create formatted chat history

app.py

@@ -8,10 +8,16 @@ from huggingface_hub import InferenceClient
 
 
 ASR_MODEL_NAME = "openai/whisper-small"
-NLP_MODEL_NAME = "mistralai/Mistral-7B-Instruct-v0.2"
-system_prompt = """"<s> [INST] You are Friday a helpful and conversational assistant. [/INST]"""
+LLM_MODEL_NAME = "mistralai/Mistral-7B-Instruct-v0.2"
 
-client = InferenceClient(NLP_MODEL_NAME)
+
+system_prompt = """"<s>[INST] You are Friday, a helpful and conversational AI assistant and You respond with one to two sentences. [/INST] Hello there! I'm friday how can I help you?</s>"""
+
+chat_history = system_prompt + """"""
+
+formatted_history = """"""
+
+client = InferenceClient(LLM_MODEL_NAME)
 
 device = 0 if torch.cuda.is_available() else "cpu"
 
@@ -22,7 +28,7 @@ pipe = pipeline(
 )
 
 
-def generate(prompt, temperature=0.1, max_new_tokens=64, top_p=0.95, repetition_penalty=1.0):
+def generate(user_prompt, temperature=0.1, max_new_tokens=128, top_p=0.95, repetition_penalty=1.0):
     temperature = float(temperature)
     if temperature < 1e-2:
         temperature = 1e-2
@@ -37,10 +43,10 @@ def generate(prompt, temperature=0.1, max_new_tokens=64, top_p=0.95, repetition_
         seed=42,
     )
 
-    formatted_prompt = system_prompt + f""" {prompt} </s>"""
+    chat_history += f""" <s>[INST] {user_prompt}  [/INST] """
 
     output = client.text_generation(
-        formatted_prompt, **generate_kwargs, stream=False, details=False, return_full_text=False)
+        chat_history, **generate_kwargs, stream=False, details=False, return_full_text=False)
 
     print(output)
     return output
@@ -54,13 +60,18 @@ def transcribe(audio):
 
     inputs = pipe({"sampling_rate": sr, "raw": y})["text"]
 
-    print("User transcription: ", inputs)
+    formatted_history += f"""Human: {inputs}\n"""
+
+    llm_response = generate(inputs)
+
+    chat_history += f""" {llm_response}</s>"""
+
+    formatted_history += f"""Friday: {llm_response}\n"""
 
-    response = generate(inputs)
-    audio_response = gTTS(response)
+    audio_response = gTTS(llm_response)
     audio_response.save("response.mp3")
 
-    print(audio_response)
+    print(formatted_history)
 
     return "response.mp3"