RAG inspired by AlphaGo: tree search over documents

What it does

PageIndex ingests long documents—financial reports, legal filings, textbooks—and builds a hierarchical tree index resembling a table of contents. Instead of chunking text and querying a vector database for semantic similarity, it uses an LLM to navigate the tree and reason about which sections are actually relevant to your query. The open-source repo handles PDF and Markdown inputs, and there is a cloud tier with enhanced OCR for messier documents.

The interesting bit

The project explicitly rejects the “similarity equals relevance” assumption that underpins most RAG pipelines. By treating retrieval as a tree-search reasoning task—borrowing the AlphaGo analogy—it aims for explainability: you get page and section references back instead of opaque embedding distances. There is even a vision-based mode that works directly on page images without OCR.

Key highlights

• Vectorless retrieval: no embeddings, no chunking, no vector DB
• Hierarchical tree index generated from document structure with summaries and node IDs
• Reports 98.7% accuracy on FinanceBench via its reasoning-based approach
• Self-hostable with standard PDF parsing, or available as a cloud API/MCP with enhanced OCR
• Scales to millions of documents via a file-level PageIndex File System layer

Caveats

• The open-source version relies on standard PDF parsing; the README notes that complex PDFs may need the cloud service’s enhanced OCR
• Markdown support requires properly formatted heading hierarchies and is explicitly not recommended for PDF-converted Markdown files
• Tree generation and retrieval are LLM-driven, so latency and cost depend on model calls rather than a fast vector lookup

Verdict

Worth a look if you are building RAG over long, structured professional documents and are skeptical of embedding-based retrieval. Skip it if you need millisecond-scale retrieval on massive unstructured text corpora where vector search already works fine.