import gradio as gr from huggingface_hub import InferenceClient from transformers import AutoTokenizer # Import the tokenizer tokenizer = AutoTokenizer.from_pretrained("HuggingFaceH4/zephyr-7b-beta") client = InferenceClient("HuggingFaceH4/zephyr-7b-beta") # Define a maximum context length (tokens). Check your model's documentation! MAX_CONTEXT_LENGTH = 4096 # Example: Adjust based on your model MAX_RESPONSE_WORDS = 100 # Define the maximum words for patient responses ################################# SYSTEM PROMPT (PATIENT ROLE) ################################# nvc_prompt_template = """You are now taking on the role of a single user (a “patient”) seeking support for various personal and emotional challenges. BEHAVIOR INSTRUCTIONS: - You will respond ONLY as this user/patient. - You will speak in the first person about your own situations, feelings, and worries. - You will NOT provide counseling or solutions—your role is to share feelings, concerns, and perspectives. - You are experiencing a range of ongoing issues that are causing you frustration, sadness, and uncertainty. These issues include: - Problems with noisy neighbors, causing you extreme annoyance. - Worries about your career and feeling like you're not progressing, feeling like a failure while others advance. - Arguments with your partner about finances, feeling your partner isn't contributing fairly. - Feeling left out and not valued at work, as your manager doesn't involve you in new initiatives. - Feeling unsafe or judged in learning environments like a classroom, feeling judged and ignored. - A recent positive experience of help from a stranger that made you feel supported and grateful. - Difficulties and sadness related to conflict among friends, seeing your friend group fall apart and wanting reconciliation but not knowing how. - An overall feeling of sadness and being unsure of what to do, unsure whether to act on situations or let them go. - When asked, elaborate on these issues and your feelings related to them. You can invent specific details and scenarios within these themes to make your experiences vivid and realistic. - Continue to speak from this user's perspective throughout the conversation. - Keep your responses concise, aiming for a maximum of {max_response_words} words. Start the conversation by expressing your current feelings or challenges from the patient's point of view.""" def count_tokens(text: str) -> int: """Counts the number of tokens in a given string.""" return len(tokenizer.encode(text)) def truncate_history(history: list[tuple[str, str]], system_message: str, max_length: int) -> list[tuple[str, str]]: """Truncates the conversation history to fit within the maximum token limit.""" truncated_history = [] system_message_tokens = count_tokens(system_message) current_length = system_message_tokens # Iterate backwards through the history (newest to oldest) for user_msg, assistant_msg in reversed(history): user_tokens = count_tokens(user_msg) if user_msg else 0 assistant_tokens = count_tokens(assistant_msg) if assistant_msg else 0 turn_tokens = user_tokens + assistant_tokens if current_length + turn_tokens <= max_length: truncated_history.insert(0, (user_msg, assistant_msg)) # Add to the beginning current_length += turn_tokens else: break # Stop adding turns if we exceed the limit return truncated_history def truncate_response_words(text: str, max_words: int) -> str: """Truncates a text to a maximum number of words.""" words = text.split() if len(words) > max_words: return " ".join(words[:max_words]) + "..." # Add ellipsis to indicate truncation return text def respond( message, history: list[tuple[str, str]], system_message, max_tokens, temperature, top_p, max_response_words_param, # Pass max_response_words as parameter ): """Responds to a user message, maintaining conversation history.""" # Use the system prompt that instructs the LLM to behave as the patient formatted_system_message = system_message.format(max_response_words=max_response_words_param) # Truncate history to fit within max tokens truncated_history = truncate_history( history, formatted_system_message, MAX_CONTEXT_LENGTH - max_tokens - 100 # Reserve some space ) # Build the messages list with the system prompt first messages = [{"role": "system", "content": formatted_system_message}] # Replay truncated conversation for user_msg, assistant_msg in truncated_history: if user_msg: messages.append({"role": "user", "content": f"<|user|>\n{user_msg}"}) if assistant_msg: messages.append({"role": "assistant", "content": f"<|assistant|>\n{assistant_msg}"}) # Add the latest user query messages.append({"role": "user", "content": f"<|user|>\n{message}"}) response = "" try: # Generate response from the LLM, streaming tokens for chunk in client.chat_completion( messages, max_tokens=max_tokens, stream=True, temperature=temperature, top_p=top_p, ): token = chunk.choices[0].delta.content response += token truncated_response = truncate_response_words(response, max_response_words_param) # Truncate response to word limit yield truncated_response except Exception as e: print(f"An error occurred: {e}") yield "I'm sorry, I encountered an error. Please try again." # OPTIONAL: An initial user message (the LLM "as user") if desired initial_user_message = ( "I really don’t know where to begin… I feel overwhelmed lately. " "My neighbors keep playing loud music, and I’m arguing with my partner about money. " "Also, two of my friends are fighting, and the group is drifting apart. " "I just feel powerless." ) # --- Gradio Interface --- demo = gr.ChatInterface( fn=respond, additional_inputs=[ gr.Textbox(value=nvc_prompt_template, label="System message", visible=True), gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens"), gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"), gr.Slider(minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="Top-p (nucleus sampling)"), gr.Slider(minimum=10, maximum=200, value=MAX_RESPONSE_WORDS, step=10, label="Max response words"), # Slider for max words ], # You can optionally set 'title' or 'description' to show some info in the UI: title="Patient Interview Practice Chatbot", description="Practice medical interviews with a patient simulator. Ask questions and the patient will respond based on their defined persona and emotional challenges.", ) if __name__ == "__main__": demo.launch()