grecinto/tensortree

Most vector databases let you store chunks of text as numerical representations and then search for similar chunks by meaning. TensorTree is a Go library that adds a layer of structure on top of that: rather than storing everything in a flat pile of vectors, you organize knowledge into named categories, nest those categories into a tree hierarchy, and then retrieve content through either the category path or semantic similarity. The practical effect is that a query for something like "Root/Knowledge" can resolve to the more specific stored category "Root/Engineering" because the system compares the meaning of the path components rather than requiring an exact match. This means you do not have to know the precise category names your data was filed under in order to find it. The core technical approach stores items in a tree-indexed structure that the description says provides search speeds proportional to the logarithm of the number of items, meaning searches scale well as the database grows without requiring manual tuning or index rebuilds. Adding new content involves a single vectorization step that converts text to an embedding on demand, there is no separate indexing pipeline to manage. TensorTree is part of a larger system called SOP, and the repository includes a snapshot of the SOP memory subsystem, an architecture document, a conceptual article, and one working command-line demo you can run with a single Go command. The demo creates a small category hierarchy, inserts content, and shows the semantic path matching in action. The project is aimed at use cases where an AI application needs structured memory that is browsable and explainable: RAG pipelines, copilot tools, documentation search, and internal knowledge tools where the category structure itself carries meaning. The README does not state a license.