app.py

import os
import time
from langchain.embeddings import HuggingFaceEmbeddings
from huggingface_hub import InferenceClient
import langchain
from langchain import HuggingFaceHub
from langchain.cache import InMemoryCache
from langchain.document_loaders import PyPDFLoader, OnlinePDFLoader, Docx2txtLoader, UnstructuredWordDocumentLoader, UnstructuredPowerPointLoader
from langchain.text_splitter import CharacterTextSplitter,RecursiveCharacterTextSplitter
from langchain.vectorstores import Chroma
from langchain.chains import RetrievalQA
import gradio as gr

def define_embeddings_llm(openai_key):
  if openai_key != "":
    embeddings = OpenAIEmbeddings(openai_api_key=openai_key)
    llm = OpenAI(
        temperature=0, model_name="gpt-3.5-turbo-16k", openai_api_key=openai_key, verbose=False
    )
  else:
    HUGGINGFACEHUB_API_TOKEN = os.environ["HUGGINGFACEHUB_API_TOKEN"]
    embeddings = HuggingFaceEmbeddings(
                model_name="sentence-transformers/all-MiniLM-L6-v2"
            )
    llm = HuggingFaceHub(repo_id="MBZUAI/LaMini-Flan-T5-248M",
                         model_kwargs={"max_length":2048,
                                       "temperature":0.2}
                        )
  langchain.llm_cache = InMemoryCache()
  return embeddings,llm

def build_context(openai_key,files,urls):
  embeddings, llm = define_embeddings_llm(openai_key)
  documents = []
  if files is not None:
    for idx, file in enumerate(files):
      if file.name.endswith('.pdf'):
        loader = PyPDFLoader(file.name)
        documents.extend(loader.load())
      elif file.name.endswith('.docx'):
        loader = Docx2txtLoader(file.name)
        documents.extend(loader.load())
      elif file.name.endswith('.ppt') or file.name.endswith('.pptx'):
        loader = UnstructuredPowerPointLoader(file.name)
        documents.extend(loader.load())
  if urls != "":
    list_urls = urls.split(sep=",")
    for url in list_urls: 
      loader = OnlinePDFLoader(url)
      documents.extend(loader.load())
  #text_splitter = RecursiveCharacterTextSplitter(chunk_size=400,chunk_overlap=20,length_function=len,separators=["\n\n", "\n", " ", ""])
  text_splitter = CharacterTextSplitter(chunk_size=1000, chunk_overlap=200)
  chunked_documents = text_splitter.split_documents(documents)
  global vectordb
  vectordb = Chroma.from_documents(
    documents=chunked_documents,
    embedding=embeddings
  )
  global qa_chain
  qa_chain = RetrievalQA.from_chain_type(
    llm=llm,
    retriever=vectordb.as_retriever(search_kwargs={'k': 7}),
    chain_type="stuff",
    #return_source_documents=True
  )
  return "ready"

def loading():
    return "Loading..."

def respond(message, chat_history):
  result = qa_chain({"query": message})["result"]
  chat_history.append((message, result))
  time.sleep(2)
  return "", chat_history

def clear_chromadb():
  ids = vectordb.get()["ids"]
  for id in ids:
    vectordb._collection.delete(ids=id)

with gr.Blocks() as demo:
  with gr.Row():
    openai_key = gr.Textbox(label="Enter your OpenAI API Key if you want to use the gpt-3.5-turbo-16k model. If not, the open source LaMini-Flan-T5-248M is used")
  with gr.Row():
    with gr.Column():
      pdf_docs = gr.Files(label="Load pdf/docx/ppt/pptx files", file_types=['.pdf','.docx','.ppt','.pptx'], type="file")
    with gr.Column():
      urls = gr.Textbox(label="Enter one of multiple online pdf urls (comma separated if multiple)", value=None)
  with gr.Row():
    load_context = gr.Button("Load documents and urls")
  with gr.Row():
    loading_status = gr.Textbox(label="Status", placeholder="", interactive=False)
  with gr.Row():
    with gr.Column():
      hg_chatbot = gr.Chatbot()
      msg = gr.Textbox(label="User message")
      clear = gr.ClearButton([msg, hg_chatbot])
      cleardb = gr.Button(value="Réinitialiser le contexte")
  load_context.click(loading, None, loading_status, queue=False)
  load_context.click(build_context, inputs=[openai_key,pdf_docs, urls], outputs=[loading_status], queue=False)
  msg.submit(respond, [msg, hg_chatbot], [msg, hg_chatbot])
  cleardb.click(clear_chromadb)

demo.queue(concurrency_count=3)
demo.launch()