isala404/Tera

Built in Rust with the Phi-2 language model for local inference, Tera supports multi-modal input including text, audio (via ffmpeg conversion), and document uploads for knowledge augmentation. The system uses a command-line interface with persistent memory management—users can teach it facts, ask questions against learned content, and manage a personalized knowledge base without external API dependencies. Architecture plans include vector indexing via SurrealDB and GPU acceleration support (CUDA/Metal) for improved performance on resource-constrained devices.