Read specialized documents with LLM reasoning without a vector database
When finding key information in long reports or legal documents, PageIndex offers an experience where the system follows the table of contents like an expert, reading only the necessary parts instead of simply finding similar words. It is a search tool that organizes documents into a tree structure and allows an LLM to follow that structure to find answers. This enables accurate, context-aware search without a vector database or text chunking.
Beyond the limits of vector search
Existing vector-based search finds information based on word similarity, but in specialized documents, similar expressions are not necessarily relevant content. Inspired by AlphaGo, PageIndex indexes documents into a hierarchical tree structure and is designed for the LLM to explore this tree. This makes the search process transparent and allows you to track the basis for the answer. It is particularly effective for documents requiring complex context, such as financial reports, legal documents, and technical manuals.
Recent updates and key features
In October 2026, the Page Search API was added, which finds the page that resolves a question in about one second and reduces token costs by 97% for document Q&A by sending only that page to the LLM.
- Page Search API Quickly identifies the page that answers a question to minimize LLM input tokens.
- PageIndex SDK Handles everything from indexing to chat in a single client for local and cloud environments.
- PageIndex Flash Increases the tree indexing speed for text-based PDFs and is used by default in local mode.
Performance and cost comparison
On the FinanceBench benchmark, PageIndex recorded 98.7% accuracy, significantly outperforming vector-based RAG. The local indexing cost is approximately $0.001 per page, meaning it costs just over $1 and takes a few minutes to index a 1,000-page textbook. Compared to feeding the entire PDF to the model each time, you can get the same answer at 2.1 times less cost for a 52-page document and 16.6 times less cost for a 420-page document. For documents of 805 pages or more, the entire document cannot fit in the context window, making the PageIndex approach essential.
Differences between cloud and local
The open-source version is suitable for text-centric PDFs and local workflows, handling indexing and storage on the user's machine. PageIndex Cloud supports scanned or image-centric documents and provides OCR, image understanding, metadata management, folder structure, and MCP server functionality. To switch to cloud mode, you only need to set the PageIndex API key, and the chat and search layers are compatible with your existing model provider. For handling large document collections, you can utilize the cloud-exclusive PageIndex File System.
Things to check before applying
PageIndex is written in Python and can be installed with the pip install -U pageindex command. The default model is sufficient for indexing, but using the best possible model for the chat model responsible for search is beneficial for accuracy. If you have a Claude subscription, you can install pageindex[claude] to use the Claude model without an API key. It is suitable for environments that do not require a vector database or specialized document processing tasks where tracing the basis of search results is important. However, documents containing images or scans are only supported in the cloud version, so you must choose between local or cloud mode depending on your usage environment.