PageIndex

Search long documents like a human with Jev and PageIndex Learn more in our open-source repo: https://lnkd.in/ggwGMhXY

On X, TypeSafe AI's CTO and Co-founder, Erik Spock Gafni, shared an interesting idea: using Jev for search over a PageIndex tree. Inspired by that discussion, we open-sourced a small project showing how PageIndex can help scale Jev-style document search to long documents: https://lnkd.in/ggwGMhXY 🌲 PageIndex builds the hierarchy. It turns a long report into a tree of sections, each with a title, summary, and page range. 🎯 Jev searches the tree. Instead of choosing from hundreds of pages at once, it makes a small choice at each level: first the relevant section, then the page within that section. This approach avoids vector databases and embeddings while making Jev practical for much longer documents. We'd love your feedback, especially if you work with long financial, legal, or technical documents. #RAG #DocumentAI #LLM #FinancialAnalysis #PageIndex

Block references are now live for all PageIndex Cloud users. Every citation points to the exact block, not just the page. Docs → https://lnkd.in/gHGgEkpS

As we scale up PageIndex Cloud, we’re introducing a new, simpler pricing model: • $0.01 per page indexed • $0.001 per active page / month • Unlimited retrieval on active pages Simple usage-based billing. No monthly commitment. Get started with $10 in free credits.

An illustration of three ways to ask a document: • Vector DB: Split into chunks and search by vector similarity • PageIndex: Navigate a document tree and retrieve by node relevance • Full context: Put the whole document into the context— through an LLM provider’s file API

LLMs can solve the Navier–Stokes problem, but they’re still very poor at understanding PDFs. GPT-6 Astra scores just 32.2% on the GDP-PDF benchmark: https://lnkd.in/gz3_nW3u

One limitation of classic RAG is that retrieval quality is tightly coupled to the embedding model used at indexing time. When a better embedding model comes along, improving retrieval often means re-embedding and re-indexing the entire knowledge base. PageIndex works differently. It builds a persistent semantic representation of the knowledge once, while retrieval is performed by the LLM at query time. You can switch to the latest LLM without re-indexing. Retrieval quality automatically scales with the model.

A simple visualization of how PageIndex works. Index: generate a tree index for each document. Retrieve: agentically tree search with LLM Learn more: https://lnkd.in/gjhPKMxi

Introducing PageIndex Flash ⚡ Our fastest tree-indexing engine for long, text-based PDFs. → Index 1,000 pages for under $1 → Reduce query costs by up to 16× → Run locally with your own LLM API keys Now in the new PageIndex SDK. Read the full update: https://lnkd.in/guF6aX4C

PageIndex is #1 on Product Hunt. Thank you to everyone who tried it, voted, and sent feedback. Bring in your entire document set, ask your hardest question, then click any citation to jump to the exact highlighted source line and verify it in seconds. #ProductHunt #productoftheday