app.py

import os
import re 

from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.document_loaders import DirectoryLoader, PyPDFLoader

from langchain.vectorstores import Chroma
from langchain.embeddings import SentenceTransformerEmbeddings
from langchain.prompts import PromptTemplate
from langchain.llms import HuggingFaceHub
from langchain.chains import RetrievalQA
from config import HUGGINGFACEHUB_API_TOKEN

from transformers import pipeline

import gradio as gr


os.environ["HUGGINGFACEHUB_API_TOKEN"] = HUGGINGFACEHUB_API_TOKEN

# Vous pouvez choisir parmi les nombreux midèles disponibles sur HugginFace (https://huggingface.co/models)
model_name = "llmware/industry-bert-insurance-v0.1"

def remove_special_characters(string):
        return re.sub(r"\n", " ", string)
    

def RAG_Langchain(query):
    embeddings = SentenceTransformerEmbeddings(model_name=model_name)
    repo_id = "llmware/bling-sheared-llama-1.3b-0.1"

    loader = DirectoryLoader('data/', glob="**/*.pdf", show_progress=True, loader_cls=PyPDFLoader)

    documents = loader.load()
    
    # La taille des chunks est un paramètre important pour la qualité de l'information retrouvée. Il existe plusieurs méthodes
    # pour en choisir la valeur. 
    # L'overlap correspond au nombre de caractères partagés entre un chunk et le chunk suivant
    text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=100)

    texts = text_splitter.split_documents(documents)

    chunk = texts[0]
    chunk.page_content = remove_special_characters(chunk.page_content)

    #Data Preparation
    for chunks in texts:
        chunks.page_content = remove_special_characters(chunks.page_content)

    # On charge tous les documents dans la base de données vectorielle, pour les utiliser ensuite
    vector_stores=Chroma.from_documents(texts, embeddings, collection_metadata = {"hnsw:space": "cosine"}, persist_directory="stores/insurance_cosine")

    #Retrieval
    load_vector_store=Chroma(persist_directory="stores/insurance_cosine", embedding_function=embeddings)

    #On prend pour l'instant k=1, on verra plus tard comment sélectionner les résultats de contexte
    docs = load_vector_store.similarity_search_with_score(query=query, k=1)
    results = {"Score":[],"Content":[],"Metadata":[]};

    for i in docs:
        doc, score = i
        #print({"Score":score, "Content":doc.page_content, "Metadata":doc.metadata})
        results['Score'].append(score)
        results['Content'].append(doc.page_content)
        results['Metadata'].append(doc.metadata)
        
    context = results['Content']
    
    return results
    

def generateResponseBasedOnContext(model_name, context_string, query):
    
    question_answerer = pipeline("question-answering", model=model_name)
    context_prompt = "You are a sports expert. Answer the user's question by using following context: "

    context = context_prompt + context_string
    print("context : ", context)

    result = question_answerer(question=query, context=context)
    return result['answer']


def gradio_adapted_RAG(model_name, query):
    context = str(RAG_Langchain(query)['Content'])
    generated_answer = generateResponseBasedOnContext(str(model_name), 
                               context, 
                               query)
    return generated_answer


dropdown = gr.Dropdown(choices=["distilbert-base-uncased-distilled-squad",
                                       "impira/layoutlm-document-qa", 
                                       "impira/layoutlm-invoices"], label="Choose a model")

iface = gr.Interface(fn=gradio_adapted_RAG, inputs=[dropdown, "text"], outputs="text")
iface.launch()