app.py

import langchain
from langchain.embeddings import SentenceTransformerEmbeddings
from langchain.chains.question_answering import load_qa_chain
from langchain.document_loaders import UnstructuredPDFLoader,UnstructuredWordDocumentLoader
from langchain.indexes import VectorstoreIndexCreator
from langchain.vectorstores import FAISS
from langchain import HuggingFaceHub
from langchain import PromptTemplate
from langchain.chat_models import ChatOpenAI
from zipfile import ZipFile
import gradio as gr
import openpyxl
import os
import shutil
from langchain.schema import Document
from langchain.text_splitter import RecursiveCharacterTextSplitter
import tiktoken
import secrets

tokenizer = tiktoken.encoding_for_model("gpt-3.5-turbo")

# create the length function
def tiktoken_len(text):
    tokens = tokenizer.encode(
        text,
        disallowed_special=()
    )
    return len(tokens)

text_splitter = RecursiveCharacterTextSplitter(
    chunk_size=400,
    chunk_overlap=40,
    length_function=tiktoken_len,
    separators=["\n\n", "\n", " ", ""]
)

embeddings = SentenceTransformerEmbeddings(model_name="all-MiniLM-L6-v2")
foo = Document(page_content='foo is fou!',metadata={"source":'foo source'})

def reset_database(ui_session_id):
  session_id = f"PDFAISS-{ui_session_id}"
  if 'drive' in session_id:
    print("RESET DATABASE: session_id contains 'drive' !!")
    return None

  try:
    shutil.rmtree(session_id)
  except:
    print(f'no {session_id} directory present')
  
  try:
    os.remove(f"{session_id}.zip")
  except:
    print("no {session_id}.zip present")

  return None

def is_duplicate(split_docs,db):
  epsilon=0.0
  print(f"DUPLICATE: Treating: {split_docs[0].metadata['source'].split('/')[-1]}")
  for i in range(min(3,len(split_docs))):
    query = split_docs[i].page_content
    docs = db.similarity_search_with_score(query,k=1)
    _ , score = docs[0]
    epsilon += score
  print(f"DUPLICATE: epsilon: {epsilon}")
  return epsilon < 0.05

def merge_split_docs_to_db(split_docs,db,progress,progress_step=0.1):
  progress(progress_step,desc="merging docs")
  if len(split_docs)==0:
    print("MERGE to db: NO docs!!")
    return

  filename = split_docs[0].metadata['source']
  if is_duplicate(split_docs,db):
    print(f"MERGE: Document is duplicated: {filename}")
    return
  print(f"MERGE: number of split docs: {len(split_docs)}")
  batch = 20
  for i in range(0, len(split_docs), batch):
    progress(i/len(split_docs),desc=f"added {i} chunks of {len(split_docs)} chunks")
    db1 = FAISS.from_documents(split_docs[i:i+batch], embeddings)
    db.merge_from(db1)
  return db

def merge_pdf_to_db(filename,db,progress,progress_step=0.1):
  progress_step+=0.05
  progress(progress_step,'unpacking pdf')
  doc = UnstructuredPDFLoader(filename).load()
  doc[0].metadata['source'] = filename.split('/')[-1]
  split_docs = text_splitter.split_documents(doc)
  progress_step+=0.3
  progress(progress_step,'docx unpacked')
  return merge_split_docs_to_db(split_docs,db,progress,progress_step)

def merge_docx_to_db(filename,db,progress,progress_step=0.1):
  progress_step+=0.05
  progress(progress_step,'unpacking docx')
  doc = UnstructuredWordDocumentLoader(filename).load()
  doc[0].metadata['source'] = filename.split('/')[-1]
  split_docs = text_splitter.split_documents(doc)
  progress_step+=0.3
  progress(progress_step,'docx unpacked')
  return merge_split_docs_to_db(split_docs,db,progress,progress_step)

def merge_txt_to_db(filename,db,progress,progress_step=0.1):
  progress_step+=0.05
  progress(progress_step,'unpacking txt')
  with open(filename) as f:
      docs = text_splitter.split_text(f.read())
      split_docs = [Document(page_content=doc,metadata={'source':filename.split('/')[-1]}) for doc in docs]
  progress_step+=0.3
  progress(progress_step,'txt unpacked')
  return merge_split_docs_to_db(split_docs,db,progress,progress_step)

def unpack_zip_file(filename,db,progress):
    with ZipFile(filename, 'r') as zipObj:
        contents = zipObj.namelist()
        print(f"unpack zip: contents: {contents}")
        tmp_directory = filename.split('/')[-1].split('.')[-2]
        shutil.unpack_archive(filename, tmp_directory)

        if 'index.faiss' in [item.lower() for item in contents]:
            db2 = FAISS.load_local(tmp_directory, embeddings)
            db.merge_from(db2)
            return db
        
        for file in contents:
            if file.lower().endswith('.docx'):
              db = merge_docx_to_db(f"{tmp_directory}/{file}",db,progress)
            if file.lower().endswith('.pdf'):
              db = merge_pdf_to_db(f"{tmp_directory}/{file}",db,progress)
            if file.lower().endswith('.txt'):
              db = merge_txt_to_db(f"{tmp_directory}/{file}",db,progress)
        return db

def add_files_to_zip(session_id):
    zip_file_name = f"{session_id}.zip"
    with ZipFile(zip_file_name, "w") as zipObj:
        for root, dirs, files in os.walk(session_id):
            for file_name in files:
                file_path = os.path.join(root, file_name)
                arcname = os.path.relpath(file_path, session_id)
                zipObj.write(file_path, arcname)

#### UI Functions ####

def embed_files(files,ui_session_id,progress=gr.Progress(),progress_step=0.05):
    progress(progress_step,desc="Starting...")
    split_docs=[]
    if len(ui_session_id)==0:
      ui_session_id = secrets.token_urlsafe(16)
    session_id = f"PDFAISS-{ui_session_id}"

    try:
      db = FAISS.load_local(session_id,embeddings)
    except:
      print(f"SESSION: {session_id} database does not exist, create a FAISS db")
      db =  FAISS.from_documents([foo], embeddings)
      db.save_local(session_id)
      print(f"SESSION: {session_id} database created")
    
    print("EMBEDDED, before embeddeding: ",session_id,len(db.index_to_docstore_id))
    for file_id,file in enumerate(files):
        file_type = file.name.split('.')[-1].lower()
        source = file.name.split('/')[-1]
        print(f"current file: {source}")
        progress(file_id/len(files),desc=f"Treating {source}")

        if file_type == 'pdf':
            db = merge_pdf_to_db(file.name,db,progress)
            db.save_local(session_id)
        
        if file_type == 'txt':
            db = merge_txt_to_db(file.name,db,progress)
            db.save_local(session_id)
        
        if file_type == 'docx':
            db = merge_docx_to_db(file.name,db,progress)
            db.save_local(session_id)

        if file_type == 'zip':
            db = unpack_zip_file(file.name,db,progress)
            db.save_local(session_id)
    
        ### move file to store ###
        progress(progress_step, desc = 'moving file to store')
        directory_path = f"{session_id}/store/"
        if not os.path.exists(directory_path):
            os.makedirs(directory_path)
        shutil.move(file.name, directory_path)

    ### load the updated db and zip it ###
    progress(progress_step, desc = 'loading db')
    db = FAISS.load_local(session_id,embeddings)
    print("EMBEDDED, after embeddeding: ",session_id,len(db.index_to_docstore_id))
    progress(progress_step, desc = 'zipping db for download')
    add_files_to_zip(session_id)
    print(f"EMBEDDED: db zipped")
    progress(progress_step, desc = 'db zipped')
    return f"{session_id}.zip",ui_session_id

def display_docs(docs):
  output_str = ''
  for i, doc in enumerate(docs):
      source = doc.metadata['source'].split('/')[-1]
      output_str += f"Ref: {i+1}\n{repr(doc.page_content)}\nSource: {source}\n\n"
  return output_str

def ask_gpt(query, apikey,history,ui_session_id):
    session_id = f"PDFAISS-{ui_session_id}"
    try:
      db = FAISS.load_local(session_id,embeddings)
      print("ASKGPT after loading",session_id,len(db.index_to_docstore_id))
    except:
      print(f"SESSION: {session_id} database does not exist")
      return f"SESSION: {session_id} database does not exist","",""

    docs = db.similarity_search(query)
    history += f"[query]\n{query}\n[answer]\n"
    if(apikey==""):
        history += f"None\n[references]\n{display_docs(docs)}\n\n"
        return "No answer from GPT", display_docs(docs),history
    else:
        llm = ChatOpenAI(temperature=0, model_name = 'gpt-3.5-turbo', openai_api_key=apikey)
        chain = load_qa_chain(llm, chain_type="stuff")
        answer = chain.run(input_documents=docs, question=query, verbose=True)
        history += f"{answer}\n[references]\n{display_docs(docs)}\n\n"
        return answer,display_docs(docs),history

with gr.Blocks() as demo:
    gr.Markdown("Upload your documents and question them.")
    with gr.Tab("Upload PDF & TXT"): 
        tb_session_id = gr.Textbox(label='session id')
        docs_input = gr.File(file_count="multiple", file_types=[".txt", ".pdf",".zip",".docx"])
        db_output = gr.outputs.File(label="Download zipped database")
        btn_generate_db = gr.Button("Generate database")
        btn_reset_db = gr.Button("Reset database")

    with gr.Tab("Ask PDF"):
        with gr.Column():
            api_key = gr.Textbox(placeholder="Leave blank if you don't have any", label="OpenAI API Key",type='password')
            query_input = gr.Textbox(placeholder="Type your question", label="Question")
            btn_askGPT = gr.Button("Answer")
            answer_output = gr.Textbox(label='GPT 3.5 answer')
            answer_output.style(show_copy_button=True)
            sources = gr.Textbox(label='Sources')
            sources.style(show_copy_button=True)
            history = gr.Textbox(label='History')
            history.style(show_copy_button=True)

    btn_generate_db.click(embed_files, inputs=[docs_input,tb_session_id], outputs=[db_output,tb_session_id])
    btn_reset_db.click(reset_database,inputs=[tb_session_id],outputs=[db_output])
    btn_askGPT.click(ask_gpt, inputs=[query_input,api_key,history,tb_session_id], outputs=[answer_output,sources,history])

demo.queue(concurrency_count=10)
demo.launch(debug=False,share=False)