Jevbox routes document search through a tree

Extend has released Jevbox, an open-source document library that automatically files uploads, answers questions with page- and section-level citations, and checks source permissions during retrieval. Its index is a hierarchy of categories, documents, and sections, with no embeddings or vector database.

The team behind Extend Parse introduced the project in a launch post. Jevbox packages ingestion, organization, retrieval, chat, authorization, and agent access into one TypeScript application. A decision model called Jev handles document filing and search routing.

The tree becomes the index

Conventional retrieval-augmented generation systems split documents into chunks, convert those chunks into numeric embeddings, and search a vector database for nearby matches. Jevbox builds a navigable document tree and searches it with hierarchical beam search. Jev scores candidate categories, documents, and sections while retaining several promising paths at each level.

After traversal, Jevbox scores source passages independently for usefulness before generating an answer. This second pass helps filter passages that appeared on a strong route but do not support the specific question.

Jev assigns calibrated probabilities to multiple-choice options. A related search project demonstrates the mechanism with questions such as “Which page answers this query?” Jevbox applies the same pattern repeatedly, narrowing the available choices at each layer so the model does not need to read the entire corpus in one context window.