services/chatbot.py

import os
import csv
from services import audio
import random
import pinecone
import gradio as gr
from openai import OpenAI


OPENAI_CLIENT = OpenAI(api_key=os.getenv("OPENAI_API_KEY"))
OPENAI_MODEL = os.getenv("OPENAI_MODEL")
pinecone.init(api_key=os.getenv("PINECONE_API_KEY"), environment=os.getenv("PINECONE_ENV"))
INDEX = pinecone.Index(os.getenv("PINECONE_INDEX"))


def start_chat(client_name: str) -> tuple[list[list[str | None]], gr.helpers, gr.helpers]:
    """
    Initialize chat with greeting text and audio in spanish
    :param client_name: name of the client
    :return: (chat history with greeting, audio with updated file and gradio update with visible=True)
    """
    # Get greeting text and audio, the first one available in spanish
    with open(f'assets/{client_name}/greetings/es.csv', mode='r', encoding='utf-8') as infile:
        reader = csv.reader(infile)
        greeting = next(reader)[0]
    audio_name = f'assets/{client_name}/media/audio/greeting_es_0.wav'

    # Initialize chat
    chat_history = [['', greeting]]

    return chat_history, gr.update(value=f'{audio_name}'), gr.update(visible=True)


def get_random_data(client_name: str) -> gr.helpers:
    """
    Returns an audio with a random data in spanish
    :param client_name: name of the client for this chatbot
    :return: gradio audio updated with a random data from the client
    """
    random_options = []
    path_audios = f'assets/{client_name}/media/audio'
    for random_audio in os.listdir(path_audios):
        if random_audio.startswith('random') and 'es' in random_audio:
            random_options.append(os.path.join(path_audios, random_audio))

    # Get any of the found random files
    num = random.randint(0, len(random_options) - 1)
    return gr.update(value=random_options[num])


def get_answer(
        chat_history: list[tuple[str, str]], user_input: str, client_name: str, general_prompt: str, context_prompt: str
) -> tuple[list[tuple[str, str]], str, gr.helpers]:
    """
    Gets the answer from the chatbot and returns it as an audio and text
    :param chat_history: previous chat history
    :param user_input: user question
    :param client_name: name of the client
    :param general_prompt: prompt used for answering the questions
    :param context_prompt: prompt used for finding the context in the vectorstore
    :return:
    """
    # Format chat history to OpenAI format msg history
    msg_history = [{'role': 'system', 'content': general_prompt}]
    for i, (user, bot) in enumerate(chat_history):
        if i == 0:
            msg_history.append({'role': 'assistant', 'content': bot})
        else:
            msg_history.append({'role': 'user', 'content': user})
            msg_history.append({'role': 'assistant', 'content': bot})

    # Get standalone question
    standalone_question = _get_standalone_question(user_input, msg_history, context_prompt)

    # Get context
    context = _get_context(standalone_question, client_name)

    # Get answer from chatbot
    response = _get_response(context, msg_history, user_input, general_prompt)

    # Get audio:
    audio.get_audio(response, 'es')

    # Update chat_history
    chat_history.append((user_input, response))

    return chat_history, "", gr.update(value='output.wav')


def _get_response(context: str, message_history: list[dict], question: str, prompt: str) -> str:
    """
    Gets the response from ChatGPT
    :param context: text obtained from the vectorstore
    :param message_history: chat history in the format used by OpenAI
    :param question: user question
    :param prompt: prompt used to answer the questions
    :return: response from ChatGPT
    """
    message_history[0]['content'] = prompt.replace('CONTEXT', context)
    message_history.append({'role': 'user', 'content': question})
    return _call_api(message_history, 0.7)


def _get_embedding(text: str) -> list[float]:
    """
    Gets the embedding of a given text
    :param text: input text
    :return: embedding of the text
    """
    response = OPENAI_CLIENT.embeddings.create(
        input=text,
        model='text-embedding-ada-002'
    )
    return response.data[0].embedding


def _call_api(message_history: list[dict], temperature: float) -> str:
    """
    Gets response form OpenAI API
    :param message_history: chat history in the format used by OpenAI
    :param temperature: randomness of the output
    :return: ChatGPT answer
    """
    response = OPENAI_CLIENT.chat.completions.create(
        model=OPENAI_MODEL,
        temperature=temperature,
        messages=message_history
    )
    return response.choices[0].message.content


def _get_standalone_question(question: str, message_history: list[dict], prompt_q: str) -> str:
    """
    Gets a standalone question/phrase based on the user's question and the previous messages. Used since
    some questions are too simple like "yes, please"
    :param question: user question
    :param message_history: msg history in the format used by OpenAI
    :param prompt_q: prompt used to get a text that will be used in the vectorstore
    :return: string with the standalone phrase
    """
    # Format the message history like: Human: blablablá \nAssistant: blablablá
    history = ''
    for i, msg in enumerate(message_history):
        if i == 0:
            continue  # Omit the prompt
        if i % 2 == 0:
            history += f'Human: {msg["content"]}\n'
        else:
            history += f'Assistant: {msg["content"]}\n'

    # Add history and question to the prompt and call chatgpt
    prompt = [{'role': 'system', 'content': ''}]
    content = prompt_q.replace('HISTORY', history).replace('QUESTION', question)
    prompt[0]['content'] = content

    return _call_api(prompt, 0.01)


def _get_context(question: str, client_name: str) -> str:
    """
    Gets the 10 nearest vectors to the given question
    :param question: standalone text
    :param client_name: name of the client, used as namespace in the vectorstore
    :return: formatted text with the nearest vectors
    """
    q_embedding = _get_embedding(question)

    # Get most similar vectors
    result = INDEX.query(
        vector=q_embedding,
        top_k=10,
        include_metadata=True,
        namespace=f'{client_name}-context'
    )['matches']

    # Crete a string based on the text of each vector
    context = ''
    for r in result:
        context += r['metadata']['Text'] + '\n'
    return context