app.py

import pandas as pd
import gradio as gr
from datasets import load_dataset
from langchain.document_loaders import DataFrameLoader
from langchain.text_splitter import CharacterTextSplitter
from langchain.vectorstores import FAISS
from langchain.embeddings import OpenAIEmbeddings
from langchain.chains import RetrievalQA
from langchain.chat_models import ChatOpenAI
from langchain_mistralai.chat_models import ChatMistralAI

dataset_names = ["zelros/pj-ce", "zelros/pj-da", 
                 "zelros/pj-groupama", "zelros/pj-sg", 
                 "zelros/pj-lbp", "zelros/pj-maif", "zelros/pj-ca"]

insurers = ["Caisse d'Epargne","Direct Assurance","Groupama",
            "Société Générale","La Banque Postale", "MAIF", "Crédit Agricole"]

db_dict = {}

def llm_response(question, insurer1, insurer2, gpt):
    if(gpt == "mistral-large-latest"):
        llm1 = ChatMistralAI(model=gpt)
        llm2 = ChatMistralAI(model=gpt)
    else:
        llm1 = ChatOpenAI(model_name=gpt)
        llm2 = ChatOpenAI(model_name=gpt)
    qa_chain1 = RetrievalQA.from_chain_type(llm1, retriever=db_dict[insurer1].as_retriever(search_kwargs={'k': 8}))
    qa_chain2 = RetrievalQA.from_chain_type(llm2, retriever=db_dict[insurer2].as_retriever(search_kwargs={'k': 8}))
    return qa_chain1({"query": question})['result'], qa_chain2({"query": question})['result']

examples = [
    ["Qui contacter en cas besoin ?", None, None, None],
    ["Les problèmes de divorce sont-ils couverts ?", None, None, None],
    ["En cas violences conjugales, puis-je être assisté ?", None, None, None],
    ["Les problèmes d'usurpation d'identité sont-ils couverts ?", None, None, None],
    ["Quel est le montant maximum couvert pour un litige ?", None, None, None],
    ["Quels frais sont pris en charges, et quels frais ne sont pas pris en charge ?", None, None, None],
    ["En tant que membre d'une association, suis-je couvert ?", None, None, None],
    ["J'ai un litige concernant un brevet et un sujet de propriété intellectuelle, suis-je couvert ?", None, None, None],
    ["Quels sont les moments où le contrat peut être résilié ?", None, None, None]
]

for i, name in enumerate(dataset_names):
    dataset = load_dataset(name)

    df = dataset['train'].to_pandas()
    df['text'] = df["title"] + df["content"]

    loader = DataFrameLoader(df, 'text')
    documents = loader.load()
    text_splitter = CharacterTextSplitter(chunk_size=5000, chunk_overlap=0)
    texts = text_splitter.split_documents(documents)
    embeddings = OpenAIEmbeddings()

    db = FAISS.from_documents(texts, embeddings)
    db_dict[insurers[i]] = db

demo = gr.Interface(llm_response, 
                    inputs=[
                        gr.Textbox(label="Question",
                                   info="More examples below :)",
                                   value="Quel est le montant maximum couvert pour un litige ?"),
                        gr.Dropdown(choices=insurers, 
                                    label="Insurer 1",
                                    value="Société Générale",
                                    info="More insurers available soon !"),
                        gr.Dropdown(choices=insurers, 
                                    label="Insurer 2",
                                    value="La Banque Postale"),
                        gr.Dropdown(choices=["gpt-4o-mini-2024-07-18","gpt-4o-2024-08-06","mistral-large-latest"], 
                                    label="LLM",
                                    value="gpt-4o-2024-08-06",
                                    info="Compare gpt-4o-mini, gpt-4o, and mistral-large!"),
                    ],
                    outputs=[gr.Textbox(label="Answer insurer 1"), 
                             gr.Textbox(label="Answer insurer 2")
                    ],
                    title='Towards more accessible and inclusive insurances',
                    description='### <center>Compare french legal protection insurances - Research project, non-binding information - Please refer to a professional for advice.',
                    examples=examples,
                    cache_examples=False)

demo.launch()