The latest release of vectifyai/pageindex, v0.4.3, is a Python library that uses a per-document hierarchical tree as its index instead of a vector store. Retrieval-augmented generation (RAG) is an approach that feeds retrieved evidence to a model to produce answers. Rather than turning documents into vectors and searching by similarity, this project has the LLM follow the tree structure to pick out relevant content. The license is MIT.
The README explains that the key is using the full context during retrieval — conversation history, domain knowledge and so on — so that results irrelevant to the question come back less often even when they resemble it. It puts local indexing cost at about $0.001 per page, and notes that a 1,000-page textbook comes to a little over $1 with gpt-5.6-luna. The package includes flash, classic and markdown indexing paths, an MCP bridge, and cloud and local APIs.
The performance comparison is based on a self-run benchmark in a separate repository. It evaluated 34 PDFs, 1,945 pages and 62 lookup questions, and reports that on spans that yield the same answer, native PDF input costs 2.1x at 52 pages and 16.6x at 420 pages, and does not fit in the context window at 805 pages. This is a self-evaluation, and because the benchmark lives in its own repository, it is hard to reproduce from the pageindex repository alone. Cost figures are therefore given, but they are hard to treat as independently verified performance evidence.