fix: Final definitive fix for model initialization race condition

api/main.py

@@ -27,10 +27,13 @@ from core.strategist import AIStrategist
 from core.predictor import rank_influencers_by_match
 from core.utils import get_supabase_client
 from core.anomaly_detector import find_anomalies
-from core.matcher import load_embedding_model, rank_documents_by_similarity
+from core.matcher import rank_documents_by_similarity
 from core.utils import get_supabase_client, extract_colors_from_url
 from core.document_parser import parse_pdf_from_url
-from core.creative_chat import CreativeDirector 
+from core.creative_chat import CreativeDirector
+from core.strategist import AIStrategist
+from core.support_agent import SupportAgent
+from core.matcher import load_embedding_model
 
 try:
     from core.rag.store import VectorStore
@@ -48,7 +51,7 @@ 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_PATH = os.path.join(ROOT_DIR, 'embedding_model') # This path is correct
+EMBEDDING_MODEL_PATH = os.path.join(ROOT_DIR, 'embedding_model')
 DB_PATH = os.path.join(os.environ.get("WRITABLE_DIR", "/tmp"), "vector_db_persistent")
 
 _llm_instance: Optional[Llama] = None
@@ -70,18 +73,9 @@ _performance_scorer = None
 def to_snake(name: str) -> str:
     return re.sub(r'(?<!^)(?=[A-Z])', '_', name).lower()
 
-class ChatMessage(BaseModel):
-    role: str # "user" or "assistant"
-    content: str
-
-class CreativeChatRequest(BaseModel):
-    message: str
-    history: List[ChatMessage]
-    task_context: str 
-
-class FinalizeScriptRequest(BaseModel):
-    history: List[ChatMessage]
-    task_context: str
+class CreativeChatRequest(BaseModel): message: str; history: list; task_context: str
+class FinalizeScriptRequest(BaseModel): history: list; task_context: str
+class FinalScriptResponse(BaseModel): hook: str; script: str; visuals: List[str]; tools: List[str]
 
 class FinalScriptResponse(BaseModel):
     hook: str
@@ -498,37 +492,94 @@ app = FastAPI(title="Reachify AI Service (Deploy-Ready)", version="11.0.0")
 
 @app.on_event("startup")
 def startup_event():
-    global _llm_instance, _creative_director, _ai_strategist, _support_agent
+    # 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
     
+    # --- STEP 1: DOWNLOAD AND LOAD THE LLM MODEL ---
     print("--- 🚀 AI Service Starting Up... ---")
-    
-    # STEP 1: DOWNLOAD AND LOAD THE LLM
     try:
         os.makedirs(MODEL_SAVE_DIRECTORY, exist_ok=True)
         if not os.path.exists(LLAMA_MODEL_PATH):
-            print(f"   - Downloading LLM: {MODEL_FILENAME}...")
-            hf_hub_download(repo_id=MODEL_REPO, filename=MODEL_FILENAME, local_dir=MODEL_SAVE_DIRECTORY)
-            print("   - ✅ Download complete.")
+            print(f"   - LLM model not found locally. Downloading '{MODEL_FILENAME}'...")
+            hf_hub_download(
+                repo_id=MODEL_REPO,
+                filename=MODEL_FILENAME,
+                local_dir=MODEL_SAVE_DIRECTORY,
+                local_dir_use_symlinks=False
+            )
+            print("   - ✅ Model downloaded successfully.")
+        else:
+            print(f"   - LLM model found locally at {LLAMA_MODEL_PATH}. Skipping download.")
         
-        print("   - Loading LLM 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)
-        print("   - ✅ LLM Loaded.")
+        print("   - ✅ LLM Loaded successfully.")
     
     except Exception as e:
-        print(f"   - ❌ FATAL ERROR: Could not load LLM. Error: {e}")
-        return
+        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 global variable is None on failure
+        # We don't return here, so that other non-LLM models can still load.
 
-    # 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}")
+    # --- STEP 2: INITIALIZE ALL AI COMPONENTS THAT NEED THE LLM ---
+    # This part only runs if the LLM was loaded successfully
+    if _llm_instance:
+        try:
+            print("   - Initializing AI components that depend on LLM...")
+            
+            # Initialize CreativeDirector
+            _creative_director = CreativeDirector(llm_instance=_llm_instance)
+            
+            # Initialize VectorStore (if it exists)
+            if VectorStore:
+                _vector_store = VectorStore()
+                print("   - RAG Engine Ready.")
+            
+            # Initialize AIStrategist
+            _ai_strategist = AIStrategist(llm_instance=_llm_instance, store=_vector_store)
+            
+            # Initialize SupportAgent
+            _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("\n--- ✅ AI Service startup sequence finished! ---")
+        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)
 
+    print("\n--- ✅ AI Service startup sequence finished! ---")
+    
 @app.get("/")
 def health_check():
     if _llm_instance:

core/creative_chat.py

@@ -1,12 +1,9 @@
-
 import os
 import sys
 from llama_cpp import Llama
 import json
 import re
 
-# Path setup to import VectorStore from the parent directory
-# This logic is correct and remains the same
 current_dir = os.path.dirname(os.path.abspath(__file__))
 parent_dir = os.path.dirname(current_dir)
 sys.path.append(parent_dir)
@@ -14,30 +11,11 @@ sys.path.append(parent_dir)
 from core.rag.store import VectorStore
 
 class CreativeDirector:
-    def __init__(self):
-        """Initialize Model and Memory once to save time."""
-        
-        # ✅ THE FIX IS HERE: We now look for the model in the writable directory
-        # where main.py downloads it, not in the read-only application directory.
-        model_name = "tinyllama-1.1b-chat-v1.0.Q4_K_M.gguf"
-        # Hugging Face Spaces provides '/data' as a writable persistent directory.
-        writable_dir = os.environ.get("WRITABLE_DIR", "/data") 
-        model_path = os.path.join(writable_dir, "llm_model", model_name)
-        # =====================================================================
-        
-        if not os.path.exists(model_path):
-            # This error will now correctly point to the writable directory
-            raise FileNotFoundError(f"❌ Model not found at: {model_path}. Please check the download logic in main.py.")
-
-        print("🧠 Loading AI Director (TinyLlama - SUPER FAST MODE)...")
-        
-        self.llm = Llama(
-            model_path=model_path,
-            n_ctx=512,       
-            n_batch=32,      
-            n_threads=4,     
-            verbose=False
-        )
+    def __init__(self, llm_instance: Llama):
+        if not llm_instance:
+            raise ValueError("CreativeDirector received an invalid LLM instance.")
+        print("🧠 Initializing AI Director with pre-loaded LLM...")
+        self.llm = llm_instance
         self.memory = VectorStore(collection_name="creative_mind")
         print("✅ AI Director is Online.")