app.py

import json
import random

from smolagents import CodeAgent,DuckDuckGoSearchTool, HfApiModel,load_tool,tool
import datetime
import requests
import pytz
import yaml
import numpy as np

from tools.QCMTool import QCMTool
from tools.final_answer import FinalAnswerTool
from tools.visit_webpage import VisitWebpageTool
from tools.web_search import DuckDuckGoSearchTool
from typing import Optional, Tuple

from Gradio_UI import GradioUI

# Below is an example of a tool that does nothing. Amaze us with your creativity !
@tool
def calculate_risk_metrics(

    returns: np.ndarray,

    var_level: float = 0.95,

    n_simulations: int = 10000,

    bootstrap: bool = False,

    random_seed: Optional[int] = None

) -> Tuple[float, float]:
    """

    Calculate Value at Risk (VaR) and Expected Shortfall (ES) using the historical method, with an option for bootstrapped historical simulation.



    Args:

        returns: Array of daily returns. Each value represents the percentage return for a single day.

        var_level: VaR level (e.g., 0.95 for 95% confidence). Defaults to 0.95.

        n_simulations: Number of bootstrap simulations. Defaults to 10000.

        bootstrap: If True, use bootstrapped historical simulation. Defaults to False.

        random_seed: Seed for random number generation to ensure reproducibility. Defaults to None.



    Returns:

        Tuple[float, float]: A tuple containing the VaR and Expected Shortfall (ES) values.

    """
    if random_seed is not None:
        np.random.seed(random_seed)

    if bootstrap:
        # Perform bootstrapped historical simulation
        simulated_var = np.zeros(n_simulations)
        simulated_es = np.zeros(n_simulations)

        for i in range(n_simulations):
            # Resample returns with replacement
            resampled_returns = np.random.choice(returns, size=len(returns), replace=True)
            # Sort the resampled returns
            sorted_returns = np.sort(resampled_returns)
            # Determine the index for the VaR level
            index = int((1 - var_level) * len(sorted_returns))
            # Calculate VaR for this simulation
            simulated_var[i] = sorted_returns[index]
            # Calculate ES for this simulation (average of returns below VaR)
            simulated_es[i] = np.mean(sorted_returns[:index])

        # Calculate the average VaR and ES across all simulations
        var_value = np.mean(simulated_var)
        es_value = np.mean(simulated_es)
    else:
        # Use the standard historical method
        sorted_returns = np.sort(returns)
        index = int((1 - var_level) * len(returns))
        # Calculate VaR
        var_value = sorted_returns[index]
        # Calculate ES (average of returns below VaR)
        es_value = np.mean(sorted_returns[:index])

    return var_value, es_value

@tool
def provide_my_information(query: str) -> str:
    """

    Provide information about me (Tianqing LIU)based on the user's query.



    Args:

        query: The user's question or request for information.



    Returns:

        str: A response containing the requested information.

    """
    # Convert the query to lowercase for case-insensitive matching
    query = query.lower()
    my_info = None
    with open("info/info.json", 'r') as file:
        my_info = json.load(file)
    # Check for specific keywords in the query and return the corresponding information
    if "who"  in query or "about" in query or "introduce" in query or "presentation" in query:
        return f" {my_info['introduction']}."
    if "name" in query:
        return f"My name is {my_info['name']}."
    elif "location" in query:
        return f"I am located in {my_info['location']}."
    elif "occupation" in query or "job" in query or "work" in query:
        return f"I work as a {my_info['occupation']}."
    elif "education" in query or "educational" in query:
        return f"I have a {my_info['education']}."
    elif "skills" in query or "what can you do" in query:
        return f"My skills include: {', '.join(my_info['skills'])}."
    elif "hobbies" in query or "interests" in query:
        return f"My hobbies are: {', '.join(my_info['hobbies'])}."
    elif "contact" in query or "email" in query or "linkedin" in query or "github" in query or "cv" in query or "resume" in query:
        contact_info = my_info["contact"]
        return (
            f"You can contact me via email at {contact_info['email']}, "
            f"connect with me on LinkedIn at {contact_info['linkedin']}, "
            f"or check out my GitHub profile at {contact_info['github']}."
            f"or check out my website at {contact_info['website']}."
        )
    else:
        return "I'm sorry, I don't have information on that. Please ask about my name, location, occupation, education, skills, hobbies, or contact details."


qcm_tool = QCMTool("info/questions.json")
final_answer = FinalAnswerTool()
visit_webpage = VisitWebpageTool()
web_search = DuckDuckGoSearchTool()

# If the agent does not answer, the model is overloaded, please use another model or the following Hugging Face Endpoint that also contains qwen2.5 coder:
# model_id='https://pflgm2locj2t89co.us-east-1.aws.endpoints.huggingface.cloud' 

model = HfApiModel(
    max_tokens=2096,
    temperature=0.5,
    model_id='Qwen/Qwen2.5-Coder-32B-Instruct',  # it is possible that this model may be overloaded
    custom_role_conversions=None,
)


# Import tool from Hub
#image_generation_tool = load_tool("agents-course/text-to-image", trust_remote_code=True)

with open("prompts.yaml", 'r') as stream:
    prompt_templates = yaml.safe_load(stream)
    
agent = CodeAgent(
    model=model,
    tools=[final_answer,calculate_risk_metrics,qcm_tool,visit_webpage,web_search,provide_my_information], ## add your tools here (don't remove final answer)
    max_steps=6,
    verbosity_level=1,
    grammar=None,
    planning_interval=None,
    name=None,
    description=None,
    prompt_templates=prompt_templates
)


GradioUI(agent).launch()