Update app.py

app.py

@@ -146,15 +146,22 @@ def generate(prompt, history):
 
         response = ""
 
-        while(len(response) < 1):
-            output = llm(context, max_tokens=400, stop=["Nurse:"], echo=False)
-            response = output["choices"][0]["text"]
-            response = response.strip()
+#        while(len(response) < 1):
+ #           output = llm(context, max_tokens=400, stop=["Nurse:"], echo=False)
+ #           response = output["choices"][0]["text"]
+  #          response = response.strip()
+
+        for output in llm(input, stream=True, max_tokens=100, ):
+            piece = output['choices'][0]['text']
+            response += piece
+            chatbot[-1] = (chatbot[-1][0], response)
+
+            yield  response
 
         context += response
         print (context)
 
-        history.append((prompt,response))
+        history.append(response)
         yield response
         
     else: