PageIndex
Retrieval with no vector database: build the document a table of contents and let the model read its way in
What is PageIndex?
Vector search finds passages that resemble the question, and resemblance is not relevance — on a long regulatory filing the paragraph that answers a question often shares almost no vocabulary with it. PageIndex drops the vector store. It builds a tree of the document's real structure, section by section, and then has a model walk that tree the way a person would use a table of contents: narrowing down, opening the part that should hold the answer, reading it. Nothing is chunked, and every answer points at the section it came from. The trade is per-query cost, since retrieval now involves reasoning rather than a lookup, and the design targets long individual documents rather than a corpus of short ones.
What can you do with PageIndex?
- Retrieve by reasoning, not by resemblance — A model decides which section should hold the answer and opens it, so a passage that answers the question in different words is still found.
- Keep natural sections instead of fixed chunks — The unit of retrieval is a real section of the document, so a definition and the clause that depends on it do not end up in separate fragments.
- Show where an answer came from — Every result names the part of the tree it was read from, which is the difference between an answer you can check and one you have to trust.
- Skip the vector database entirely — There is no index to build, embed, tune or keep in sync — a document is submitted and the tree is what gets stored.
- Build the tree in seconds when structure is obvious — A faster mode extracts the structure from the document itself rather than asking a model to infer it, for material with a clean table of contents.
- Run it locally or against the hosted service — The same client indexes and retrieves on your own machine with your own model key, or points at the maintainers' cloud without a code change.
Before you choose PageIndex
- Retrieval itself costs model calls, so each query is slower and more expensive than a vector lookup — the calculation changes if you serve many cheap queries rather than a few expensive ones.
- It is aimed at long, well-structured professional documents; a corpus of short unstructured pages has little tree to reason over, and the vendor's headline accuracy figure comes from a financial-document benchmark.
Star history
27 Aug to 30 Oct · +4.6k
Frequently asked questions
Is PageIndex free for commercial use?
PageIndex is released under the MIT licence — OSI-approved open source, which permits commercial use.
How can PageIndex be deployed?
PageIndex is available as Runs locally / Self-hosted / Managed cloud.