PageIndex Says You Don't Need Vector Databases for RAG

PageIndex gained 943 stars in a single day, hitting 29.9K total. The pitch: build hierarchical tree indexes from documents and use LLM reasoning to navigate them. No vector database. No chunking. No embeddings.

The approach mimics how a human expert navigates complex documents. Instead of splitting text into chunks, embedding them, and doing cosine similarity search, PageIndex builds a tree structure representing the document's logical hierarchy. At retrieval time, the LLM reasons about which branches to follow, producing context-aware, explainable retrieval.

This challenges a fundamental assumption I've held about RAG architecture. I built Rayni on pgvector with chunk-and-embed retrieval. It works. But chunking always felt like a lossy compression step, something we tolerated because vector similarity was the best tool we had. PageIndex suggests another path entirely.

The project already supports OpenAI Agents SDK integration, a FileSystem layer for corpus-scale retrieval, and vision-based RAG for documents with visual content. The star velocity suggests real demand, not just curiosity. Builders are tired of tuning chunk sizes, overlap parameters, and embedding models and getting mediocre retrieval quality.

I'm not ready to say vector databases are dead. For many workloads, embeddings plus approximate nearest neighbor search is still the right choice, especially at scale with sub-second latency across millions of documents. But PageIndex opens a design space the RAG community has been ignoring. Reasoning-based retrieval trades compute at query time for precision, and with inference costs dropping, that trade-off gets more attractive every quarter.

What builders should try: if your RAG pipeline struggles with complex, hierarchically structured documents (legal contracts, technical specs, codebases), test PageIndex against your current vector-based approach. The comparison will tell you whether your retrieval problems are embedding problems or structure problems.