Extend Debuts Jevbox to Replace Vector Search With Trees

AI infrastructure startup Extend has released Jevbox, an open-source, permission-aware document library designed to bypass traditional vector search methods. Instead of relying on embeddings and vector databases, Jevbox organizes uploaded documents into a hierarchical tree structure of categories, documents, and sections. A specialized decision model named Jev manages document filing and routes search queries through this tree using a hierarchical beam search.

During a query, the Jev model scores candidate categories, documents, and sections, keeping several promising paths open at each level. This narrowing process allows the system to locate relevant information without loading an entire corpus into a single context window. After traversing the tree, Jevbox performs a second pass to score individual passages for usefulness, filtering out irrelevant text that may have been caught along a strong search path. Every generated answer includes precise page- and section-level citations.

Built as a TypeScript application, Jevbox integrates Postgres, SpiceDB, pg-boss, and the company's own Extend Parse engine. It supports more than 20 chat model providers and includes a Model Context Protocol (MCP) server, allowing AI agents to automatically inherit a user's specific permission scope. Security is tightly integrated; if a user's access to a source document is revoked, any dependent conversations are immediately blocked.

For developers and system administrators, Jevbox simplifies deployment by offering self-hosting options via Docker Compose, a Helm chart for Kubernetes, or a one-click deployment on Render. By eliminating the need for complex vector databases and embedding pipelines, the library provides practitioners with a more deterministic, auditable, and secure alternative for retrieval-augmented generation tasks.