app.py

## Setup
# Import the necessary Libraries
import json
import gradio as gr
import uuid
import os
import pandas as pd

from openai import OpenAI

from langchain_community.embeddings.sentence_transformer import (
    SentenceTransformerEmbeddings
)
from langchain_community.vectorstores import Chroma

from huggingface_hub import CommitScheduler
from dotenv import load_dotenv
from pathlib import Path

# Create Client


load_dotenv()
os.environ["anyscale_api_key"]=os.getenv("anyscale_api_key")

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

# Define the embedding model and the vectorstore
embedding_model = SentenceTransformerEmbeddings(model_name='thenlper/gte-large')
# Load the persisted vectorDB
collection_name = 'Dataset-10k'

dataset_db = Chroma(
    collection_name=collection_name,
    persist_directory='./dataset_db',
    embedding_function=embedding_model
)

# Prepare the logging functionality

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

scheduler = CommitScheduler(
    repo_id="RAG-10K-log",
    repo_type="dataset",
    folder_path=log_folder,
    path_in_repo="data",
    every=2
)

# Define the Q&A system message

qna_system_message = """
You are an assistant to a financial services firm who answers user queries on annual reports.
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):
    sample = {
    'user_input':user_input,
    'company':company_input
    }

    filter = "dataset/"+company+"-10-k-2023.pdf"

    # Create context_for_query
    relevant_document_chunks = reportsdb.similarity_search(user_question, k=5, filter = {"source":filter})
    context_list = [d.page_content for d in relevant_document_chunks]
    context_for_query = ". ".join(context_list)

    prompt = [
        {'role':'system', 'content': qna_system_message},
        {'role': 'user', 'content': qna_user_message_template.format(
             context=context_for_query,
             question=user_question
            )
        }
    ]

    # Create messages
    try:
      model_name = "mistralai/Mixtral-8x7B-Instruct-v0.1"
      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}'



    # Get response from the LLM

    # 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 [0]

# Set-up the Gradio UI
user_input = gr.Textbox (label = 'Query')
company_input = gr.Radio(
    ['aws','google','IBM','Meta','msft'],
    label = 'company'
    )

model_output = gr.Text (label = 'Response')

# 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.



# Create the interface
# For the inputs parameter of Interface provide [textbox,company]

demo = gr.Interface(
    fn=predict_output,
    inputs=[user_input,company],
    outputs=prediction,
    title="RAG on 10k-reports",
    description="This API allows you to query on annaul reports",
    concurrency_limit=16
)

demo.queue()
demo.launch()