Update app.py

app.py

@@ -11,21 +11,23 @@ llm = Llama(model_path= hf_hub_download(repo_id="TheBloke/orca_mini_3B-GGML", fi
 
 history = []
 
+preprompt = " The user and the AI are having a conversation : "
+
 def generate_text(input_text, history):
     print("history ",history)
     print("input ", input_text)
     if history == []:
-        input_text_with_history = f"Q: {input_text} \n A:"
+        input_text_with_history = f"{pre_prompt}"+ "\n" + f"Q: {input_text} " + "\n" +" A:"
     else:
         input_text_with_history = f"{history[-1][1]}"+ "\n"
-        input_text_with_history += f"Q: {input_text} \n A:"
+        input_text_with_history += f"Q: {input_text}" + "\n" +" A:"
     print("new input", input_text_with_history)
-    output = llm(input_text_with_history, max_tokens=1024, stop=["Q:", "\n"], stream=True, echo=True)
+    output = llm(input_text_with_history, max_tokens=1024, stop=["Q:", "\n"], stream=True)
 
     for out in output:
      stream = copy.deepcopy(out)
      print(stream["choices"][0]["text"])
-    yield output
+    yield stream["choices"][0]["text"]
     
     history =["init",input_text_with_history]