PageIndex
worthwhileTree-based reasoning retrieval that beats vector search on structured documents — but limited to well-formatted PDFs and GPT-4-class models.
What it is
A vectorless RAG system that builds a hierarchical tree index (like a table of contents) from long PDFs, then uses LLM reasoning via tree search to retrieve relevant sections. No embeddings, no chunks. Achieves 98.7% on FinanceBench.
How it differs from vanilla Claude
Vanilla Claude can be given a full document and asked to reason over it, but for 100+ page PDFs, context limits and cost make that impractical. A smart agent could be prompted to create a ToC-like summary and search iteratively, but PageIndex automates this with deterministic tree building and structured search, saving significant prompt engineering and token waste.
Skill, plugin, or workflow shift?
PDF parsing quality (standard vs enhanced OCR), tree depth parameters, LLM model choice, optional page images for vision-based RAG, MCP/API integration. Overhead: per-document tree generation costs tokens (GPT-4) and takes time.
Devil's advocate — is this just complexity?
1) The core insight is a clever prompt pattern — ask an LLM to generate a tree index and search it. A capable agent with tool use can replicate this ad-hoc. 2) Tree building has a single-pass structure; long or poorly formatted documents may fail silently. 3) Claims 'no chunking' but tree nodes are effectively semantic chunks with boundaries defined by document structure, not size. 4) Benchmark on FinanceBench is impressive but narrow (financial tables); generalizability to diverse PDFs is unproven. 5) The open-source self-host version uses basic PDF parsing — cloud service with 'enhanced OCR' is the real product, making the repo a demo.
What would make it better
Open-source the enhanced OCR pipeline; add support for more formats (Word, HTML, scanned PDFs); provide incremental/streaming tree updates for large corpora; reduce token cost by caching tree structures; add open-source reference implementation that works with local LLMs (e.g., Llama) to avoid vendor lock-in.
The honest case for it
PageIndex is not just a wrapper — it concretely solves a fundamental failure of vector RAG: similarity ≠ relevance. For structured long documents (SEC filings, legal contracts, manuals), the hierarchical tree + reasoning search provides traceable, high-precision retrieval that dense embeddings cannot match. The tree approach is deterministic and explainable, and the 98.7% FinanceBench score (with open evaluation) shows real-world impact.
Who it's for
Audience fit
Depth and leverage for a technical engineer who wants to understand it and level up their workflow — not just offload work.
Value for someone who wants a more capable tool without the technical depth — accessible, does-it-for-you.
Strong for engineers who build RAG pipelines and need explainable, structure-aware retrieval; vibe coders will hit complexity with tree search and agent setup.