error response handling

app.py

@@ -79,17 +79,24 @@ def chat_completions():
         response.raise_for_status()
 
         if stream:
-            iterator = response.iter_content(chunk_size=8192)
+            iterator = response.iter_lines()
             try:
-                first_chunk = next(iterator)
+                first_line = next(iterator)
             except StopIteration:
                 return Response(status=204)
 
-            if not first_chunk.strip().startswith(b'data:'):
-                error_message = first_chunk.decode('utf-8')
+            # The first line might be empty, skip it if so
+            if not first_line:
+                try:
+                    first_line = next(iterator)
+                except StopIteration:
+                    return Response(status=204)
+
+            if not first_line.strip().startswith(b'data:'):
+                error_message = first_line.decode('utf-8')
                 # Try to get more of the error message if it exists
                 try:
-                    error_message += "".join(chunk.decode('utf-8') for chunk in iterator)
+                    error_message += "\n" + "\n".join(line.decode('utf-8') for line in iterator)
                 except Exception:
                     pass
                 return jsonify({
@@ -102,9 +109,9 @@ def chat_completions():
                 }), 400
 
             def generate():
-                yield first_chunk
-                for chunk in iterator:
-                    yield chunk
+                yield first_line + b'\n'
+                for line in iterator:
+                    yield line + b'\n'
             
             return Response(generate(), content_type=response.headers['content-type'])
         else: