fix: Definitive fix for model initialization order

api/main.py

@@ -48,13 +48,9 @@ MODEL_REPO = "TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF"
 MODEL_FILENAME = "tinyllama-1.1b-chat-v1.0.Q4_K_M.gguf"
 MODEL_SAVE_DIRECTORY = os.path.join(os.environ.get("WRITABLE_DIR", "/data"), "llm_model")
 LLAMA_MODEL_PATH = os.path.join(MODEL_SAVE_DIRECTORY, MODEL_FILENAME)
-
-EMBEDDING_MODEL_NAME = 'sentence-transformers/all-MiniLM-L6-v2'
-EMBEDDING_MODEL_PATH = os.path.join(ROOT_DIR, 'embedding_model')
+EMBEDDING_MODEL_PATH = os.path.join(ROOT_DIR, 'embedding_model') # This path is correct
 DB_PATH = os.path.join(os.environ.get("WRITABLE_DIR", "/tmp"), "vector_db_persistent")
 
-FINAL_EMBEDDING_PATH = EMBEDDING_MODEL_PATH if os.path.exists(EMBEDDING_MODEL_PATH) else EMBEDDING_MODEL_NAME
-
 _llm_instance: Optional[Llama] = None
 _vector_store: Optional[Any] = None
 _ai_strategist: Optional[AIStrategist] = None
@@ -502,100 +498,43 @@ app = FastAPI(title="Reachify AI Service (Deploy-Ready)", version="11.0.0")
 
 @app.on_event("startup")
 def startup_event():
-    # Make sure we can modify the global variables
-    global _llm_instance, _creative_director, _support_agent, _ai_strategist, _vector_store, \
-           _budget_predictor, _influencer_matcher, _performance_predictor, _payout_forecaster, \
-           _earnings_optimizer, _earnings_encoder, _likes_predictor, _comments_predictor, \
-           _revenue_forecaster, _performance_scorer
+    global _llm_instance, _creative_director, _ai_strategist, _support_agent
     
-    # --- STEP 1: DOWNLOAD AND LOAD THE LLM MODEL ---
     print("--- 🚀 AI Service Starting Up... ---")
+    
+    # STEP 1: DOWNLOAD AND LOAD THE LLM
     try:
-        # Create the directory where the model will be saved if it doesn't exist
         os.makedirs(MODEL_SAVE_DIRECTORY, exist_ok=True)
-        
-        # Check if the model file already exists before trying to download it
         if not os.path.exists(LLAMA_MODEL_PATH):
-            print(f"   - LLM model not found locally. Downloading '{MODEL_FILENAME}'...")
-            # This function downloads the file from the Hub to the specified directory
-            hf_hub_download(
-                repo_id=MODEL_REPO,
-                filename=MODEL_FILENAME,
-                local_dir=MODEL_SAVE_DIRECTORY,
-                local_dir_use_symlinks=False # Important for container environments
-            )
-            print("   - ✅ Model downloaded successfully.")
-        else:
-            print(f"   - LLM model found locally at {LLAMA_MODEL_PATH}. Skipping download.")
+            print(f"   - Downloading LLM: {MODEL_FILENAME}...")
+            hf_hub_download(repo_id=MODEL_REPO, filename=MODEL_FILENAME, local_dir=MODEL_SAVE_DIRECTORY)
+            print("   - ✅ Download complete.")
         
-        # Now that the file is guaranteed to be there, load it into memory
-        print("   - Loading Llama LLM into memory...")
-        _llm_instance = Llama(model_path=LLAMA_MODEL_PATH, n_gpu_layers=0, n_ctx=2048, verbose=False, use_mmap=False)
-        print("   - ✅ LLM Loaded successfully.")
+        print("   - Loading LLM into memory...")
+        _llm_instance = Llama(model_path=LLAMA_MODEL_PATH, n_gpu_layers=0, n_ctx=2048, verbose=False)
+        print("   - ✅ LLM Loaded.")
     
     except Exception as e:
-        # If anything in this block fails, the LLM is not usable.
-        print(f"   - ❌ FATAL ERROR: Could not download or load the LLM model. LLM-dependent features will be disabled.")
-        traceback.print_exc()
-        _llm_instance = None # Ensure the global variable is None
-
-    # --- STEP 2: INITIALIZE ALL AI COMPONENTS THAT NEED THE LLM ---
-    # This part only runs if the LLM was loaded successfully (_llm_instance is not None)
-    if _llm_instance:
-        try:
-            print("   - Initializing AI components...")
-            _creative_director = CreativeDirector(llm_instance=_llm_instance)
-            
-            if VectorStore:
-                _vector_store = VectorStore()
-                print("   - RAG Engine Ready.")
-            
-            _ai_strategist = AIStrategist(llm_instance=_llm_instance, store=_vector_store)
-            _support_agent = SupportAgent(llm_instance=_llm_instance, embedding_path=EMBEDDING_MODEL_PATH, db_path=DB_PATH)
-            
-            print("   - ✅ Core AI components (Director, Strategist, Agent) are online.")
+        print(f"   - ❌ FATAL ERROR: Could not load LLM. Error: {e}")
+        return
 
-        except Exception as e:
-            print(f"   - ❌ FAILED to initialize core AI components: {e}")
-            traceback.print_exc()
-    else:
-        print("   - ⚠️ SKIPPING initialization of LLM-dependent components because LLM failed to load.")
-
-    # --- STEP 3: LOAD ALL OTHER MODELS (These don't depend on the LLM) ---
-    print("   - Loading ML models from joblib files...")
-    model_paths = {
-        'budget': ('_budget_predictor', 'budget_predictor_v1.joblib'),
-        'matcher': ('_influencer_matcher', 'influencer_matcher_v1.joblib'),
-        'performance': ('_performance_predictor', 'performance_predictor_v1.joblib'),
-        'payout': ('_payout_forecaster', 'payout_forecaster_v1.joblib'),
-        'earnings': ('_earnings_optimizer', 'earnings_model.joblib'),
-        'earnings_encoder': ('_earnings_encoder', 'earnings_encoder.joblib'),
-        'likes_predictor': ('_likes_predictor', 'likes_predictor_v1.joblib'),
-        'comments_predictor': ('_comments_predictor', 'comments_predictor_v1.joblib'),
-        'revenue_forecaster': ('_revenue_forecaster', 'revenue_forecaster_v1.joblib'),
-        'performance_scorer': ('_performance_scorer', 'performance_scorer_v1.joblib'),
-    }
-    for name, (var, file) in model_paths.items():
-        path = os.path.join(MODELS_DIR, file)
-        try:
-            globals()[var] = joblib.load(path)
-            print(f"     - Loaded {name} model.")
-        except FileNotFoundError:
-            globals()[var] = None
-            print(f"   - ⚠️ WARNING: Model '{name}' not found at {path}. Endpoint will be disabled.")
-    
-    print("   - Initializing Text Embedding Model...")
-    load_embedding_model(EMBEDDING_MODEL_PATH)
+    # STEP 2: INITIALIZE AI COMPONENTS that need the LLM
+    print("   - Initializing AI components...")
+    try:
+        _creative_director = CreativeDirector(llm_instance=_llm_instance)
+        _ai_strategist = AIStrategist(llm_instance=_llm_instance, store=None) # Assuming store is optional
+        _support_agent = SupportAgent(llm_instance=_llm_instance, embedding_path=EMBEDDING_MODEL_PATH, db_path=DB_PATH)
+    except Exception as e:
+        print(f"   - ❌ Error initializing AI agents: {e}")
 
     print("\n--- ✅ AI Service startup sequence finished! ---")
 
-@app.get("/", summary="Health Check")
-def read_root():
-    # We add a check here to see if the LLM loaded successfully during startup.
-    # This helps with debugging on the live server.
-    if _llm_instance is None:
-        return {"status": "AI Service is running, but the Core LLM FAILED to load. Check logs."}
-    return {"status": "AI Service is running and all models are loaded."}
+@app.get("/")
+def health_check():
+    if _llm_instance:
+        return {"status": "AI Service is Running"}
+    else:
+        return {"status": "AI Service is in a degraded state: Core LLM failed to load."}
 
 def _cleanup_llm_response(data: dict) -> dict:
     """A robust helper to clean common messy JSON outputs from smaller LLMs."""