Fix: skip Qwen3 thinking mode by pre-filling empty think block + reduce MAX_TOKENS to 150

App: load pre-built index only, no Docling dependency at runtime

App: load pre-built index instantly, skip Docling at runtime — fast startup"

Speed fix: single-pass agent with minimal prompt, MAX_TOKENS=200, shorter context

Add main app: Agentic RAG with Qwen3 1.7B + tool calling + BGE embeddings