Process response length

app.py

@@ -110,8 +110,14 @@ def fit_to_length(text, min_length=3, max_length=100):
         return None
     return text
 
+def get_completion_text(messages_think: List[ChatMessage]):
+    txt = ""
+    for cm in messages_think:
+        txt = " ".join((txt, cm.content))
+    return txt
 
-def get_model_reponse(messages_think):
+
+def get_model_reponse(messages_think: List[ChatMessage]):
     """Process the text content."""
 
     # Prepare the model input
@@ -148,7 +154,9 @@ async def completion(data: Completion):
         raise HTTPException(status_code=503, detail="Model not loaded")
 
     try:
-        result = get_model_reponse(data)
+        mt = data.messages_think
+        text = get_completion_text(mt)
+        result = get_model_reponse(mt)
 
         return {
             "choices": [