app.py

## Setup
# Import the necessary Libraries
import json
import tiktoken 
import pandas as pd
from openai import OpenAI
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_community.document_loaders import PyPDFDirectoryLoader
from langchain_community.embeddings.sentence_transformer import (
    SentenceTransformerEmbeddings
)
from langchain_community.vectorstores import Chroma

import os
import uuid
import joblib
import json

import gradio as gr
from dotenv import load_dotenv

from huggingface_hub import CommitScheduler
from pathlib import Path

# Create Client
# load_dotenv()

# os.environ['API_KEY_PROJ3'] = os.getenv('API_KEY_PROJ3')

# Create Client
client = OpenAI(
    base_url="https://api.endpoints.anyscale.com/v1",
    api_key=os.environ['Anyscale_Colab_key2']
)

# Define the embedding model and the vectorstore
embedding_model = SentenceTransformerEmbeddings(model_name='thenlper/gte-large')

collection_name = 'collection'

# Load the persisted vectorDB
vectorstore_persisted = Chroma(
    collection_name=collection_name,
    persist_directory='./proj3_db',
    embedding_function=embedding_model
)

retriever = vectorstore_persisted.as_retriever(
    search_type = 'similarity',
    search_kargs = {'k':5}
)

# persisted_vectordb_location = './proj3_db'

# Prepare the logging functionality

log_file = Path("logs/") / f"data_{uuid.uuid4()}.json"
log_folder = log_file.parent

scheduler = CommitScheduler(
    repo_id="mgchavez/Finsights_Grey",
    repo_type="dataset",
    folder_path=log_folder,
    path_in_repo="data",
    every=2
)

# Define the Q&A system message
qna_system_message = """
               
        User input will have the context required by you to answer user questions.
        This context will begin with the token: ###Context
        The context contains references to specific portions of a document relevant to the user query.
        
        User questions will begin with the token: ###Question
        
        Please answer only using the context provided in the input. Do not mention anything about the context in your final answer.
        
        If the answer is not found in the context, respond "I don't know".
"""

# Define the user message template
qna_user_message_template = """
###Context
Here are some documents that are relevant to the question mentioned below.
{context}

###Question
{question}
"""


# Define the predict function that runs when 'Submit' is clicked or when a API request is made
def predict(user_input, company):

    filter = "dataset/"+company+"-10-k-2023.pdf"
    relevant_document_chunks = vectorstore_persisted.similarity_search(user_input, k=5, filter={"source":filter})

    # Create context_for_query
    context_for_query = ". ".join(relevant_document_chunks)

    # Create messages
    prompt = [
        {'role': 'system', 'content': qna_system_message},
        {'role': 'user', 'content': qna_user_message_template.format(
            context=context_for_query,
            question=user_input
        )
         }
    ]
    # model_name = 'mlabonne/NeuralHermes-2.5-Mistral-7B'
    model_name = 'mistralai/Mixtral-8x7B-Instruct-v0.1'
    # model_name = 'thenlper/gte-large'
    # Get response from the LLM
    try:
        response = client.chat.completions.create(
            model=model_name,
            messages=prompt,
            temperature=0
        )

        prediction = response.choices[0].message.content.strip()
    except Exception as e:
        prediction = f'Sorry, I encountered the following error: \n {e}'

    # While the prediction is made, log both the inputs and outputs to a local log file
    # While writing to the log file, ensure that the commit scheduler is locked to avoid parallel
    # access

    with scheduler.lock:
        with log_file.open("a") as f:
            f.write(json.dumps(
                {
                    'user_input': user_input,
                    'retrieved_context': context_for_query,
                    'model_response': prediction
                }
            ))
            f.write("\n")

    return prediction


# Set-up the Gradio UI
# Add text box and radio button to the interface
# The radio button is used to select the company 10k report in which the context needs to be retrieved.
lst_companies = ['aws', 'google', 'IBM', 'Meta', 'msft']
textbox = gr.Textbox('Input user')
company = gr.Radio(lst_companies, label='Company')

model_output = gr.Label(label="Charge predictor")

# Create the interface
# For the inputs parameter of Interface provide [textbox,company]
demo = gr.Interface(
    fn=predict,
    inputs=[textbox, company],
    outputs=model_output,
    title="Charge Predictor",
    description="This API allows you to predict the charge of insurace",
    allow_flagging="auto",
    concurrency_limit=8
)

demo.queue()
demo.launch()