from the document structure and uses LLM reasoning to navigate to the most relevant section. The approach is inspired by AlphaGo's tree search. The result: 98.7% accuracy on FinanceBench via the Mafin 2.5 financial analysis system, significantly outperforming traditional vector-based RAG.

No chunking. No embeddings. No vector DB. Just document structure and LLM reasoning with full page and section traceability.

I've spent months building and maintaining vector RAG pipelines. pgvector in Rayni, embedding generation, chunk size tuning, retrieval quality debugging. The entire infrastructure is non-trivial. If a reasoning-based approach can match or beat vector retrieval on a serious financial benchmark, that's not just an academic curiosity. It's a potential infrastructure elimination.

The timing is interesting. Milvus just shipped a 3.0 release candidate with data lake architecture, external collections, and entity-level TTL. The vector database ecosystem is maturing. And a project with 31K stars is saying you might not need any of it.

I'm not ready to rip out my vector pipelines yet. FinanceBench is one benchmark, and document-structured financial reports are an ideal case for hierarchical navigation. I don't know how well this works on messy, unstructured content where there's no clean ToC to build. But for anyone working with structured documents, PDFs, technical specs, legal filings, financial reports, this is a "try it this weekend" story. If it works for your use case, you just eliminated an entire infrastructure dependency.