Function calling with local LLMs using Ollama and llama.cpp. Qwen 2.5 7B matches GPT-4 accuracy for tool selection. Working code and agentic loop patterns.