Top AI Product

So here’s something I didn’t expect to say in 2026: maybe we’ve been doing RAG wrong this whole time.

[PageIndex](https://github.com/VectifyAI/PageIndex) popped up on GitHub Trending today with over 16,000 stars, and after spending some time with it, I get why people are excited. Built by [VectifyAI](https://github.com/vectifyai), it takes a completely different approach to retrieval-augmented generation — one that throws out vector embeddings, chunking, and vector databases altogether. Yeah, all of it.

Instead of converting your documents into vectors and hoping cosine similarity finds the right passage, PageIndex builds a hierarchical tree index from your documents — basically a smart, structured table of contents. Then it uses the LLM itself to reason through that tree, navigating section by section the way you or I would flip through a long financial report looking for a specific number. It reads the structure, picks the relevant section, digs deeper, checks if it has enough context, and keeps going until it finds what it needs. It’s surprisingly intuitive once you see it in action.

The numbers back it up too. On FinanceBench, PageIndex hit [98.7% accuracy](https://pageindex.ai/blog/pageindex-intro) — compared to roughly 50% for traditional vector-based RAG on the same benchmark. That’s not a marginal improvement; that’s a different league. The reason is straightforward: financial documents are full of terms that look similar to an embedding model but mean very different things in context. Vector search struggles there. Reasoning doesn’t.

What really caught my attention is the MCP integration — there’s a separate [pageindex-mcp](https://github.com/VectifyAI/pageindex-mcp) server that lets tools like Claude and Cursor reason over document structure directly. No OCR pipeline needed either, since it can work straight from PDF images through its vision-based retrieval mode.

The [Hacker News thread](https://news.ycombinator.com/item?id=43548690) has some good back-and-forth about trade-offs. The honest criticism is that reasoning-based retrieval is slower and more expensive per query than a vector lookup. That’s fair. But if you’ve ever spent days debugging why your RAG pipeline keeps pulling the wrong chunk from a 200-page SEC filing, you might happily trade a bit of latency for actually getting the right answer.

Is this the end of vector databases for document retrieval? Probably not for every use case. But for complex, domain-heavy documents where accuracy matters more than milliseconds, PageIndex makes a pretty compelling argument that we’ve been overcomplicating things.