Siddhesh2377/ToolNeuron

Implements a hybrid RAG pipeline combining FTS4 BM25, vector search, and Reciprocal Rank Fusion across PDFs, docs, and chat history, with grammar-constrained JSON for function calling. Built on llama.cpp for inference, LocalDream for Stable Diffusion, and a custom Unified Memory System (UMS) for efficient binary storage, with AES-256-GCM encryption backed by Android KeyStore and hardware TEE where available. Includes 7 callable plugins (web search, file ops, calculator) and adaptive hardware tuning that auto-detects CPU topology to optimize thread allocation and context windows.

272 stars.