app.py

import argparse
import logging
from pathlib import Path
from threading import Thread
from typing import List
import os
import gradio as gr

logger = logging.getLogger(__name__)

COMET_API_KEY = os.getenv("COMET_API_KEY")
COMET_WORKSPACE = os.getenv("COMET_WORKSPACE")
COMET_PROJECT_NAME = os.getenv("COMET_PROJECT_NAME")
QDRANT_URL = os.getenv("QDRANT_URL")
QDRANT_API_KEY = os.getenv("QDRANT_API_KEY")

def parseargs() -> argparse.Namespace:
    """
    Parses command line arguments for the Financial Assistant Bot.

    Returns:
        argparse.Namespace: An object containing the parsed arguments.
    """

    parser = argparse.ArgumentParser(description="Financial Assistant Bot")

    parser.add_argument(
        "--env-file-path",
        type=str,
        default=".env",
        help="Path to the environment file",
    )

    parser.add_argument(
        "--logging-config-path",
        type=str,
        default="logging.yaml",
        help="Path to the logging configuration file",
    )

    parser.add_argument(
        "--model-cache-dir",
        type=str,
        default="./model_cache",
        help="Path to the directory where the model cache will be stored",
    )

    parser.add_argument(
        "--embedding-model-device",
        type=str,
        default="cuda:0",
        help="Device to use for the embedding model (e.g. 'cpu', 'cuda:0', etc.)",
    )

    parser.add_argument(
        "--debug",
        action="store_true",
        default=False,
        help="Enable debug mode",
    )

    return parser.parse_args()


args = parseargs()


# === Load Bot ===


def load_bot(
#    env_file_path: str = ".env",
    logging_config_path: str = "logging.yaml",
    model_cache_dir: str = "./model_cache",
    embedding_model_device: str = "cuda:0",
    debug: bool = False,
):
    """
    Load the financial assistant bot in production or development mode based on the `debug` flag

    In DEV mode the embedding model runs on CPU and the fine-tuned LLM is mocked.
    Otherwise, the embedding model runs on GPU and the fine-tuned LLM is used.

    Args:
        env_file_path (str): Path to the environment file.
        logging_config_path (str): Path to the logging configuration file.
        model_cache_dir (str): Path to the directory where the model cache is stored.
        embedding_model_device (str): Device to use for the embedding model.
        debug (bool): Flag to indicate whether to run the bot in debug mode or not.

    Returns:
        FinancialBot: An instance of the FinancialBot class.
    """

    from financial_bot import initialize

    # Be sure to initialize the environment variables before importing any other modules.
    # initialize(logging_config_path=logging_config_path, env_file_path=env_file_path)
    initialize(logging_config_path=logging_config_path)

    from financial_bot import utils
    from financial_bot.langchain_bot import FinancialBot

    logger.info("#" * 100)
    utils.log_available_gpu_memory()
    utils.log_available_ram()
    logger.info("#" * 100)

    bot = FinancialBot(
        model_cache_dir=Path(model_cache_dir) if model_cache_dir else None,
        embedding_model_device=embedding_model_device,
        streaming=True,
        debug=debug,
    )

    return bot


bot = load_bot(
#    env_file_path=args.env_file_path,
    logging_config_path=args.logging_config_path,
    model_cache_dir=args.model_cache_dir,
    embedding_model_device=args.embedding_model_device,
    debug=args.debug,
)


# === Gradio Interface ===


def predict(message: str, history: List[List[str]], about_me: str) -> str:
    """
    Predicts a response to a given message using the financial_bot Gradio UI.

    Args:
        message (str): The message to generate a response for.
        history (List[List[str]]): A list of previous conversations.
        about_me (str): A string describing the user.

    Returns:
        str: The generated response.
    """

    generate_kwargs = {
        "about_me": about_me,
        "question": message,
        "to_load_history": history,
    }

    if bot.is_streaming:
        t = Thread(target=bot.answer, kwargs=generate_kwargs)
        t.start()

        for partial_answer in bot.stream_answer():
            yield partial_answer
    else:
        yield bot.answer(**generate_kwargs)


demo = gr.ChatInterface(
    predict,
    textbox=gr.Textbox(
        placeholder="Ask me a financial question",
        label="Financial Question",
        container=False,
        scale=7,
    ),
    additional_inputs=[
        gr.Textbox(
            "I am a 30 year old graphic designer. I want to invest in something with potential for high returns.",
            label="About Me",
        )
    ],
    title="Friendly Financial Bot 🤑",
    description="Ask me any financial or crypto market questions, and I will do my best to provide useful insight. My advice is based on current \
    finance news, stored as embeddings in a Qdrant vector db. I run on a 4bit quantized Mistral-7B-Instruct-v0.2 model with a QLoRa \
    adapter fine-tuned for providing financial guidance. Some sample questions and additional background scenarios are below.",
    theme="soft",
    examples=[
        [
            "Do you think investing in Boeing is a good idea right now?",
            "I'm a 31 year old pilot. I'm curious about the potential of investing in certain airlines.",
        ],
        [
            "What's your opinion on investing in AI-related companies?",
            "I'm a 25 year old entrepreneur interested in emerging technologies. \
             I'm willing to take calculated risks for potential high returns.",
        ],
        [
            "What's your opinion on investing in the Chinese stock market?",
            "I am a risk-averse 40-year old and would like to avoid risky investments.",
        ],
    ],
    cache_examples=False,
    retry_btn=None,
    undo_btn=None,
    clear_btn="Clear",
)


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