ai-agents-masterclass/cost-saving-ai-router/cost-saving-task-agent.py

import asana
from asana.rest import ApiException
from dotenv import load_dotenv
from datetime import datetime
import streamlit as st
import json
import os

from langchain_groq import ChatGroq
from langchain_openai import ChatOpenAI
from langchain_core.tools import tool
from langchain_core.messages import SystemMessage, AIMessage, HumanMessage, ToolMessage

load_dotenv()

groq_model = os.getenv('GROQ_MODEL', 'llama3-groq-70b-8192-tool-use-preview')
openai_model = os.getenv('OPENAI_MODEL', 'gpt-4o')

configuration = asana.Configuration()
configuration.access_token = os.getenv('ASANA_ACCESS_TOKEN', '')
api_client = asana.ApiClient(configuration)

# create an instance of the different Asana API classes
projects_api_instance = asana.ProjectsApi(api_client)
tasks_api_instance = asana.TasksApi(api_client)

workspace_gid = os.getenv("ASANA_WORKPLACE_ID", "")

# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# ~~~~~~~~~~~~~~~~~~~~~ AI Agent Tool Functions ~~~~~~~~~~~~~~~~~~~~~~~~
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

@tool
def create_asana_task(task_name, project_gid, due_on="today"):
    """
    Creates a task in Asana given the name of the task and when it is due

    Example call:

    create_asana_task("Test Task", "2024-06-24")
    Args:
        task_name (str): The name of the task in Asana
        project_gid (str): The ID of the project to add the task to (not the name but the project ID)
        due_on (str): The date the task is due in the format YYYY-MM-DD. If not given, the current day is used
    Returns:
        str: The API response of adding the task to Asana or an error message if the API call threw an error
    """
    if due_on == "today":
        due_on = str(datetime.now().date())

    task_body = {
        "data": {
            "name": task_name,
            "due_on": due_on,
            "projects": [project_gid]
        }
    }

    try:
        api_response = tasks_api_instance.create_task(task_body, {})
        # return json.dumps(api_response, indent=2)
        return "Task created successfully!"
    except ApiException as e:
        return f"Exception when calling TasksApi->create_task: {e}"  

@tool
def get_asana_projects():
    """
    Gets all of the projects in the user's Asana workspace

    Returns:
        str: The API response from getting the projects or an error message if the projects couldn't be fetched.
        The API response is an array of project objects, where each project object looks like:
        {'gid': '1207789085525921', 'name': 'Project Name', 'resource_type': 'project'}
    """    
    opts = {
        'limit': 50, # int | Results per page. The number of objects to return per page. The value must be between 1 and 100.
        'workspace': workspace_gid, # str | The workspace or organization to filter projects on.
        'archived': False # bool | Only return projects whose `archived` field takes on the value of this parameter.
    }

    try:
        api_response = projects_api_instance.get_projects(opts)
        return json.dumps(list(api_response), indent=2)
    except ApiException as e:
        return "Exception when calling ProjectsApi->create_project: %s\n" % e

@tool
def create_asana_project(project_name, due_on=None):
    """
    Creates a project in Asana given the name of the project and optionally when it is due

    Example call:

    create_asana_project("Test Project", "2024-06-24")
    Args:
        project_name (str): The name of the project in Asana
        due_on (str): The date the project is due in the format YYYY-MM-DD. If not supplied, the project is not given a due date
    Returns:
        str: The API response of adding the project to Asana or an error message if the API call threw an error
    """    
    body = {
        "data": {
            "name": project_name, "due_on": due_on, "workspace": workspace_gid
        }
    } # dict | The project to create.

    try:
        # Create a project
        api_response = projects_api_instance.create_project(body, {})
        return json.dumps(api_response, indent=2)
    except ApiException as e:
        return "Exception when calling ProjectsApi->create_project: %s\n" % e  

@tool
def get_asana_tasks(project_gid):
    """
    Gets all the Asana tasks in a project

    Example call:

    get_asana_tasks("1207789085525921")
    Args:
        project_gid (str): The ID of the project in Asana to fetch the tasks for
    Returns:
        str: The API response from fetching the tasks for the project in Asana or an error message if the API call threw an error
        The API response is an array of tasks objects where each task object is in the format:
        {'gid': '1207780961742158', 'created_at': '2024-07-11T16:25:46.380Z', 'due_on': None or date in format "YYYY-MM-DD", 'name': 'Test Task'}
    """        
    opts = {
        'limit': 50, # int | Results per page. The number of objects to return per page. The value must be between 1 and 100.
        'project': project_gid, # str | The project to filter tasks on.
        'opt_fields': "created_at,name,due_on", # list[str] | This endpoint returns a compact resource, which excludes some properties by default. To include those optional properties, set this query parameter to a comma-separated list of the properties you wish to include.
    }

    try:
        # Get multiple tasks
        api_response = tasks_api_instance.get_tasks(opts)
        return json.dumps(list(api_response), indent=2)
    except ApiException as e:
        return "Exception when calling TasksApi->get_tasks: %s\n" % e

@tool
def update_asana_task(task_gid, data):
    """
    Updates a task in Asana by updating one or both of completed and/or the due date

    Example call:

    update_asana_task("1207780961742158", {"completed": True, "due_on": "2024-07-13"})
    Args:
        task_gid (str): The ID of the task to update
        data (dict): A dictionary with either one or both of the keys 'completed' and/or 'due_on'
                    If given, completed needs to be either True or False.
                    If given, the due date needs to be in the format 'YYYY-MM-DD'.
    Returns:
        str: The API response of updating the task or an error message if the API call threw an error
    """      
    # Data: {"completed": True or False, "due_on": "YYYY-MM-DD"}
    body = {"data": data} # dict | The task to update.

    try:
        # Update a task
        api_response = tasks_api_instance.update_task(body, task_gid, {})
        return json.dumps(api_response, indent=2)
    except ApiException as e:
        return "Exception when calling TasksApi->update_task: %s\n" % e

@tool
def delete_task(task_gid):
    """
    Deletes a task in Asana

    Example call:

    delete_task("1207780961742158")
    Args:
        task_gid (str): The ID of the task to delete
    Returns:
        str: The API response of deleting the task or an error message if the API call threw an error
    """        
    try:
        # Delete a task
        api_response = tasks_api_instance.delete_task(task_gid)
        return json.dumps(api_response, indent=2)
    except ApiException as e:
        return "Exception when calling TasksApi->delete_task: %s\n" % e   

# Maps the function names to the actual function object in the script
# This mapping will also be used to create the list of tools to bind to the agent
available_functions = {
    "create_asana_task": create_asana_task,
    "get_asana_projects": get_asana_projects,
    "create_asana_project": create_asana_project,
    "get_asana_tasks": get_asana_tasks,
    "update_asana_task": update_asana_task,
    "delete_task": delete_task
}     


# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# ~~~~~~~~~~~~~~~~~~~~~ AI Prompting Functions ~~~~~~~~~~~~~~~~~~~~~~~~~
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

def get_latest_messages(messages, num_messages=3):
    return "\n\n".join([message.content for message in messages[-num_messages:]])

def decide_model_from_prompt(messages):
    ai_router = ChatGroq(model=groq_model)
    latest_messages = get_latest_messages(messages)

    router_prompt = f"""
        Your only job is to take requests from users (as the last message in chat history), and determine the complexity of the 
        request to route it to a more powerful and more expensive LLM if the request is complicated, and a less powerful
        and cheaper LLM if the request is not complicated.

        A request is complicated if it requires the LLM to take more than one action (create a task, search for tasks, etc.). 
        A request is not complicated if it will involve the LLM taking zero or one action.

        The last three messages in the conversation (ending with the user's message/request) is:

        {latest_messages}

        Output CHEAP if the request is not complicated and can be routed to the cheaper LLM.
        Output EXPENSIVE if the request is complicated (involves more than one action likely)
        and needs to be routed to the more expensive LLM.

        Your output needs to be CHEAP or EXPENSIVE, nothing else.
    """

    response = ai_router.invoke(router_prompt)
    return response.content

def prompt_ai(messages, router_decided_model, nested_calls=0):
    if nested_calls > 5:
        raise "AI is tool calling too much!"

    # First, prompt the AI with the latest user message
    tools = [tool for _, tool in available_functions.items()]
    asana_chatbot = ChatGroq(model=groq_model) if router_decided_model.lower() == "cheap" else ChatOpenAI(model=openai_model)
    asana_chatbot_with_tools = asana_chatbot.bind_tools(tools)

    stream = asana_chatbot_with_tools.stream(messages)
    first = True
    for chunk in stream:
        if first:
            gathered = chunk
            first = False
        else:
            gathered = gathered + chunk

        yield chunk

    has_tool_calls = len(gathered.tool_calls) > 0

    # Second, see if the AI decided it needs to invoke a tool
    if has_tool_calls:
        # Add the tool request to the list of messages so the AI knows later it invoked the tool
        messages.append(gathered)

        # If the AI decided to invoke a tool, invoke it
        # For each tool the AI wanted to call, call it and add the tool result to the list of messages
        for tool_call in gathered.tool_calls:
            tool_name = tool_call["name"].lower()
            selected_tool = available_functions[tool_name]
            tool_output = selected_tool.invoke(tool_call["args"])
            messages.append(ToolMessage(tool_output, tool_call_id=tool_call["id"]))                

        # Call the AI again so it can produce a response with the result of calling the tool(s)
        additional_stream = prompt_ai(messages, router_decided_model, nested_calls + 1)
        for additional_chunk in additional_stream:
            yield additional_chunk


# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# ~~~~~~~~~~~~~~~~~~ Main Function with UI Creation ~~~~~~~~~~~~~~~~~~~~
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

system_message = f"""
You are a personal assistant who helps manage tasks in Asana. 
You never give IDs to the user since those are just for you to keep track of. 
When a user asks to create a task and you don't know the project to add it to for sure, clarify with the user.
The current date is: {datetime.now().date()}
"""

def main():
    st.title("Asana Chatbot (With AI Router)")

    # Initialize chat history
    if "messages" not in st.session_state:
        st.session_state.messages = [
            SystemMessage(content=system_message)
        ]    

    # Display chat messages from history on app rerun
    for message in st.session_state.messages:
        message_json = json.loads(message.json())
        message_type = message_json["type"]
        if message_type in ["human", "ai", "system"]:
            with st.chat_message(message_type):
                st.markdown(message_json["content"])        

    # React to user input
    if prompt := st.chat_input("What would you like to do today?"):
        # Display user message in chat message container
        st.chat_message("user").markdown(prompt)
        # Add user message to chat history
        st.session_state.messages.append(HumanMessage(content=prompt))

        # Display assistant response in chat message container
        with st.chat_message("assistant"):
            router_decided_model = decide_model_from_prompt(st.session_state.messages)
            print(f"Going with {router_decided_model} LLM model...")

            stream = prompt_ai(st.session_state.messages, router_decided_model)
            response = st.write_stream(stream)
        
        st.session_state.messages.append(AIMessage(content=response))


if __name__ == "__main__":
    main()