helper/utils.py

import os
from datetime import datetime
import json
from typing import Any, Dict, List, Tuple, Union
import requests

import numpy as np
import pandas as pd
import PyPDF2
from openai import OpenAI
from together import Together


# Credit
def current_year():
    now = datetime.now()
    return now.year


def read_and_textify(
    files: List[str], chunk_size: int = 2  # Default chunk size set to 50
) -> Tuple[List[str], List[str]]:
    """
    Reads PDF files and extracts text from each page, breaking the text into specified segments.

    This function iterates over a list of uploaded PDF files, extracts text from each page,
    and compiles a list of texts and corresponding source information, segmented into smaller parts
    of approximately 'chunk_size' words each.

    Args:
    files (List[st.uploaded_file_manager.UploadedFile]): A list of uploaded PDF files.
    chunk_size (int): The number of words per text segment. Default is 50.

    Returns:
    Tuple[List[str], List[str]]: A tuple containing two lists:
        1. A list of strings, where each string is a segment of text extracted from a PDF page.
        2. A list of strings indicating the source of each text segment (file name, page number, and segment number).
    """

    text_list = []  # List to store extracted text segments
    sources_list = []  # List to store source information

    # Iterate over each file
    for file in files:
        pdfReader = PyPDF2.PdfReader(file)  # Create a PDF reader object
        # Iterate over each page in the PDF
        for i in range(len(pdfReader.pages)):
            pageObj = pdfReader.pages[i]  # Get the page object
            text = pageObj.extract_text()  # Extract text from the page
            if text:
                # Split text into chunks of approximately 'chunk_size' words
                words = text.split(". ")
                for j in range(0, len(words), chunk_size):
                    chunk = ". ".join(words[j : j + chunk_size]) + "."
                    text_list.append(chunk)
                    # Create a source identifier for each chunk and add it to the list
                    sources_list.append(f"{file.name}_page_{i}_chunk_{j // chunk_size}")
            else:
                # If no text extracted, still add a placeholder
                text_list.append("")
                sources_list.append(f"{file.name}_page_{i}_chunk_0")
            pageObj.clear()  # Clear the page object (optional, for memory management)

    return text_list, sources_list


def read_and_textify_advanced(
    files: List[str], chunk_size: int = 2  # Default chunk size set to 50
) -> Tuple[List[str], List[str]]:
    """
    Reads PDF files and extracts text from each page, breaking the text into specified segments.

    This function iterates over a list of uploaded PDF files, extracts text from each page,
    and compiles a list of texts and corresponding source information, segmented into smaller parts
    of approximately 'chunk_size' words each.

    Args:
    files (List[st.uploaded_file_manager.UploadedFile]): A list of uploaded PDF files.
    chunk_size (int): The number of words per text segment. Default is 50.

    Returns:
    Tuple[List[str], List[str]]: A tuple containing two lists:
        1. A list of strings, where each string is a segment of text extracted from a PDF page.
        2. A list of strings indicating the source of each text segment (file name, page number, and segment number).
    """

    text_list = []  # List to store extracted text segments
    sources_list = []  # List to store source information

    # Iterate over each file
    for file in files:
        pdfReader = PyPDF2.PdfReader(file)  # Create a PDF reader object
        # Iterate over each page in the PDF
        for i in range(len(pdfReader.pages)):
            pageObj = pdfReader.pages[i]  # Get the page object
            text = pageObj.extract_text()  # Extract text from the page
            if text:
                # Split text into chunks of approximately 'chunk_size' words
                words = text.split(". ")
                for j in range(len(words)):
                    # Get the chunk of text from j-chunk_size to j+chunk_size
                    start = max(0, j - chunk_size)
                    end = min(len(words), j + chunk_size + 1)
                    chunk = ". ".join(words[start:end]) + '.'
                    text_list.append(chunk)
                    # Create a source identifier for each chunk and add it to the list
                    sources_list.append(f"{file.name}_page_{i}_chunk_{j}")
            else:
                # If no text extracted, still add a placeholder
                text_list.append("")
                sources_list.append(f"{file.name}_page_{i}_chunk_0")
            pageObj.clear()  # Clear the page object (optional, for memory management)

    return text_list, sources_list


openai_client = OpenAI(api_key=os.environ["OPENAI_API_KEY"])


def list_to_nums(sentences: List[str]) -> List[List[float]]:
    """
    Converts a list of sentences into a list of numerical embeddings using OpenAI's embedding model.

    Args:
    - sentences (List[str]): A list of sentences (strings).

    Returns:
    - List[List[float]]: A list of lists of numerical embeddings.
    """

    # Initialize the list to store embeddings
    embeddings = []

    # Loop through each sentence to convert to embeddings
    for sentence in sentences:
        # Use the OpenAI API to get embeddings for the sentence

        response = openai_client.embeddings.create(
            input=sentence, model="text-embedding-3-small"
        )

        embeddings.append(response.data[0].embedding)

    return embeddings


def call_gpt(prompt: str, content: str) -> str:
    """
    Sends a structured conversation context including a system prompt, user prompt,
    and additional background content to the GPT-3.5-turbo model for a response.

    This function is responsible for generating an AI-powered response by interacting
    with the OpenAI API. It puts together a preset system message, a formatted user query,
    and additional background information before requesting the completion from the model.

    Args:
        prompt (str): The main question or topic that the user wants to address.
        content (str): Additional background information or details relevant to the prompt.

    Returns:
        str: The generated response from the GPT model based on the given prompts and content.

    Note: 'openai_client' is assumed to be an already created and authenticated instance of the OpenAI
          openai_client, which should be set up prior to calling this function.
    """

    # Generates a response from the model based on the interactive messages provided
    response = openai_client.chat.completions.create(
        model="gpt-3.5-turbo",  # The AI model being queried for a response
        messages=[
            # System message defining the assistant's role
            {"role": "system", "content": "You are a helpful assistant."},
            # User message containing the prompt
            {"role": "user", "content": f"I want to ask you a question: {prompt}"},
            # Assistant message asking for background content
            {"role": "assistant", "content": "What is the background content?"},
            # User providing the background content
            {"role": "user", "content": content},
        ],
    )

    # Extracts and returns the response content from the model's completion
    return response.choices[0].message.content


together_client = Together(api_key=os.environ["TOGETHER_API_KEY"])


def call_llama(prompt: str) -> str:
    """
    Send a prompt to the Llama model and return the response.
    Args:
        prompt (str): The input prompt to send to the Llama model.
    Returns:
        str: The response from the Llama model.
    """

    # Create a completion request with the prompt
    response = together_client.chat.completions.create(
        # Use the Llama-3-8b-chat-hf model
        model="meta-llama/Llama-3-8b-chat-hf",
        # Define the prompt as a user message
        messages=[{"role": "user", "content": prompt}],  # Use the input prompt
    )

    # Return the content of the first response message
    return response.choices[0].message.content


def call_llama2(prompt: str, max_new_tokens: int = 50, temperature: float = 0.9) -> str:
    """
    Calls the Llama API to generate text based on a given prompt, controlling the length and randomness.

    Args:
        prompt (str): The prompt text to send to the Llama model for text generation.
        max_new_tokens (int, optional): The maximum number of tokens that the model should generate. Defaults to 50.
        temperature (float, optional): Controls the randomness of the output. Lower values make the model more deterministic.
            A higher value increases randomness. Defaults to 0.9.

    Returns:
        str: The generated text response from the Llama model.

    Raises:
        Exception: If the API call fails and returns a non-200 status code, it raises an exception with the error details.
    """
    # API endpoint for the Llama model
    api_url = "https://v6rkdcyir7.execute-api.us-east-1.amazonaws.com/beta"

    # Configuration for the request body
    json_body = {
        "body": {
            "inputs": f"<s>[INST] {prompt} [/INST]",
            "parameters": {
                "max_new_tokens": max_new_tokens,
                "top_p": 0.9,  # Fixed probability cutoff to select tokens with cumulative probability above this threshold
                "temperature": temperature
            }
        }
    }

    # Headers to indicate that the payload is JSON
    headers = {"Content-Type": "application/json"}

    # Perform the POST request to the Llama API
    response = requests.post(api_url, headers=headers, json=json_body)

    # Parse the JSON response
    response_body = response.json()['body']

    # Convert the string response to a JSON object
    body_list = json.loads(response_body)

    # Extract the 'generated_text' from the first item in the list
    generated_text = body_list[0]['generated_text']

    # Separate the answer from the instruction
    answer = generated_text.split("[/INST]")[-1].strip()

    # Check the status code of the response
    if response.status_code == 200:
        return answer  # Return the text generated by the model
    else:
        # Raise an exception if the API did not succeed
        raise Exception(f"Error calling Llama API: {response.status_code}")


def quantize_to_kbit(arr: Union[np.ndarray, Any], k: int = 16) -> np.ndarray:
    """Converts an array to a k-bit representation by normalizing and scaling its values.

    Args:
        arr (Union[np.ndarray, Any]): The input array to be quantized.
        k (int): The number of levels to quantize to. Defaults to 16 for 4-bit quantization.
    Returns:
        np.ndarray: The quantized array with values scaled to 0 to k-1.
    """
    if not isinstance(arr, np.ndarray):  # Check if input is not a numpy array
        arr = np.array(arr)  # Convert input to a numpy array
    arr_min = arr.min()  # Calculate the minimum value in the array
    arr_max = arr.max()  # Calculate the maximum value in the array
    normalized_arr = (arr - arr_min) / (
        arr_max - arr_min
    )  # Normalize array values to [0, 1]
    return np.round(normalized_arr * (k - 1)).astype(
        int
    )  # Scale normalized values to 0-(k-1) and convert to integer


def quantized_influence(
    arr1: np.ndarray, arr2: np.ndarray, k: int = 16, use_dagger: bool = False
) -> Tuple[float, List[float]]:
    """
    Calculates a weighted measure of influence based on quantized version of input arrays and optionally applies a transformation.

    Args:
        arr1 (np.ndarray): First input array to be quantized and analyzed.
        arr2 (np.ndarray): Second input array to be quantized and used for influence measurement.
        k (int): The quantization level, defaults to 16 for 4-bit quantization.
        use_dagger (bool): Flag to apply a transformation based on local averages, defaults to False.
    Returns:
        Tuple[float, List[float]]: A tuple containing the quantized influence measure and an optional list of transformed values based on local estimates.
    """
    # Quantize both arrays to k levels
    arr1_quantized = quantize_to_kbit(arr1, k)
    arr2_quantized = quantize_to_kbit(arr2, k)

    # Find unique quantized values in arr1
    unique_values = np.unique(arr1_quantized)

    # Compute the global average of quantized arr2
    total_samples = len(arr2_quantized)
    y_bar_global = np.mean(arr2_quantized)

    # Compute weighted local averages and normalize
    weighted_local_averages = [
        (np.mean(arr2_quantized[arr1_quantized == val]) - y_bar_global) ** 2
        * len(arr2_quantized[arr1_quantized == val]) ** 2
        for val in unique_values
    ]
    qim = np.sum(weighted_local_averages) / (
        total_samples * np.std(arr2_quantized)
    )  # Calculate the quantized influence measure

    if use_dagger:
        # If use_dagger is True, compute local estimates and map them to unique quantized values
        local_estimates = [
            np.mean(arr2_quantized[arr1_quantized == val]) for val in unique_values
        ]
        daggers = {
            unique_values[i]: v for i, v in enumerate(local_estimates)
        }  # Map unique values to local estimates

        def find_val_(i: int) -> float:
            """Helper function to map quantized values to their local estimates."""
            return daggers[i]

        # Apply transformation based on local estimates
        daggered_values = list(map(find_val_, arr1_quantized))
        return qim, daggered_values
    else:
        # If use_dagger is False, return the original quantized arr1 values
        daggered_values = arr1_quantized.tolist()
        return qim


def query_search(
    prompt: str,
    sentences: list[str],
    query_database: list[list[float]],
    sources: list[str],
    levels: int,
) -> pd.DataFrame:
    """
    Takes a text prompt and searches a predefined database by converting the prompt
    and database entries to embeddings, and then calculating a quantized influence metric.

    Args:
    - prompt (str): A text prompt to search for in the database.

    Returns:
    - pd.DataFrame: A pandas DataFrame sorted by the quantized influence metric in descending order.
                     The DataFrame contains the original sentences, their embeddings, and the computed scores.
    """
    # Convert the prompt to its numerical embedding
    prompt_embed_ = list_to_nums([prompt])

    # Calculate scores for each item in the database using the quantized influence metric
    scores = [
        [
            sentences[i],  # The sentence itself
            # query_database[i],  # Embedding of the sentence
            sources[i],  # Source of the sentence
            quantized_influence(
                prompt_embed_[0], query_database[i], k=levels, use_dagger=False
            ),  # Score calculation
        ]
        for i in range(len(query_database))
    ]

    # Convert the list of scores into a DataFrame
    refs = pd.DataFrame(scores)
    # Rename columns for clarity
    refs = refs.rename(
        # columns={0: "sentences", 1: "query_embeddings", 2: "page no", 3: "qim"}
        columns={0: "sentences", 1: "page no", 2: "qim"}
    )
    # Sort the DataFrame based on the 'qim' score in descending order
    refs = refs.sort_values(by="qim", ascending=False)

    return refs