bigmacfive/turbo-graph

Turbo-graph is a search library built for AI applications that need to find relevant information quickly, a technique often called retrieval augmented generation, or RAG. It is a fork of an existing project called turbovec, which already does fast vector search, the kind of search that finds similar pieces of text based on meaning rather than exact words. Turbo-graph keeps that core search engine and adds a layer on top for handling more complicated, constrained queries. Many real world search needs are not just find the top matches. They often need to combine several conditions at once, such as only searching within a certain customer's data, filtering by tags or source, limiting to a time window, following connections in a graph of related items, and blending in traditional keyword search alongside vector search. Rebuilding that combined view from scratch on every request in ordinary Python code is slow. Turbo-graph solves this by keeping graph memory, indexed views by tag, source, and time, and cached compiled filters, so repeated constrained searches do not have to redo the same work. The project explains that plain turbovec is enough when filters are light and cheap to build. Turbo-graph is meant for cases where constraints are central to the product, such as multi tenant systems where the same filtered view of the data gets queried again and again. It also adds reranking that considers graph relationships, hybrid scoring that blends keyword and vector results, and detailed reports that explain why results were chosen and how much was served from cache. The core is written in Rust for speed, with Python bindings so it can be used from typical AI application code, and the Python side releases its lock during long operations so multiple requests can run at once. The README includes benchmark comparisons against FAISS, a well known vector search library, showing competitive recall and speed, particularly on ARM processors.