VectifyAI/PageIndex

The problem

Vector RAG breaks documents into chunks, embeds them, and retrieves by semantic similarity. On long professional documents — 10-Ks, contracts, regulatory filings, technical manuals — similarity is a poor proxy for relevance: it surfaces passages that read alike but answer nothing, and misses passages that answer the question in different words. PageIndex takes a different route: build a hierarchical tree index of the document's structure, then have an LLM reason its way down the tree to the right section, the way a human expert flips to a chapter and reads.

What's different

Against vector databases such as Chroma or pgvector, the difference is the retrieval mechanism itself, not a tuning knob: no embeddings, no chunking, and results trace back to explicit document nodes, which makes citations and audit trails natural. Against just stuffing the whole PDF into a long-context model, PageIndex reads only the nodes its reasoning reaches, so per-query cost stays flat as documents grow — the README reports native PDF input costing 2.1× at 52 pages and 16.6× at 420 pages, and at 805 pages the document no longer fits at all. The trade-off is latency and model dependence: retrieval is sequential LLM reasoning, so a weak chat model hurts quality more than it would in a vector pipeline.

Who it's for

Teams building QA, due diligence, or research tools over dense, structured documents where being able to point at the source page matters: fintech and legaltech engineers, analysts, and anyone whose corpus is dozens to thousands of long PDFs rather than millions of short web pages. It is a poor fit for high-QPS, low-latency semantic search over short snippets, for multilingual or fuzzy recall where embeddings shine, or for teams unwilling to pay LLM tokens on every query.

Cost of adoption

Python package with an MIT license; you need your own LLM API key (OpenAI or, with the optional extra, Claude via a subscription). Indexing is a one-time cost — the README cites roughly $0.001 per page and 13 seconds to 4.5 minutes per document in their tests — but query cost scales with how much tree the chat model walks, so you should benchmark against your own corpus. The project is young (created April 2025) and the README is heavily oriented toward the hosted cloud and app; local SDK details, agent integration, and multi-document search live in external docs, so read those before committing.

Verdict

PageIndex is a genuinely different retrieval design rather than another vector-DB wrapper, and the reported FinanceBench accuracy plus the 39.6k-star count signal real interest: it gained about 3,867 stars in the month HotRepo tracked, reaching #20 on the monthly board. The caveats are maturity and economics — it is barely over a year old, Hacker News discussion is thin (one 192-point Show HN thread), and per-query LLM reasoning is priced very differently from a similarity lookup. Worth a serious prototype if your documents are long and your answers must be traceable; not yet a drop-in replacement for a mature vector stack.