Update app.py

app.py

@@ -5,20 +5,35 @@ from langchain.text_splitter import RecursiveCharacterTextSplitter
 from langchain.vectorstores import Chroma
 from langchain.chains import ConversationalRetrievalChain
 from langchain.embeddings import HuggingFaceEmbeddings 
+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_names = ["mistralai/Mixtral-8x7B-Instruct-v0.1"]
-llm_names_simple = [os.path.basename(llm) for llm in llm_names]
+llm_name0 = "mistralai/Mixtral-8x7B-Instruct-v0.1"
+list_llm = [llm_name0]
+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):
     loaders = [PyPDFLoader(x) for x in list_file_path]
-    pages = [loader.load() for loader in loaders]
-    text_splitter = RecursiveCharacterTextSplitter(chunk_size=chunk_size, chunk_overlap=chunk_overlap)
+    pages = []
+    for loader in loaders:
+        pages.extend(loader.load())
+    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()
     new_client = chromadb.EphemeralClient()
@@ -30,26 +45,35 @@ def create_db(splits, collection_name):
     )
     return vectordb
 
+# Load vector database
 def load_db():
     embedding = HuggingFaceEmbeddings()
-    vectordb = Chroma(embedding_function=embedding)
+    vectordb = Chroma(
+        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...")
-    model_kwargs = {"temperature": temperature, "max_new_tokens": max_tokens, "top_k": top_k}
     if llm_model == "mistralai/Mixtral-8x7B-Instruct-v0.1":
-        model_kwargs["load_in_8bit"] = True
-    llm = HuggingFaceHub(repo_id=llm_model, model_kwargs=model_kwargs)
+        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()
+    memory = ConversationBufferMemory(
+        memory_key="chat_history",
+        output_key='answer',
+        return_messages=True
+    )
+
+    retriever=vector_db.as_retriever()
     progress(0.8, desc="Defining retrieval chain...")
     qa_chain = ConversationalRetrievalChain.from_llm(
         llm,
         retriever=retriever,
-        chain_type="stuff",
+        chain_type="stuff", 
         memory=memory,
         return_source_documents=True,
     )
@@ -67,14 +91,18 @@ def initialize_database(list_file_obj, chunk_size, chunk_overlap, progress=gr.Pr
     return vector_db, collection_name, "Complete!"
 
 def initialize_LLM(llm_option, llm_temperature, max_tokens, top_k, vector_db, progress=gr.Progress()):
-    llm_name = llm_names[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 = [f"User: {user_message}\nAssistant: {bot_message}" for user_message, bot_message in chat_history]
+    formatted_chat_history = []
+    for user_message, bot_message in chat_history:
+        formatted_chat_history.append(f"User: {user_message}")
+        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)
     response = qa_chain({"question": message, "chat_history": formatted_chat_history})
@@ -86,9 +114,12 @@ def conversation(qa_chain, message, history):
     response_source2_page = response_sources[1].metadata["page"] + 1
     new_history = history + [(message, response_answer)]
     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 = [file_obj.name for _ in file_obj]
+    list_file_path = []
+    for idx, file in enumerate(file_obj):
+        file_path = file_obj.name
+        list_file_path.append(file_path)
     return list_file_path
 
 def demo():
@@ -96,43 +127,79 @@ def demo():
         vector_db = gr.State()
         qa_chain = gr.State()
         collection_name = gr.State()
-
-        gr.Markdown("""<center><h2>ChatPDF</center></h2>""")
-
-        with gr.Tab("Step 1 - Selezione PDF"):
-            document = gr.Files(height=100, file_count="multiple", file_types=["pdf"], interactive=True, label="Upload your PDF documents (single or multiple)")
-            db_btn = gr.Radio(["ChromaDB"], label="Vector database type", value="ChromaDB", type="index", info="Choose your vector database")
+        
+        gr.Markdown(
+        """<center><h2>PDF-based chatbot (powered by LangChain and open-source LLMs)</center></h2>""")
+        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)")
+            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):
-                slider_chunk_size = gr.Slider(minimum=100, maximum=1000, value=600, step=20, label="Chunk size", info="Chunk size", interactive=True)
-                slider_chunk_overlap = gr.Slider(minimum=10, maximum=200, value=40, step=10, label="Chunk overlap", info="Chunk overlap", interactive=True)
-            db_progress = gr.Textbox(label="Vector database initialization", value="None")
-            db_btn.click(initialize_database, inputs=[document, slider_chunk_size, slider_chunk_overlap], outputs=[vector_db, collection_name, db_progress])
-
-        with gr.Tab("Step 2 - Inizializzazione QA"):
-            llm_btn = gr.Radio(llm_names_simple, label="LLM models", value=llm_names_simple[0], type="index", info="Choose your LLM model")
+                with gr.Row():
+                    slider_chunk_size = gr.Slider(minimum = 100, maximum = 1000, value=600, step=20, label="Chunk size", info="Chunk size", interactive=True)
+                with gr.Row():
+                    slider_chunk_overlap = gr.Slider(minimum = 10, maximum = 200, value=40, step=10, label="Chunk overlap", info="Chunk overlap", interactive=True)
+            with gr.Row():
+                db_progress = gr.Textbox(label="Vector database initialization", value="None")
+            with gr.Row():
+                db_btn = gr.Button("Generate vector database...")
+            
+        with gr.Tab("Step 2 - QA chain initialization"):
+            with gr.Row():
+                llm_btn = gr.Radio(list_llm_simple, \
+                    label="LLM models", value = list_llm_simple[0], type="index", info="Choose your LLM model")
             with gr.Accordion("Advanced options - LLM model", open=False):
-                slider_temperature = gr.Slider(minimum=0.0, maximum=1.0, value=0.7, step=0.1, label="Temperature", info="Model temperature", interactive=True)
-                slider_maxtokens = gr.Slider(minimum=224, maximum=4096, value=1024, step=32, label="Max Tokens", info="Model max tokens", interactive=True)
-                slider_topk = gr.Slider(minimum=1, maximum=10, value=3, step=1, label="top-k samples", info="Model top-k samples", interactive=True)
-            llm_progress = gr.Textbox(value="None", label="QA chain initialization")
-            qachain_btn = gr.Button("Initialize question-answering chain...")
-            qachain_btn.click(initialize_LLM, inputs=[llm_btn, slider_temperature, slider_maxtokens, slider_topk, vector_db], outputs=[qa_chain, llm_progress]).then(lambda: [None, "", 0, "", 0], inputs=None, outputs=[chatbot, doc_source1, source1_page, doc_source2, source2_page], queue=False)
-
-        with gr.Tab("Step 3 - Conversazione con Chatbot"):
+                with gr.Row():
+                    slider_temperature = gr.Slider(minimum = 0.0, maximum = 1.0, value=0.7, step=0.1, label="Temperature", info="Model temperature", interactive=True)
+                with gr.Row():
+                    slider_maxtokens = gr.Slider(minimum = 224, maximum = 4096, value=1024, step=32, label="Max Tokens", info="Model max tokens", interactive=True)
+                with gr.Row():
+                    slider_topk = gr.Slider(minimum = 1, maximum = 10, value=3, step=1, label="top-k samples", info="Model top-k samples", interactive=True)
+            with gr.Row():
+                llm_progress = gr.Textbox(value="None",label="QA chain initialization")
+            with gr.Row():
+                qachain_btn = gr.Button("Initialize question-answering chain...")
+
+        with gr.Tab("Step 3 - Conversation with chatbot"):
             chatbot = gr.Chatbot(height=300)
-            with gr.Accordion("Advanced - Document references", open=True):
-                doc_source1 = gr.Textbox(label="Reference 1", lines=2, container=True, scale=20)
-                source1_page = gr.Number(label="Page", scale=1)
-                doc_source2 = gr.Textbox(label="Reference 2", lines=2, container=True, scale=20)
-                source2_page = gr.Number(label="Page", scale=1)
-            msg = gr.Textbox(placeholder="Type message", container=True)
-            submit_btn = gr.Button("Submit")
-            clear_btn = gr.ClearButton([msg, chatbot])
-
-        msg.submit(conversation, inputs=[qa_chain, msg, chatbot], outputs=[qa_chain, msg, chatbot, doc_source1, source1_page, doc_source2, source2_page], queue=False)
-        submit_btn.click(conversation, inputs=[qa_chain, msg, chatbot], outputs=[qa_chain, msg, chatbot, doc_source1, source1_page, doc_source2, source2_page], queue=False)
-        clear_btn.click(lambda: [None, "", 0, "", 0], inputs=None, outputs=[chatbot, doc_source1, source1_page, doc_source2, source2_page], queue=False)
-
+            with gr.Accordion("Advanced - Document references", open=False):
+                with gr.Row():
+                    doc_source1 = gr.Textbox(label="Reference 1", lines=2, container=True, scale=20)
+                    source1_page = gr.Number(label="Page", scale=1)
+                with gr.Row():
+                    doc_source2 = gr.Textbox(label="Reference 2", lines=2, container=True, scale=20)
+                    source2_page = gr.Number(label="Page", scale=1)
+            with gr.Row():
+                msg = gr.Textbox(placeholder="Type message", container=True)
+            with gr.Row():
+                submit_btn = gr.Button("Submit")
+                clear_btn = gr.ClearButton([msg, chatbot])
+            
+        # Preprocessing events
+        db_btn.click(initialize_database, \
+            inputs=[document, slider_chunk_size, slider_chunk_overlap], \
+            outputs=[vector_db, collection_name, db_progress])
+        qachain_btn.click(initialize_LLM, \
+            inputs=[llm_btn, slider_temperature, slider_maxtokens, slider_topk, vector_db], \
+            outputs=[qa_chain, llm_progress]).then(lambda:[None,"",0,"",0], \
+            inputs=None, \
+            outputs=[chatbot, doc_source1, source1_page, doc_source2, source2_page], \
+            queue=False)
+
+        # Chatbot events
+        msg.submit(conversation, \
+            inputs=[qa_chain, msg, chatbot], \
+            outputs=[qa_chain, msg, chatbot, doc_source1, source1_page, doc_source2, source2_page], \
+            queue=False)
+        submit_btn.click(conversation, \
+            inputs=[qa_chain, msg, chatbot], \
+            outputs=[qa_chain, msg, chatbot, doc_source1, source1_page, doc_source2, source2_page], \
+            queue=False)
+        clear_btn.click(lambda:[None,"",0,"",0], \
+            inputs=None, \
+            outputs=[chatbot, doc_source1, source1_page, doc_source2, source2_page], \
+            queue=False)
     demo.queue().launch(debug=True)
 
 if __name__ == "__main__":