app.py

import gradio as gr
import json
import os
import numexpr
from groq import Groq
from groq.types.chat.chat_completion_tool_param import ChatCompletionToolParam

MODEL = "llama3-groq-70b-8192-tool-use-preview"
client = Groq(api_key=os.environ["GROQ_API_KEY"])


def evaluate_math_expression(expression: str):
    return json.dumps(numexpr.evaluate(expression).tolist())


calculator_tool: ChatCompletionToolParam = {
    "type": "function",
    "function": {
        "name": "evaluate_math_expression",
        "description": "Calculator tool: use this for evaluating numeric expressions with Python. Ensure the expression is valid Python syntax (e.g., use '**' for exponentiation, not '^').",
        "parameters": {
            "type": "object",
            "properties": {
                "expression": {
                    "type": "string",
                    "description": "The mathematical expression to evaluate. Must be valid Python syntax.",
                },
            },
            "required": ["expression"],
        },
    },
}

tools = [calculator_tool]


def call_function(tool_call, available_functions):
    function_name = tool_call.function.name
    if function_name not in available_functions:
        return {
            "tool_call_id": tool_call.id,
            "role": "tool",
            "content": f"Function {function_name} does not exist.",
        }
    function_to_call = available_functions[function_name]
    function_args = json.loads(tool_call.function.arguments)
    function_response = function_to_call(**function_args)
    return {
        "tool_call_id": tool_call.id,
        "role": "tool",
        "name": function_name,
        "content": json.dumps(function_response),
    }


def get_model_response(messages, inner_messages, message, system_message):
    messages_for_model = []
    for msg in messages:
        native_messages = msg.get("metadata", {}).get("native_messages", [msg])
        if isinstance(native_messages, list):
            messages_for_model.extend(native_messages)
        else:
            messages_for_model.append(native_messages)

    messages_for_model.insert(
        0,
        {
            "role": "system",
            "content": system_message,
        },
    )
    messages_for_model.append(
        {
            "role": "user",
            "content": message,
        }
    )
    messages_for_model.extend(inner_messages)

    try:
        return client.chat.completions.create(
            model=MODEL,
            messages=messages_for_model,
            tools=tools,
            temperature=0.5,
            top_p=0.65,
            max_tokens=4096,
        )
    except Exception as e:
        print(f"An error occurred while getting model response: {str(e)}")
        print(messages_for_model)
        return None


def respond(message, history, system_message):
    inner_history = []

    available_functions = {
        "evaluate_math_expression": evaluate_math_expression,
    }

    assistant_content = ""
    assistant_native_message_list = []

    while True:
        response_message = (
            get_model_response(history, inner_history, message, system_message)
            .choices[0]
            .message
        )

        if not response_message.tool_calls and response_message.content is not None:
            break

        if response_message.tool_calls is not None:
            assistant_native_message_list.append(response_message)
            inner_history.append(response_message)

            assistant_content += (
                "```json\n"
                + json.dumps(
                    [
                        tool_call.model_dump()
                        for tool_call in response_message.tool_calls
                    ],
                    indent=2,
                )
                + "\n```\n"
            )
            assistant_message = {
                "role": "assistant",
                "content": assistant_content,
                "metadata": {"native_messages": assistant_native_message_list},
            }

            yield assistant_message

            for tool_call in response_message.tool_calls:
                function_response = call_function(tool_call, available_functions)
                assistant_content += (
                    "```json\n"
                    + json.dumps(
                        {
                            "name": tool_call.function.name,
                            "arguments": json.loads(tool_call.function.arguments),
                            "response": json.loads(function_response["content"]),
                        },
                        indent=2,
                    )
                    + "\n```\n"
                )
                native_tool_message = {
                    "tool_call_id": tool_call.id,
                    "role": "tool",
                    "content": function_response["content"],
                }
                assistant_native_message_list.append(
                    native_tool_message
                )
                tool_message = {
                    "role": "assistant",
                    "content": assistant_content,
                    "metadata": {"native_messages": assistant_native_message_list},
                }
                yield tool_message
                inner_history.append(native_tool_message)

    assistant_content += response_message.content
    assistant_native_message_list.append(response_message)

    final_message = {
        "role": "assistant",
        "content": assistant_content,
        "metadata": {"native_messages": assistant_native_message_list},
    }
    
    yield final_message

system_prompt = "You are a friendly Chatbot with access to a calculator. Don't mention that we are using functions defined in Python."

demo = gr.ChatInterface(
    respond,
    additional_inputs=[
        gr.Textbox(
            value=system_prompt,
            label="System message",
        ),
    ],
    type="messages",
    title="Groq Tool Use Chat",
    description="This chatbot uses the `llama3-groq-70b-8192-tool-use-preview` LLM with tool use capabilities, including a calculator function.",
)

if __name__ == "__main__":
    demo.launch()