PageIndex

Skipping embeddings entirely, PageIndex retrieves the way a person uses a table of contents: Vectify AI's MIT-licensed system parses a document into a hierarchical tree index, then lets an LLM navigate that tree with multi-step reasoning and tree search to locate relevant sections, with no vector database, chunking strategy, or similarity threshold anywhere in the pipeline. The approach suits long, structured material like financial reports, legal filings, and technical manuals, where the team reports 98.7 percent accuracy on the FinanceBench benchmark. A pip install pageindex gets the open source SDK, which runs in local mode against a caller-supplied LLM API key or in cloud mode where hosted OCR, tree building, and retrieval sit behind a PageIndex key from the developer dashboard. Native integrations cover OpenAI, Anthropic, the Claude Agent SDK, LangChain, and PydanticAI, and an MCP server exposes retrieval to Claude and Cursor. Documentation lives at docs.pageindex.ai. Around 35,000 GitHub stars later, it anchors the vectorless side of the ongoing RAG architecture debate.