Final production optimizations: pre-download BERT and explicit model status detection

Dockerfile

@@ -20,6 +20,11 @@ COPY hf_requirements.txt .
 RUN pip install --upgrade pip && \
     pip install -r hf_requirements.txt
 
+# Pre-download the BERT base model during build to speed up startup
+RUN python -c "from transformers import BertTokenizer, BertForSequenceClassification; \
+    BertTokenizer.from_pretrained('bert-base-uncased'); \
+    BertForSequenceClassification.from_pretrained('bert-base-uncased', num_labels=2)"
+
 # Copy the rest of the application code
 COPY . .
 

hf_api.py

@@ -68,6 +68,10 @@ def analyze(request: AnalysisRequest):
         if not content:
             return {"error": "Could not retrieve content for this article."}
 
+        # Ensure model is loaded (pre-load should have handled this, but be safe)
+        if not service.bias_model:
+            service.load_local_bias_model()
+
         if request.action == "get_summary":
             summary = service.summarize_content(content[:3000])
             return {
@@ -86,6 +90,14 @@ def analyze(request: AnalysisRequest):
             # 2. Analyze sentences in a single batch
             batch_results = service.rate_bias_batch(analysis_sentences)
             
+            # Check if all results are "Offline" or "Error"
+            is_offline = all(res.get("label") in ["Offline", "Error"] for res in batch_results)
+            if is_offline:
+                return {
+                    "error": "The bias detection model is currently offline or failing to load.",
+                    "details": batch_results[0].get("reasoning", "Unknown error")
+                }
+            
             # 3. Get Summary
             summary = service.summarize_content(content)