Update app.py

app.py

@@ -1,12 +1,15 @@
 import gradio as gr
 from huggingface_hub import InferenceClient
-from transformers import AutoTokenizer  # Import the tokenizer
+from transformers import AutoTokenizer
 
-# Import the tokenizer - No need to import twice, remove the second import
+# Load tokenizer and inference client
 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 this based on your model!
+
+# Define a maximum context length (tokens); adjust based on your model
+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.
@@ -72,87 +75,113 @@ default_nvc_prompt_template = r"""<|system|>You are Roos, an NVC (Nonviolent Com
         - “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.”</s>"""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.
+        - “Thank you for sharing with me. If you’d like to continue this conversation later, I’m here to help.”"""
+
+def count_tokens(text: str) -> int:
+    """Counts the number of tokens in a given string by encoding with the tokenizer."""
+    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/assistant tuples).
+        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.
+        The truncated history as a list of (user_msg, assistant_msg) tuples.
     """
     truncated_history = []
     system_message_tokens = count_tokens(system_message)
     current_length = system_message_tokens
-    # Iterate backwards through the history (newest to oldest)
+
+    # 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))  # Add to the beginning
+            truncated_history.insert(0, (user_msg, assistant_msg))
             current_length += turn_tokens
         else:
-            break  # Stop adding turns if we exceed the limit
-    return truncated_historydef respond(
-    message,
-    history: list[tuple[str, str]],
-    system_message,  # System message is now an argument
-    max_tokens,
-    temperature,
-    top_p,
-):
-    """Responds to a user message, maintaining conversation history, using special tokens and message list."""
-    if message.lower() == "clear memory": # Check for the clear memory command
-        return "", []  # Return empty message and empty history to reset the chat
-
-    formatted_system_message = system_message  # Use the system_message argument
-    truncated_history = truncate_history(history, formatted_system_message, MAX_CONTEXT_LENGTH - max_tokens - 100) # Reserve space for the new message and some generation
-
-    messages = [{"role": "system", "content": formatted_system_message}] # Start with system message as before
+            break
+
+    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
+    if message.lower() == "clear memory":
+        return "", []
+
+    # 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}]
+
     for user_msg, assistant_msg in truncated_history:
         if user_msg:
-            messages.append({"role": "user", "content": user_msg}) # Format history user message - Removed extra tags
+            messages.append({"role": "user", "content": user_msg})
         if assistant_msg:
-            messages.append({"role": "assistant", "content": assistant_msg}) # Format history assistant message - Removed extra tags
-    messages.append({"role": "user", "content": message}) # Format current user message - Removed extra tags
+            messages.append({"role": "assistant", "content": assistant_msg})
+
+    # Add the new user message
+    messages.append({"role": "user", "content": message})
 
     response = ""
     try:
-      for chunk in client.chat_completion(
-          messages, # Send the messages list again, but with formatted content
-          max_tokens=max_tokens,
-          stream=True,
-          temperature=temperature,
-          top_p=top_p,
-      ):
-          token = chunk.choices[0].delta.content
-          response += token
-          # Post-processing to remove prefixes (example - add to your existing yield) - Solution 3 (Fallback)
-          processed_response = response.replace("User:", "").replace("Assistant:", "").replace("Roos:", "").lstrip()
-          yield processed_response
+        # Stream the response
+        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
+            yield response
 
     except Exception as e:
-      print(f"An error occurred: {e}") # It's a good practice add a try-except block
-      yield "I'm sorry, I encountered an error. Please try again."
-
+        print(f"An error occurred: {e}")
+        yield "I'm sorry, I encountered an error. Please try again."
 
-# --- Gradio Interface ---
+# Build a Gradio chat interface
 demo = gr.ChatInterface(
-    respond,
+    fn=respond,
     additional_inputs=[
         gr.Textbox(
             value=default_nvc_prompt_template,
             label="System message",
             visible=True,
-            lines=10,  # Increased height for more space to read the prompt
+            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=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,
@@ -164,4 +193,4 @@ demo = gr.ChatInterface(
 )
 
 if __name__ == "__main__":
-    demo.launch(share=True)
+    demo.launch(share=True)