Create app.py

app.py

@@ -0,0 +1,347 @@
+import langchain
+from langchain.embeddings import SentenceTransformerEmbeddings
+from langchain.document_loaders import UnstructuredPDFLoader,UnstructuredWordDocumentLoader
+from langchain.indexes import VectorstoreIndexCreator
+from langchain.vectorstores import FAISS
+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
+import time
+import requests
+import tempfile
+
+from groq import Groq
+
+
+
+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=800,
+    chunk_overlap=400,
+    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 < 0.1
+
+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 = 10
+  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, allow_dangerous_deserialization=True)
+            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):
+    if ui_session_id not in os.environ['users'].split(', '):
+        return "README.md", ""
+    print(files)
+    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, allow_dangerous_deserialization=True)
+    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):
+        print("ID : ", file_id, "FILE : ", file)
+        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':
+            db2 = merge_pdf_to_db(file.name,db,progress)
+        
+        if file_type == 'txt':
+            db2 = merge_txt_to_db(file.name,db,progress)
+        
+        if file_type == 'docx':
+            db2 = merge_docx_to_db(file.name,db,progress)
+
+        if file_type == 'zip':
+            db2 = unpack_zip_file(file.name,db,progress)
+
+        if db2 != None:
+            db = db2
+            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)
+            try:
+                shutil.move(file.name, directory_path)
+            except:
+                pass
+
+    ### load the updated db and zip it ###
+    progress(progress_step, desc = 'loading db')
+    db = FAISS.load_local(session_id,embeddings, allow_dangerous_deserialization=True)
+    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 add_to_db(references,ui_session_id):
+    files = store_files(references)
+    return embed_files(files,ui_session_id)
+
+def export_files(references):
+    files = store_files(references, ret_names=True)
+    #paths = [file.name for file in files]
+    return files
+    
+
+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_llm(system, user_input):
+    messages = [
+        {
+            "role": "system",
+            "content": system
+        },
+        {
+            "role": "user",
+            "content": user_input,
+        }
+    ]
+    client = Groq(api_key=os.environ["GROQ_KEY"])
+    chat_completion = client.chat.completions.create(
+        messages=messages,
+        model='mixtral-8x7b-32768',
+    )
+    return chat_completion.choices[0].message.content
+
+def ask_llm_stream(system, user_input):
+    llm_response = ""
+    client = Groq(api_key=os.environ["GROQ_KEY"])
+
+    messages = [
+        {
+            "role": "system",
+            "content": system
+        },
+        {
+            "role": "user",
+            "content": user_input,
+        }
+    ]
+
+    stream = client.chat.completions.create(
+        messages=messages,
+        model="mixtral-8x7b-32768",
+        temperature=0.5,
+        max_tokens=1024,
+        top_p=1,
+        stop=None,
+        stream=True,
+    )
+
+    for chunk in stream:
+        llm_response += str(chunk.choices[0].delta.content) if chunk.choices[0].delta.content is not None else ""
+        yield llm_response
+    
+
+def ask_gpt(query, ui_session_id, history):
+    if ui_session_id not in os.environ['users'].split(', '):
+        return "Please Login", "", ""
+    session_id = f"PDFAISS-{ui_session_id}"
+    try:
+      db = FAISS.load_local(session_id,embeddings, allow_dangerous_deserialization=True)
+      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)
+
+    documents = "\n\n*-*-*-*-*-*\n\n".join(f"Content: {doc.page_content}\n" for doc in docs)
+    system = f"# Instructions\nTake a deep breath and resonate step by step.\nYou are a helpful standard assistant. Your have only one mission and that consists in answering to the user input based on the **provided documents**. If the answer to the question that is asked by the user isn't contained in the **provided documents**, say so but **don't make up an answer**. I chose you because you can say 'I don't know' so please don't do like the other LLMs and don't define acronyms that aren\'t present in the following **PROVIDED DOCUMENTS** double check if it is present before answering. If some of the information can be useful for the user you can tell him.\nFinish your response by **ONE** follow up question that the provided documents could answer.\n\nThe documents are separated by the string \'*-*-*-*-*-*\'. Do not provide any explanations or details.\n\n# **Provided documents**: {documents}."    
+    gen = ask_llm_stream(system, query)
+    last_value=""
+    displayable_docs = display_docs(docs)
+    while True:
+        try:
+            last_value = next(gen)
+            yield last_value, displayable_docs, history + f"[query]\n{query}\n[answer]\n{last_value}\n[references]\n{displayable_docs}\n\n"
+        except StopIteration as e:
+            break
+    history += f"[query]\n{query}\n[answer]\n{last_value}\n[references]\n{displayable_docs}\n\n"
+    return last_value, displayable_docs, history
+
+
+def auth_user(ui_session_id):
+    if ui_session_id in os.environ['users'].split(', '):
+        return gr.Textbox(label='Username', visible=False), gr.File(file_count="multiple", file_types=[".txt", ".pdf",".zip",".docx"], visible=True), gr.Button("Reset AI Knowledge", visible=True), gr.Markdown(label='AI Answer', visible=True), gr.Textbox(placeholder="Type your question", label="Question ❔", scale=9, visible=True), gr.Button("▶", scale=1, visible=True), gr.Textbox(label='Sources', show_copy_button=True, visible=True), gr.File(label="Zipped database", visible=True), gr.Textbox(label='History', show_copy_button=True, visible=True)
+    else:
+        return gr.Textbox(label='Username', visible=True), gr.File(file_count="multiple", file_types=[".txt", ".pdf",".zip",".docx"], visible=False), gr.Button("Reset AI Knowledge", visible=False), gr.Markdown(label='AI Answer', visible=False), gr.Textbox(placeholder="Type your question", label="Question ❔", scale=9, visible=False), gr.Button("▶", scale=1, visible=False), gr.Textbox(label='Sources', show_copy_button=True, visible=False), gr.File(label="Zipped database", visible=False), gr.Textbox(label='History', show_copy_button=True, visible=False)
+
+with gr.Blocks() as demo:
+    gr.Markdown("# Enrich an LLM knowledge with your own documents 🧠🤖")
+        
+    with gr.Column():        
+        tb_session_id = gr.Textbox(label='Username')
+        docs_input = gr.File(file_count="multiple", file_types=[".txt", ".pdf",".zip",".docx"], visible=False)
+        btn_reset_db = gr.Button("Reset AI Knowledge", visible=False)
+
+
+    with gr.Column():
+        answer_output = gr.Markdown(label='AI Answer', visible=False)
+        with gr.Row():
+            query_input = gr.Textbox(placeholder="Type your question", label="Question ❔", scale=9, visible=False)
+            btn_askGPT = gr.Button("▶", scale=1, visible=False)
+        
+        tb_sources = gr.Textbox(label='Sources', show_copy_button=True, visible=False)
+        
+
+    with gr.Accordion("Download your knowledge base and see your conversation history", open=False):
+        db_output = gr.File(label="Zipped database", visible=False)
+        tb_history = gr.Textbox(label='History', show_copy_button=True, visible=False, interactive=False)
+        
+
+    tb_session_id.submit(auth_user, inputs=tb_session_id, outputs=[tb_session_id, docs_input, btn_reset_db, answer_output, query_input, btn_askGPT, tb_sources, db_output, tb_history])
+
+    docs_input.upload(embed_files, inputs=[docs_input,tb_session_id], outputs=[db_output,tb_session_id, query_input])
+    btn_reset_db.click(reset_database,inputs=[tb_session_id],outputs=[db_output])
+    btn_askGPT.click(ask_gpt, inputs=[query_input, tb_session_id, tb_history], outputs=[answer_output, tb_sources, tb_history])
+    query_input.submit(ask_gpt, inputs=[query_input, tb_session_id, tb_history], outputs=[answer_output, tb_sources, tb_history])
+
+
+demo.launch(debug=False,share=False)