Refactor model loading: implement lazy loading for models and update .gitignore

.gitignore

@@ -1,2 +1,2 @@
 __pycache__
-venv
+.venv

app/core/models.py

@@ -1,14 +1,22 @@
+# app/core/models.py
+
 import joblib
 from sentence_transformers import SentenceTransformer
 import os
 import requests
-from huggingface_hub import snapshot_download # <-- Add this import
+from huggingface_hub import snapshot_download
 
 TEMP_DIR = "/tmp"
 
-# This function for your .pkl files is still correct and needed
-def download_and_load_model(model_url, model_filename):
-    # ... (no changes needed in this function)
+# --- Initialize models as None ---
+# They will be loaded into these global variables later.
+tfidf_vectorizer = None
+le = None
+rf_model = None
+sentence_model = None
+
+# This function remains the same
+def download_and_load_pkl_model(model_url, model_filename):
     local_path = os.path.join(TEMP_DIR, model_filename)
     if not os.path.exists(local_path):
         print(f"Downloading model from {model_url} to {local_path}...")
@@ -28,27 +36,19 @@ def download_and_load_model(model_url, model_filename):
         print(f"Failed to load model {local_path}. Error: {e}")
         return None
 
-# --- NEW FUNCTION FOR THE SENTENCE TRANSFORMER ---
+# This function remains the same
 def get_sentence_transformer(model_name='all-MiniLM-L6-v2'):
-    """
-    Downloads the SentenceTransformer model to /tmp if it doesn't exist,
-    then loads it from there.
-    """
     local_model_path = os.path.join(TEMP_DIR, model_name)
-
     if not os.path.exists(local_model_path):
         print(f"Downloading SentenceTransformer model '{model_name}' to {local_model_path}...")
-        # Use snapshot_download to get all files for the model from Hugging Face
         try:
             snapshot_download(repo_id=f"sentence-transformers/{model_name}",
                               local_dir=local_model_path,
-                              local_dir_use_symlinks=False) # This is important for Vercel
+                              local_dir_use_symlinks=False)
             print("Download complete.")
         except Exception as e:
             print(f"Failed to download SentenceTransformer model. Error: {e}")
             return None
-
-    # Load the model from the local path in /tmp
     try:
         print(f"Loading SentenceTransformer model from {local_model_path}...")
         return SentenceTransformer(local_model_path)
@@ -56,30 +56,34 @@ def get_sentence_transformer(model_name='all-MiniLM-L6-v2'):
         print(f"Failed to load SentenceTransformer model from {local_model_path}. Error: {e}")
         return None
 
-# --- Main Model Loading Logic ---
-try:
-    # Your .pkl model loading remains the same
+# --- NEW LAZY LOADING FUNCTION ---
+def load_all_models():
+    """
+    This function loads all models into the global variables.
+    It will only be called when the models are first needed.
+    """
+    global tfidf_vectorizer, le, rf_model, sentence_model
+    print("--- First request received, initiating model loading... ---")
+
+    # Load the classification models
     TFIDF_URL = "https://pub-4a389a9b2dc842a2a55678d2db0ec0c6.r2.dev/tfidf_vectorizer.pkl"
     LE_URL = "https://pub-4a389a9b2dc842a2a55678d2db0ec0c6.r2.dev/label_encoder.pkl"
     RF_MODEL_URL = "https://pub-4a389a9b2dc842a2a55678d2db0ec0c6.r2.dev/random_forest_model.pkl"
 
-    tfidf_vectorizer = download_and_load_model(TFIDF_URL, "tfidf_vectorizer.pkl")
-    le = download_and_load_model(LE_URL, "label_encoder.pkl")
-    rf_model = download_and_load_model(RF_MODEL_URL, "random_forest_model.pkl")
+    tfidf_vectorizer = download_and_load_pkl_model(TFIDF_URL, "tfidf_vectorizer.pkl")
+    le = download_and_load_pkl_model(LE_URL, "label_encoder.pkl")
+    rf_model = download_and_load_pkl_model(RF_MODEL_URL, "random_forest_model.pkl")
 
     if all([tfidf_vectorizer, le, rf_model]):
         print("Classification models loaded successfully.")
     else:
         print("One or more classification models failed to load.")
 
-    # --- THIS IS THE LINE TO CHANGE ---
-    # Old line: sentence_model = SentenceTransformer('all-MiniLM-L6-v2')
-    # New line:
+    # Load the sentence transformer
     sentence_model = get_sentence_transformer()
     if sentence_model:
         print("SentenceTransformer model loaded successfully.")
     else:
         print("SentenceTransformer model failed to load.")
 
-except Exception as e:
-    print(f"An unexpected error occurred during model loading: {e}")
+    print("--- Model loading complete. ---")

app/services/ranking.py

@@ -3,7 +3,7 @@ import json
 import os
 from geopy.geocoders import Nominatim
 
-from ..core.models import tfidf_vectorizer, le, rf_model, sentence_model
+from ..core.models import load_all_models, tfidf_vectorizer, le, rf_model, sentence_model
 from ..schemas.recommendation import RecommendationRequest
 from ..utils.text import clean_resume
 
@@ -55,6 +55,9 @@ def geo_coords(city_name: str) -> tuple | None:
 
 def get_category_prediction(profile_text: str) -> str:
     """Processes text and predicts the job category."""
+    if rf_model is None:
+        load_all_models()
+
     cleaned_text = profile_text.lower()
     vectorized_text = tfidf_vectorizer.transform([cleaned_text])
     prediction_encoded = rf_model.predict(vectorized_text)[0]
@@ -64,6 +67,9 @@ def get_category_prediction(profile_text: str) -> str:
 def  get_ranked_internships(request: RecommendationRequest) -> list[int]:
     """Performs two-stage ranking with dynamic geocoding."""
 
+    if sentence_model is None:
+        load_all_models()
+
     profile_text_to_encode = request.profile_text
 
     if request.predicted_category: