document_questioner_app.py

import openai
import os
import gradio as gr
import chromadb
from langchain.document_loaders import PyPDFLoader
from langchain.embeddings.openai import OpenAIEmbeddings
from langchain.vectorstores import Chroma
from langchain.indexes import VectorstoreIndexCreator
from langchain.chains import ConversationalRetrievalChain
from langchain.prompts import PromptTemplate
from langchain.chat_models import ChatOpenAI
from langchain.llms import OpenAI 

def load_document(Document):
    
    # loads a PDF document
    if not Document:
        return "Merci de fournir un document PDF"
    if not Document.name.endswith('.pdf'):
        return ("Merci de fournir un document PDF")

    loader = PyPDFLoader(Document.name)
    docs = loader.load()
    global k
    k = len(docs)
    
    # Create embeddings
    embeddings = OpenAIEmbeddings(openai_api_key = os.environ['OpenaiKey'])
    
    # Write in DB
    global docsearch
    docsearch = Chroma.from_documents(docs, embeddings, ids=["page" + str(d.metadata["page"]) for d in docs], k=1)
    global chat_history
    chat_history = []

    return "Endodage créé"

def get_chat_history(inputs) -> str:
    res = []
    for human, ai in inputs:
        res.append(f"Question : {human}\nRéponse : {ai}")
    return "\n".join(res)

def question_document(Question):
    
    if "docsearch" not in globals():
        return "Merci d'encoder un document PDF"

    # Define LLM
    turbo = ChatOpenAI(model_name="gpt-3.5-turbo", temperature=0, openai_api_key = os.environ['OpenaiKey'])
    davinci = OpenAI(model_name = "text-davinci-003", openai_api_key = os.environ['OpenaiKey'])

    # Customize map_reduce prompts
    #question_template = """{context}
    #Precise the number starting the above text in your answer. It corresponds to its page number in the document it is from. Label this number as "page".
    #Also make sure to answer in the same langage than the following question.
    #QUESTION : {question}
    #ANSWER :
    #"""
    
    #combine_template = """{summaries}
    #Note that the above text is based on transient extracts from one source document.
    #So make sure to not mention different documents or extracts or passages or portions or texts. There is only one, entire document. 
    #Also make sure to answer in the same langage than the following question. 
    #QUESTION : {question}.
    #ANSWER :
    #"""
    
    #question_prompt = PromptTemplate(template = question_template, input_variables=['context', 'question'])
    #combine_prompt = PromptTemplate(template = combine_template, input_variables=['summaries', 'question'])
    
    # Define chain  
    #chain_type_kwargs = { "combine_prompt" : combine_prompt, "question_prompt" : question_prompt} #, "return_intermediate_steps" : True}
    #qa = RetrievalQAWithSourcesChain.from_chain_type(llm = llm, chain_type = "map_reduce", chain_type_kwargs = chain_type_kwargs, retriever=docsearch.as_retriever(), return_source_documents = True)
   
    vectordbkwargs = {"search_distance": 10}
    search_kwargs={"k" : k}

    qa = ConversationalRetrievalChain.from_llm(llm = turbo, chain_type = "map_reduce",retriever=docsearch.as_retriever(search_kwargs = search_kwargs), get_chat_history = get_chat_history, return_source_documents = True)
    answer = qa({"question" : Question,"chat_history":chat_history, "vectordbkwargs": vectordbkwargs}, return_only_outputs = True)
    chat_history.append((Question, answer["answer"]))
    #answer = qa({"question" : Question}, )
    print(answer)
    return "".join(get_chat_history(chat_history))

with gr.Blocks() as demo:
    
    gr.Markdown(
    """
    # Interrogateur de PDF
    par Nicolas et Alex
    """)
    
    with gr.Row():
      
      with gr.Column():
        input_file = gr.inputs.File(label="Charger un document")
        greet_btnee = gr.Button("Encoder le document")
        output_words = gr.outputs.Textbox(label="Encodage")
        greet_btnee.click(fn=load_document, inputs=input_file, outputs = output_words)

      with gr.Column():
        text = gr.inputs.Textbox(label="Question")
        greet_btn = gr.Button("Poser une question")
        answer = gr.Textbox(label = "Réponse", lines = 8)
        greet_btn.click(fn = question_document, inputs = text, outputs = answer)
        
      
demo.launch()