Update expert mode LLM to mistral-medium-3.5-128b with reasoning effort

agents/nvidia_llm.py

@@ -29,7 +29,7 @@ _QUEUE_MAX_SIZE    = int(os.getenv("NVIDIA_QUEUE_MAX", "8"))
 _REQUEST_TIMEOUT_S = int(os.getenv("NVIDIA_LLM_TIMEOUT", "600"))
 
 # DEFAULT_MODEL = "google/diffusiongemma-26b-a4b-it"
-DEFAULT_MODEL = "sarvamai/sarvam-m"
+DEFAULT_MODEL = "mistralai/mistral-medium-3.5-128b"
 
 # ── Flask app ─────────────────────────────────────────────────────────────────
 app = Flask(__name__)
@@ -61,12 +61,14 @@ def _run_inference(data: dict) -> dict:
         chat_template_kwargs = None
         temperature = float(data.get("temperature", 1.0))
         top_p = float(data.get("top_p", 0.95))
+        reasoning_effort = None
     else:
         model_name = data.get("model", DEFAULT_MODEL)
         max_tokens = int(data.get("max_tokens", 16384))
         chat_template_kwargs = None
-        temperature = float(data.get("temperature", 0.5))
-        top_p = float(data.get("top_p", 1.0))
+        temperature = float(data.get("temperature", 0.70))
+        top_p = float(data.get("top_p", 1.00))
+        reasoning_effort = "high"
 
     invoke_url = "https://integrate.api.nvidia.com/v1/chat/completions"
     headers = {
@@ -99,6 +101,8 @@ def _run_inference(data: dict) -> dict:
             }
             if chat_template_kwargs:
                 payload["chat_template_kwargs"] = chat_template_kwargs
+            if reasoning_effort:
+                payload["reasoning_effort"] = reasoning_effort
 
             response = requests.post(invoke_url, headers=headers, json=payload)
             response.raise_for_status()