Update app.py

app.py

@@ -1,5 +1,6 @@
 import gradio as gr
 import os
+
 from langchain.document_loaders import PyPDFLoader
 from langchain.text_splitter import RecursiveCharacterTextSplitter
 from langchain.vectorstores import Chroma
@@ -9,30 +10,39 @@ from langchain.llms import HuggingFacePipeline
 from langchain.chains import ConversationChain
 from langchain.memory import ConversationBufferMemory
 from langchain.llms import HuggingFaceHub
+
 from pathlib import Path
 import chromadb
+
 from transformers import AutoTokenizer
 import transformers
 import torch
 import tqdm 
 import accelerate
 
+
+
 llm_name0 = "mistralai/Mixtral-8x7B-Instruct-v0.1"
-list_llm = [llm_name0]
+list_llm = [llm_name0, llm_name1, llm_name2, llm_name3, llm_name4, llm_name5, llm_name6, llm_name7, llm_name8]
 list_llm_simple = [os.path.basename(llm) for llm in list_llm]
 
 # Load PDF document and create doc splits
 def load_doc(list_file_path, chunk_size, chunk_overlap):
+    # Processing for one document only
+    # loader = PyPDFLoader(file_path)
+    # pages = loader.load()
     loaders = [PyPDFLoader(x) for x in list_file_path]
     pages = []
     for loader in loaders:
         pages.extend(loader.load())
+    # text_splitter = RecursiveCharacterTextSplitter(chunk_size = 600, chunk_overlap = 50)
     text_splitter = RecursiveCharacterTextSplitter(
         chunk_size = chunk_size, 
         chunk_overlap = chunk_overlap)
     doc_splits = text_splitter.split_documents(pages)
     return doc_splits
 
+
 # Create vector database
 def create_db(splits, collection_name):
     embedding = HuggingFaceEmbeddings()
@@ -42,32 +52,38 @@ def create_db(splits, collection_name):
         embedding=embedding,
         client=new_client,
         collection_name=collection_name,
+        # persist_directory=default_persist_directory
     )
     return vectordb
 
+
 # Load vector database
 def load_db():
     embedding = HuggingFaceEmbeddings()
     vectordb = Chroma(
+        # persist_directory=default_persist_directory, 
         embedding_function=embedding)
     return vectordb
 
+
 # Initialize langchain LLM chain
 def initialize_llmchain(llm_model, temperature, max_tokens, top_k, vector_db, progress=gr.Progress()):
     progress(0.1, desc="Initializing HF tokenizer...")
     progress(0.5, desc="Initializing HF Hub...")
+    # URL: https://github.com/langchain-ai/langchain/issues/6080
     if llm_model == "mistralai/Mixtral-8x7B-Instruct-v0.1":
         llm = HuggingFaceHub(
             repo_id=llm_model, 
             model_kwargs={"temperature": temperature, "max_new_tokens": max_tokens, "top_k": top_k, "load_in_8bit": True}
         )
+    
     progress(0.75, desc="Defining buffer memory...")
     memory = ConversationBufferMemory(
         memory_key="chat_history",
         output_key='answer',
         return_messages=True
     )
-
+    # retriever=vector_db.as_retriever(search_type="similarity", search_kwargs={'k': 3})
     retriever=vector_db.as_retriever()
     progress(0.8, desc="Defining retrieval chain...")
     qa_chain = ConversationalRetrievalChain.from_llm(
@@ -75,27 +91,42 @@ def initialize_llmchain(llm_model, temperature, max_tokens, top_k, vector_db, pr
         retriever=retriever,
         chain_type="stuff", 
         memory=memory,
+        # combine_docs_chain_kwargs={"prompt": your_prompt})
         return_source_documents=True,
+        # return_generated_question=True,
+        # verbose=True,
     )
     progress(0.9, desc="Done!")
     return qa_chain
 
+
+# Initialize database
 def initialize_database(list_file_obj, chunk_size, chunk_overlap, progress=gr.Progress()):
+    # Create list of documents (when valid)
+    #file_path = file_obj.name
     list_file_path = [x.name for x in list_file_obj if x is not None]
     collection_name = Path(list_file_path[0]).stem
+    # print('list_file_path: ', list_file_path)
+    # print('Collection name: ', collection_name)
     progress(0.25, desc="Loading document...")
+    # Load document and create splits
     doc_splits = load_doc(list_file_path, chunk_size, chunk_overlap)
+    # Create or load Vector database
     progress(0.5, desc="Generating vector database...")
+    # global vector_db
     vector_db = create_db(doc_splits, collection_name)
     progress(0.9, desc="Done!")
     return vector_db, collection_name, "Complete!"
 
+
 def initialize_LLM(llm_option, llm_temperature, max_tokens, top_k, vector_db, progress=gr.Progress()):
+    # print("llm_option",llm_option)
     llm_name = list_llm[llm_option]
     print("llm_name: ",llm_name)
     qa_chain = initialize_llmchain(llm_name, llm_temperature, max_tokens, top_k, vector_db, progress)
     return qa_chain, "Complete!"
 
+
 def format_chat_history(message, chat_history):
     formatted_chat_history = []
     for user_message, bot_message in chat_history:
@@ -103,25 +134,39 @@ def format_chat_history(message, chat_history):
         formatted_chat_history.append(f"Assistant: {bot_message}")
     return formatted_chat_history
     
+
 def conversation(qa_chain, message, history):
     formatted_chat_history = format_chat_history(message, history)
+    #print("formatted_chat_history",formatted_chat_history)
+   
+    # Generate response using QA chain
     response = qa_chain({"question": message, "chat_history": formatted_chat_history})
     response_answer = response["answer"]
     response_sources = response["source_documents"]
     response_source1 = response_sources[0].page_content.strip()
     response_source2 = response_sources[1].page_content.strip()
+    # Langchain sources are zero-based
     response_source1_page = response_sources[0].metadata["page"] + 1
     response_source2_page = response_sources[1].metadata["page"] + 1
+    # print ('chat response: ', response_answer)
+    # print('DB source', response_sources)
+    
+    # Append user message and response to chat history
     new_history = history + [(message, response_answer)]
+    # return gr.update(value=""), new_history, response_sources[0], response_sources[1] 
     return qa_chain, gr.update(value=""), new_history, response_source1, response_source1_page, response_source2, response_source2_page
     
+
 def upload_file(file_obj):
     list_file_path = []
     for idx, file in enumerate(file_obj):
         file_path = file_obj.name
         list_file_path.append(file_path)
+    # print(file_path)
+    # initialize_database(file_path, progress)
     return list_file_path
 
+
 def demo():
     with gr.Blocks(theme="base") as demo:
         vector_db = gr.State()
@@ -129,10 +174,16 @@ def demo():
         collection_name = gr.State()
         
         gr.Markdown(
-        """<center><h2>PDF-based chatbot (powered by LangChain and open-source LLMs)</center></h2>""")
+        """<center><h2>PDF-based chatbot (powered by LangChain and open-source LLMs)</center></h2>
+        <h3>Ask any questions about your PDF documents, along with follow-ups</h3>
+        <b>Note:</b> This AI assistant performs retrieval-augmented generation from your PDF documents. \
+        When generating answers, it takes past questions into account (via conversational memory), and includes document references for clarity purposes.</i>
+        <br><b>Warning:</b> This space uses the free CPU Basic hardware from Hugging Face. Some steps and LLM models used below (free inference endpoints) can take some time to generate an output.<br>
+        """)
         with gr.Tab("Step 1 - Document pre-processing"):
             with gr.Row():
                 document = gr.Files(height=100, file_count="multiple", file_types=["pdf"], interactive=True, label="Upload your PDF documents (single or multiple)")
+                # upload_btn = gr.UploadButton("Loading document...", height=100, file_count="multiple", file_types=["pdf"], scale=1)
             with gr.Row():
                 db_btn = gr.Radio(["ChromaDB"], label="Vector database type", value = "ChromaDB", type="index", info="Choose your vector database")
             with gr.Accordion("Advanced options - Document text splitter", open=False):
@@ -177,6 +228,7 @@ def demo():
                 clear_btn = gr.ClearButton([msg, chatbot])
             
         # Preprocessing events
+        #upload_btn.upload(upload_file, inputs=[upload_btn], outputs=[document])
         db_btn.click(initialize_database, \
             inputs=[document, slider_chunk_size, slider_chunk_overlap], \
             outputs=[vector_db, collection_name, db_progress])
@@ -202,5 +254,6 @@ def demo():
             queue=False)
     demo.queue().launch(debug=True)
 
+
 if __name__ == "__main__":
     demo()