add KV-cache and move system promt to the config

server/models.py

@@ -68,16 +68,18 @@ class ModelManager:
     def _load_llm(self):
         from server.config import config
 
-        backend = config.get("llm", {}).get("backend", "local")
+        llm_config = config.get("llm", {})
+        backend = llm_config.get("backend", "local")
+        system_prompt = llm_config.get("system_prompt", "You are a helpful assistant.")
 
         if backend == "lmstudio":
             from server.llm import LMStudioEngine
 
-            lms = config.get("llm", {}).get("lmstudio", {})
+            lms = llm_config.get("lmstudio", {})
             url = lms.get("url", "http://host.docker.internal:1234")
             model = lms.get("model", "") or ""
             log.info(f"Using LM Studio backend at {url} (model={model or 'server default'})")
-            self.llm_engine = LMStudioEngine(url, model)
+            self.llm_engine = LMStudioEngine(url, model, system_prompt)
         else:
             log.info("Loading Qwen3-4B (GPTQ 4-bit)...")
             from transformers import AutoModelForCausalLM, AutoTokenizer
@@ -90,7 +92,8 @@ class ModelManager:
                 model_name,
                 device_map=device,
             )
-            self.llm_engine = LLMEngine(model, tokenizer)
+            max_cache_tokens = llm_config.get("max_cache_tokens", 4096)
+            self.llm_engine = LLMEngine(model, tokenizer, system_prompt, max_cache_tokens)
             log.info("Qwen3-4B-GPTQ-Int4 loaded (~2.5GB VRAM).")
 
     def _load_tts(self):