Fix model caching - use HF_HOME for all models to prevent re-download on restart

Dockerfile

@@ -35,11 +35,11 @@ RUN python -c "from transformers import AutoTokenizer, AutoModelForSequenceClass
     AutoModelForSequenceClassification.from_pretrained('Cyberlace/swara-structure-model', cache_dir='/.cache'); \
     print('✅ Structure Model cached!')" && chmod -R 777 /.cache
 
-# 2. Whisper will be downloaded on first use (lazy loading to avoid OOM during build)
-# RUN python -c "import whisper; \
-#     print('📥 Downloading Whisper medium model...'); \
-#     whisper.load_model('medium', download_root='/.cache'); \
-#     print('✅ Whisper medium cached!')"
+# 2. Download Whisper medium model (~1.5GB) - with memory optimization
+RUN python -c "import whisper; \
+    print('📥 Downloading Whisper medium model (1.5GB)...'); \
+    whisper.load_model('medium', download_root='/.cache'); \
+    print('✅ Whisper medium cached!')" && chmod -R 777 /.cache
 
 # 3. Download Sentence Transformer for Keywords (~420MB)
 RUN python -c "from sentence_transformers import SentenceTransformer; \

app/services/keywords.py

@@ -42,7 +42,10 @@ class KeywordService:
         if EMBEDDINGS_AVAILABLE:
             print(f"📦 Loading BERT model: {model_name}...")
             device = get_device()
-            self.model = SentenceTransformer(model_name, device=device)
+            # Use HF_HOME cache directory (set in Dockerfile)
+            import os
+            cache_dir = os.environ.get('HF_HOME', '/.cache')
+            self.model = SentenceTransformer(model_name, device=device, cache_folder=cache_dir)
             print("✅ Model loaded!")
         else:
             self.model = None