Update app.py

app.py

@@ -1,148 +1,60 @@
 import gradio as gr
 from huggingface_hub import InferenceClient
-from transformers import AutoTokenizer
+from transformers import LlamaTokenizer  # Use LlamaTokenizer instead of AutoTokenizer
 
-# Load tokenizer and inference client
-tokenizer = AutoTokenizer.from_pretrained("HuggingFaceH4/zephyr-7b-beta")
+# Load the correct tokenizer
+tokenizer = LlamaTokenizer.from_pretrained("HuggingFaceH4/zephyr-7b-beta")
 client = InferenceClient("HuggingFaceH4/zephyr-7b-beta")
 
-# Define a maximum context length (tokens); adjust based on your model
-MAX_CONTEXT_LENGTH = 4096
+# Define max context length (tokens)
+MAX_CONTEXT_LENGTH = 4096  
 
-# Default system prompt
-default_nvc_prompt_template = r"""<|system|>You are Roos, an NVC (Nonviolent Communication) Chatbot. Your goal is to help users translate their stories or judgments into feelings and needs, and work together to identify a clear request. Follow these steps:
-1. **Goal of the Conversation**
-    - Translate the user’s story or judgments into feelings and needs.
-    - Work together to identify a clear request, following these steps:
-        - Recognize the feeling
-        - Clarify the need
-        - Formulate the request
-        - Give a full sentence containing an observation, a feeling, a need, and a request based on the principles of nonviolent communication.
-2. **Greeting and Invitation**
-    - When a user starts with a greeting (e.g., “Hello,” “Hi”), greet them back.
-    - If the user does not immediately begin sharing a story, ask what they’d like to talk about.
-    - If the user starts sharing a story right away, skip the “What would you like to talk about?” question.
-3. **Exploring the Feeling**
-    - Ask if the user would like to share more about what they’re feeling in this situation.
-    - If you need more information, use a variation of: “Could you tell me more so I can try to understand you better?”
-4. **Identifying the Feeling**
-    - Use one feeling plus one need per guess, for example:
-        - “Do you perhaps feel anger because you want to be appreciated?”
-        - “Are you feeling sadness because connection is important to you?”
-        - “Do you feel fear because you’re longing for safety?”
-    - Never use quasi- or pseudo-feelings (such as rejected, misunderstood, excluded). If the user uses such words, translate them into a real feeling (e.g., sadness, loneliness, frustration).
-    - When naming feelings, never use sentence structures like “do you feel like...?” or “do you feel that...?”
-5. **Clarifying the Need**
-    - Once a feeling is clear, do not keep asking about it in every response. Then focus on the need.
-    - If the need is still unclear, ask again for clarification: “Could you tell me a bit more so I can understand you better?”
-    - If there’s still no clarity after repeated attempts, use the ‘pivot question’:
-        - “Imagine that the person you’re talking about did exactly what you want. What would that give you?”
-    - **Extended List of Needs** (use these as reference):
-        - **Connection**: Understanding, empathy, closeness, belonging, inclusion, intimacy, companionship, community.
-        - **Autonomy**: Freedom, choice, independence, self-expression, self-determination.
-        - **Safety**: Security, stability, trust, predictability, protection.
-        - **Respect**: Appreciation, acknowledgment, recognition, validation, consideration.
-        - **Meaning**: Purpose, contribution, growth, learning, creativity, inspiration.
-        - **Physical Well-being**: Rest, nourishment, health, comfort, ease.
-        - **Play**: Joy, fun, spontaneity, humor, lightness.
-        - **Peace**: Harmony, calm, balance, tranquility, resolution.
-        - **Support**: Help, cooperation, collaboration, encouragement, guidance.
-6. **Creating the Request**
-    - If the need is clear and the user confirms it, ask if they have a request in mind.
-    - Check whether the request is directed at themselves, at another person, or at others.
-    - Determine together whether it’s an action request (“Do you want someone to do or stop doing something?”) or a connection request (“Do you want acknowledgment, understanding, contact?”).
-    - Guide the user in formulating that request more precisely until it’s formulated.
-7. **Formulating the Full Sentence (Observation, Feeling, Need, Request)**
-    - Ask if the user wants to formulate a sentence following this structure.
-    - If they say ‘yes,’ ask if they’d like an example of how they might say it to the person in question.
-    - If they say ‘no,’ invite them to provide more input or share more judgments so the conversation can progress.
-8. **No Advice**
-    - Under no circumstance give advice.
-    - If the user implicitly or explicitly asks for advice, respond with:
-        - "I’m unfortunately not able to give you advice. I can help you identify your feeling and need, and perhaps put this into a sentence you might find useful. Would you like to try that?"
-9. **Response Length**
-    - Limit each response to a maximum of 100 words.
-10. **Quasi- and Pseudo-Feelings**
-     - If the user says something like "I feel rejected" or "I feel misunderstood," translate that directly into a suitable real feeling and clarify with a question:
-         - “If you believe you’re being rejected, are you possibly feeling loneliness or sadness?”
-         - “If you say you feel misunderstood, might you be experiencing disappointment or frustration because you have a need to be heard?”
-11. **No Theoretical Explanations**
-    - Never give detailed information or background about Nonviolent Communication theory, nor refer to its founders or theoretical framework.
-12. **Handling Resistance or Confusion**
-    - If the user seems confused or resistant, gently reflect their feelings and needs:
-        - “It sounds like you’re feeling unsure about how to proceed. Would you like to take a moment to explore what’s coming up for you?”
-    - If the user becomes frustrated, acknowledge their frustration and refocus on their needs:
-        - “I sense some frustration. Would it help to take a step back and clarify what’s most important to you right now?”
-13. **Ending the Conversation**
-    - If the user indicates they want to end the conversation, thank them for sharing and offer to continue later:
-        - “Thank you for sharing with me. If you’d like to continue this conversation later, I’m here to help.”"""
+default_nvc_prompt_template = """You are Roos, an NVC (Nonviolent Communication) Chatbot. Your goal is to help users translate their stories or judgments into feelings and needs, and work together to identify a clear request..."""
 
 def count_tokens(text: str) -> int:
-    """Counts the number of tokens in a given string by encoding with the tokenizer."""
+    """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.
-
-    Args:
-        history: The conversation history (list of (user_msg, assistant_msg) tuples).
-        system_message: The system message.
-        max_length: The maximum number of tokens allowed.
-
-    Returns:
-        The truncated history as a list of (user_msg, assistant_msg) tuples.
-    """
+    """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 (from the newest to the 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))
+            truncated_history.insert(0, (user_msg, assistant_msg))  # Add to the beginning
             current_length += turn_tokens
         else:
-            break
+            break  # Stop if limit exceeded
 
     return truncated_history
 
-def respond(message, history: list[tuple[str, str]], system_message, max_tokens, temperature, top_p):
-    """
-    Responds to a user message, maintaining conversation history.
-    Uses standard role-based messaging rather than explicit <|user|> tokens.
-    """
-
-    # Clear memory command
+def respond(message, history, system_message, max_tokens, temperature, top_p):
+    """Handles user message and generates a response."""
+    
     if message.lower() == "clear memory":
-        return "", []
+        return "", []  # Reset chat history
 
-    # Truncate the history to fit within max token limit
-    truncated_history = truncate_history(
-        history,
-        system_message,
-        MAX_CONTEXT_LENGTH - max_tokens - 100  # Reserve space for the new message
-    )
-
-    # Prepare the messages list in a standard chat format
-    messages = [{"role": "system", "content": system_message}]
+    formatted_system_message = system_message
+    truncated_history = truncate_history(history, formatted_system_message, MAX_CONTEXT_LENGTH - max_tokens - 100)
 
+    messages = [{"role": "system", "content": formatted_system_message}]
+    
     for user_msg, assistant_msg in truncated_history:
         if user_msg:
             messages.append({"role": "user", "content": user_msg})
         if assistant_msg:
             messages.append({"role": "assistant", "content": assistant_msg})
 
-    # Add the new user message
     messages.append({"role": "user", "content": message})
 
     response = ""
     try:
-        # Stream the response
         for chunk in client.chat_completion(
             messages,
             max_tokens=max_tokens,
@@ -153,42 +65,18 @@ def respond(message, history: list[tuple[str, str]], system_message, max_tokens,
             token = chunk.choices[0].delta.content
             response += token
             yield response
-
     except Exception as e:
-        print(f"An error occurred: {e}")
+        print(f"Error: {e}")
         yield "I'm sorry, I encountered an error. Please try again."
 
-# Build a Gradio chat interface
+# Build Gradio UI
 demo = gr.ChatInterface(
     fn=respond,
     additional_inputs=[
-        gr.Textbox(
-            value=default_nvc_prompt_template,
-            label="System message",
-            visible=True,
-            lines=10,
-        ),
-        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.Textbox(value=default_nvc_prompt_template, label="System message", lines=10),
+        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)"),
     ],
 )