Fix NameError for get_lazy_llm and enable async trend decoding

api/main.py

@@ -54,11 +54,11 @@ EMBEDDING_MODEL_PATH = os.path.join(ROOT_DIR, 'embedding_model')
 DB_PATH = os.path.join(os.environ.get("WRITABLE_DIR", "/tmp"), "vector_db_persistent")
 
 # --- Global Instances ---
-_llm_instance: Optional[Llama] = None
-_vector_store: Optional[Any] = None
-_ai_strategist: Optional[AIStrategist] = None
-_creative_director: CreativeDirector | None = None 
-_support_agent: Optional[SupportAgent] = None
+_llm_instance = None 
+_vector_store = None
+_ai_strategist = None
+_creative_director = None 
+_support_agent = None
 _budget_predictor = None
 _influencer_matcher = None
 _performance_predictor = None
@@ -69,11 +69,31 @@ _likes_predictor = None
 _comments_predictor = None
 _revenue_forecaster = None
 _performance_scorer = None
-_community_brain: Optional[CommunityBrain] = None 
+_community_brain = None
 
 def to_snake(name: str) -> str:
     return re.sub(r'(?<!^)(?=[A-Z])', '_', name).lower()
 
+def get_lazy_llm():
+    """Wakes up the AI model only when it's needed."""
+    global _llm_instance
+    if _llm_instance:
+        return _llm_instance
+    
+    print("⏳ Awakening AI Brain (Loading LLM on-demand)...")
+    try:
+        from llama_cpp import Llama
+        if not os.path.exists(LLAMA_MODEL_PATH):
+            print("   - Downloading model (first-time only)...")
+            hf_hub_download(repo_id=MODEL_REPO, filename=MODEL_FILENAME, local_dir=MODEL_SAVE_DIRECTORY)
+        
+        _llm_instance = Llama(model_path=LLAMA_MODEL_PATH, n_ctx=1024, n_threads=2, verbose=False)
+        print("✅ AI Brain is Active.")
+        return _llm_instance
+    except Exception as e:
+        print(f"❌ Failed to load AI: {e}")
+        return None
+
 # ==============================================================
 # 🎯 FIX 1: DEFINE NESTED CLASSES FIRST
 # These MUST come before they are used in ForecastResponse.
@@ -1770,14 +1790,10 @@ async def decode_trend_endpoint(req: TrendAnalysisRequest):
         # 2. Process the request
         from core.thunderbird_engine import decode_market_trend
         
-        # We don't need to await this because the llm_instance call itself is synchronous
-        # The benefit of async on the endpoint is that FastAPI can handle other requests
-        # while this one is waiting for the AI.
         result = decode_market_trend(req.topic, ai_brain)
         
         return result
     
     except Exception as e:
         print(f"❌ AI Decoding Error in Endpoint: {e}")
-        # traceback.print_exc()
         raise HTTPException(status_code=500, detail="An internal error occurred in the AI.")