Update app.py

app.py

@@ -1,179 +1,304 @@
-from langchain_huggingface import HuggingFaceEmbeddings
 import gradio as gr
 import os
-from googletrans import Translator
+
+from langchain_community.document_loaders import PyPDFLoader
+from langchain.text_splitter import RecursiveCharacterTextSplitter
 from langchain_community.vectorstores import Chroma
-from langchain_community.document_loaders import UnstructuredPDFLoader, PyPDFLoader
-from langchain.text_splitter import CharacterTextSplitter
 from langchain.chains import ConversationalRetrievalChain
-from langchain.schema import Document
+from langchain_community.embeddings import HuggingFaceEmbeddings 
+from langchain_community.llms import HuggingFacePipeline
+from langchain.chains import ConversationChain
 from langchain.memory import ConversationBufferMemory
-from langchain.callbacks.manager import CallbackManager
-from langchain.callbacks.streaming_stdout import StreamingStdOutCallbackHandler
-from langchain.llms.base import LLM
-from typing import List, Dict, Any, Optional
-from pydantic import BaseModel
-from langchain.llms.base import LLM
-from transformers import AutoTokenizer, AutoModelForCausalLM
+from langchain_community.llms import HuggingFaceEndpoint
+
+from pathlib import Path
+import chromadb
+from unidecode import unidecode
+
+from transformers import AutoTokenizer
+import transformers
 import torch
-import logging
+import tqdm 
+import accelerate
+import re
 
-# Configurazione del logging
-logging.basicConfig(level=logging.INFO)
-logger = logging.getLogger(__name__)
 
-# Aggiornamento dell'inizializzazione di HuggingFaceEmbeddings
-embedding_function = HuggingFaceEmbeddings(model_name="sentence-transformers/all-MiniLM-L6-v2")
 
-# Definizione della lista di modelli LLM
-list_llm_simple = ["meta/llama-7b-hf", "meta/llama-7b-hf"]
-list_llm = ["meta/llama-7b-hf", "meta/llama-7b-hf"]
+# default_persist_directory = './chroma_HF/'
+list_llm = ["mistralai/Mistral-7B-Instruct-v0.2", "mistralai/Mixtral-8x7B-Instruct-v0.1", "mistralai/Mistral-7B-Instruct-v0.1", \
+    "google/gemma-7b-it","google/gemma-2b-it", \
+    "HuggingFaceH4/zephyr-7b-beta", "HuggingFaceH4/zephyr-7b-gemma-v0.1", \
+    "meta-llama/Llama-2-7b-chat-hf", "microsoft/phi-2", \
+    "TinyLlama/TinyLlama-1.1B-Chat-v1.0", "mosaicml/mpt-7b-instruct", "tiiuae/falcon-7b-instruct", \
+    "google/flan-t5-xxl"
+]
+list_llm_simple = [os.path.basename(llm) for llm in list_llm]
 
-def initialize_database(document, chunk_size, chunk_overlap, progress=gr.Progress()):
-    logger.info("Initializing database...")
-    documents = []
-    splitter = CharacterTextSplitter(chunk_size=chunk_size, chunk_overlap=chunk_overlap)
-    
-    for file in document:
-        try:
-            loader = UnstructuredPDFLoader(file.name)
-            docs = loader.load()
-        except ImportError:
-            logger.warning("UnstructuredPDFLoader non disponibile. Tentativo di utilizzo di PyPDFLoader.")
-            try:
-                loader = PyPDFLoader(file.name)
-                docs = loader.load()
-            except ImportError:
-                logger.error("Impossibile caricare il documento PDF. Assicurati di aver installato 'unstructured' o 'pypdf'.")
-                return None, None, "Errore: Pacchetti necessari non installati. Esegui 'pip install unstructured pypdf' e riprova."
-        
-        for doc in docs:
-            text_chunks = splitter.split_text(doc.page_content)
-            for chunk in text_chunks:
-                documents.append(Document(page_content=chunk, metadata={"filename": file.name, "page": doc.metadata.get("page", 0)}))
-    
-    if not documents:
-        return None, None, "Errore: Nessun documento caricato correttamente."
-    
-    vectorstore = Chroma.from_documents(documents, embedding_function)
-    progress.update(0.5)
-    logger.info("Database initialized successfully.")
-    return vectorstore, None, "Initialized"  # Aggiunto None come secondo output
+# 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
 
-def initialize_LLM(llm_option, llm_temperature, max_tokens, top_k, vector_db, progress=gr.Progress(), language="italiano"):
-    logger.info("Initializing LLM chain...")
 
-    # Definizione del modello LLM
-    if language == "italiano":
-        model = AutoModelForCausalLM.from_pretrained("meta/llama-7b-hf")
-    else:
-        model = AutoModelForCausalLM.from_pretrained("meta/llama-7b-hf")
+# Create vector database
+def create_db(splits, collection_name):
+    embedding = HuggingFaceEmbeddings()
+    new_client = chromadb.EphemeralClient()
+    vectordb = Chroma.from_documents(
+        documents=splits,
+        embedding=embedding,
+        client=new_client,
+        collection_name=collection_name,
+        # persist_directory=default_persist_directory
+    )
+    return vectordb
 
-    tokenizer = AutoTokenizer.from_pretrained("meta/llama-7b-hf" if language == "italiano" else "meta/llama-7b-hf")
 
+# 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...")
+    # HuggingFacePipeline uses local model
+    # Note: it will download model locally...
+    # tokenizer=AutoTokenizer.from_pretrained(llm_model)
+    # progress(0.5, desc="Initializing HF pipeline...")
+    # pipeline=transformers.pipeline(
+    #     "text-generation",
+    #     model=llm_model,
+    #     tokenizer=tokenizer,
+    #     torch_dtype=torch.bfloat16,
+    #     trust_remote_code=True,
+    #     device_map="auto",
+    #     # max_length=1024,
+    #     max_new_tokens=max_tokens,
+    #     do_sample=True,
+    #     top_k=top_k,
+    #     num_return_sequences=1,
+    #     eos_token_id=tokenizer.eos_token_id
+    #     )
+    # llm = HuggingFacePipeline(pipeline=pipeline, model_kwargs={'temperature': temperature})
+    
+    # HuggingFaceHub uses HF inference endpoints
+    progress(0.5, desc="Initializing HF Hub...")
+    # Use of trust_remote_code as model_kwargs
+    # Warning: langchain issue
+    # URL: https://github.com/langchain-ai/langchain/issues/6080
+    if llm_model == "mistralai/Mixtral-8x7B-Instruct-v0.1":
+        llm = HuggingFaceEndpoint(
+            repo_id=llm_model, 
+            # model_kwargs={"temperature": temperature, "max_new_tokens": max_tokens, "top_k": top_k, "load_in_8bit": True}
+            temperature = temperature,
+            max_new_tokens = max_tokens,
+            top_k = top_k,
+            load_in_8bit = True,
+        )
+    elif llm_model in ["HuggingFaceH4/zephyr-7b-gemma-v0.1","mosaicml/mpt-7b-instruct"]:
+        raise gr.Error("LLM model is too large to be loaded automatically on free inference endpoint")
+        llm = HuggingFaceEndpoint(
+            repo_id=llm_model, 
+            temperature = temperature,
+            max_new_tokens = max_tokens,
+            top_k = top_k,
+        )
+    elif llm_model == "microsoft/phi-2":
+        # raise gr.Error("phi-2 model requires 'trust_remote_code=True', currently not supported by langchain HuggingFaceHub...")
+        llm = HuggingFaceEndpoint(
+            repo_id=llm_model, 
+            # model_kwargs={"temperature": temperature, "max_new_tokens": max_tokens, "top_k": top_k, "trust_remote_code": True, "torch_dtype": "auto"}
+            temperature = temperature,
+            max_new_tokens = max_tokens,
+            top_k = top_k,
+            trust_remote_code = True,
+            torch_dtype = "auto",
+        )
+    elif llm_model == "TinyLlama/TinyLlama-1.1B-Chat-v1.0":
+        llm = HuggingFaceEndpoint(
+            repo_id=llm_model, 
+            # model_kwargs={"temperature": temperature, "max_new_tokens": 250, "top_k": top_k}
+            temperature = temperature,
+            max_new_tokens = 250,
+            top_k = top_k,
+        )
+    elif llm_model == "meta-llama/Llama-2-7b-chat-hf":
+        raise gr.Error("Llama-2-7b-chat-hf model requires a Pro subscription...")
+        llm = HuggingFaceEndpoint(
+            repo_id=llm_model, 
+            # model_kwargs={"temperature": temperature, "max_new_tokens": max_tokens, "top_k": top_k}
+            temperature = temperature,
+            max_new_tokens = max_tokens,
+            top_k = top_k,
+        )
+    else:
+        llm = HuggingFaceEndpoint(
+            repo_id=llm_model, 
+            # model_kwargs={"temperature": temperature, "max_new_tokens": max_tokens, "top_k": top_k, "trust_remote_code": True, "torch_dtype": "auto"}
+            # model_kwargs={"temperature": temperature, "max_new_tokens": max_tokens, "top_k": top_k}
+            temperature = temperature,
+            max_new_tokens = max_tokens,
+            top_k = top_k,
+        )
+    
+    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(
-        llm=model,
-        retriever=vector_db.as_retriever(),
-        chain_type="stuff",
-        temperature=llm_temperature,
+        llm,
+        retriever=retriever,
+        chain_type="stuff", 
+        memory=memory,
+        # combine_docs_chain_kwargs={"prompt": your_prompt})
+        return_source_documents=True,
+        #return_generated_question=False,
         verbose=False,
     )
-    progress.update(1.0)
-    logger.info("LLM chain initialized successfully.")
+    progress(0.9, desc="Done!")
+    return qa_chain
+
+
+# Generate collection name for vector database
+#  - Use filepath as input, ensuring unicode text
+def create_collection_name(filepath):
+    # Extract filename without extension
+    collection_name = Path(filepath).stem
+    # Fix potential issues from naming convention
+    ## Remove space
+    collection_name = collection_name.replace(" ","-") 
+    ## ASCII transliterations of Unicode text
+    collection_name = unidecode(collection_name)
+    ## Remove special characters
+    #collection_name = re.findall("[\dA-Za-z]*", collection_name)[0]
+    collection_name = re.sub('[^A-Za-z0-9]+', '-', collection_name)
+    ## Limit length to 50 characters
+    collection_name = collection_name[:50]
+    ## Minimum length of 3 characters
+    if len(collection_name) < 3:
+        collection_name = collection_name + 'xyz'
+    ## Enforce start and end as alphanumeric character
+    if not collection_name[0].isalnum():
+        collection_name = 'A' + collection_name[1:]
+    if not collection_name[-1].isalnum():
+        collection_name = collection_name[:-1] + 'Z'
+    print('Filepath: ', filepath)
+    print('Collection name: ', collection_name)
+    return collection_name
+
+
+# Initialize database
+def initialize_database(list_file_obj, chunk_size, chunk_overlap, progress=gr.Progress()):
+    # Create list of documents (when valid)
+    list_file_path = [x.name for x in list_file_obj if x is not None]
+    # Create collection_name for vector database
+    progress(0.1, desc="Creating collection name...")
+    collection_name = create_collection_name(list_file_path[0])
+    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, history):
-    chat_history = ""
-    for item in history:
-        chat_history += f"\nUser: {item[0]}\nAI: {item[1]}"
-    chat_history += f"\n\nUser: {message}"
-    return chat_history
 
-def translate_text(text, src_lang, dest_lang):
-    translator = Translator()
-    result = translator.translate(text, src=src_lang, dest=dest_lang)
-    return result.text
+def format_chat_history(message, 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, language):
+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"]
-    if response_answer.find("Helpful Answer:")!= -1:
+    if response_answer.find("Helpful Answer:") != -1:
         response_answer = response_answer.split("Helpful Answer:")[-1]
-    
-    if language != "italian":
-        try:
-            translated_response = translate_text(response_answer, src="en", dest="it")
-        except Exception as e:
-            logger.error(f"Error translating response: {e}")
-            translated_response = response_answer
-    else:
-        translated_response = response_answer
-    
     response_sources = response["source_documents"]
     response_source1 = response_sources[0].page_content.strip()
     response_source2 = response_sources[1].page_content.strip()
     response_source3 = response_sources[2].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
     response_source3_page = response_sources[2].metadata["page"] + 1
+    # print ('chat response: ', response_answer)
+    # print('DB source', response_sources)
     
-    new_history = history + [(message, translated_response)]
+    # 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, response_source3, response_source3_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()
         qa_chain = gr.State()
         collection_name = gr.State()
-        language = gr.State(value="italian")  # Modifica qui
+        
         gr.Markdown(
-        """<center><h2>Chatbot basato su PDF</center></h2>
-        <h3>Fai domande sui tuoi documenti PDF</h3>""")
+        """<center><h2>Creatore di chatbot basato su PDF</center></h2>
+        <h3>Potete fare domande su i vostri documenti PDF</h3>""")
+        
         gr.Markdown(
-        """<b>Note:</b> Questo assistente AI, utilizzando Langchain e LLM open-source, esegue retrieval-augmented generation (RAG) dai tuoi documenti PDF. \
-        L'interfaccia utente mostra esplicitamente più passaggi per aiutare a comprendere il flusso di lavoro RAG. 
-        Questo chatbot tiene conto delle domande precedenti quando genera risposte (tramite memoria conversazionale), e include riferimenti al documento per scopi di chiarezza.<br>
-        <br><b>Avviso:</b> Questo spazio utilizza l'hardware CPU Basic gratuito da Hugging Face. Alcuni passaggi e modelli LLM utilizzati qui sotto (endpoint di inferenza gratuiti) possono richiedere del tempo per generare una risposta.
+        """<b>Nota:</b> Questo assistente IA, utilizzando Langchain e modelli LLM open source, esegue generazione aumentata da recupero (RAG) dai vostri documenti PDF. \
+        L'interfaccia utente esplicitamente mostra i passaggi multipli per aiutare a comprendere il flusso di lavoro RAG. 
+        Questo chatbot tiene conto delle domande passate nel generare le risposte (tramite memoria conversazionale), e include riferimenti ai documenti per scopi di chiarezza.<br>
+        <br><b>Avviso:</b> Questo spazio utilizza l'hardware di base CPU gratuito da Hugging Face. Alcuni passaggi e modelli LLM usati qui sotto (endpoint di inferenza gratuiti) possono richiedere del tempo per generare una risposta.
         """)
         
-        with gr.Tab("Step 1 - Carica PDF"):
-            with gr.Row():
-                document = gr.Files(height=100, file_count="multiple", file_types=["pdf"], interactive=True, label="Carica i tuoi documenti PDF (singolo o multiplo)")
-        
-        with gr.Tab("Step 2 - Processa documento"):
-            with gr.Row():
-                db_btn = gr.Radio(["ChromaDB"], label="Tipo di database vettoriale", value = "ChromaDB", type="index", info="Scegli il tuo database vettoriale")
-            with gr.Accordion("Opzioni avanzate - Divisore testo documento", open=False):
-                with gr.Row():
-                    slider_chunk_size = gr.Slider(minimum = 100, maximum = 1000, value=600, step=20, label="Dimensione chunk", info="Dimensione chunk", interactive=True)
-                with gr.Row():
-                    slider_chunk_overlap = gr.Slider(minimum = 10, maximum = 200, value=40, step=10, label=" Sovrapposizione chunk", info="Sovrapposizione chunk", interactive=True)
-            with gr.Row():
-                db_progress = gr.Textbox(label="Inizializzazione database vettoriale", value="Nessuno")
-            with gr.Row():
-                db_btn = gr.Button("Genera database vettoriale")
-            
-        with gr.Tab("Step 3 - Inizializza catena QA"):
-            with gr.Row():
-                llm_btn = gr.Radio(list_llm_simple, \
-                    label="Modelli LLM", value = list_llm_simple[0], type="index", info="Scegli il tuo modello LLM")
-            with gr.Accordion("Opzioni avanzate - Modello LLM", open=False):
-                with gr.Row():
-                    slider_temperature = gr.Slider(minimum = 0.01, maximum = 1.0, value=0.7, step=0.1, label="Temperatura", info="Temperatura del modello", interactive=True)
-                with gr.Row():
-                    slider_maxtokens = gr.Slider(minimum = 224, maximum = 4096, value=1024, step=32, label="Token massimi", info="Token massimi del modello", interactive=True)
-                with gr.Row():
-                    slider_topk = gr.Slider(minimum = 1, maximum = 10, value=3, step=1, label="Campioni top-k", info="Campioni top-k del modello", interactive=True)
-            with gr.Row():
-                llm_progress = gr.Textbox(value="Nessuno",label="Inizializzazione catena QA")
-            with gr.Row():
-                qachain_btn = gr.Button("Inizializza catena Question Answering")
+        # ... (resto del codice rimane invariato)
 
-        with gr.Tab("Step 4 - Chatbot"):
+        with gr.Tab("Passo 4 - Chatbot"):
             chatbot = gr.Chatbot(height=300)
-            with gr.Accordion("Avanzate - Riferimenti documento", open=False):
+            with gr.Accordion("Opzioni avanzate - Riferimenti ai documenti", open=False):
                 with gr.Row():
                     doc_source1 = gr.Textbox(label="Riferimento 1", lines=2, container=True, scale=20)
                     source1_page = gr.Number(label="Pagina", scale=1)
@@ -184,19 +309,18 @@ def demo():
                     doc_source3 = gr.Textbox(label="Riferimento 3", lines=2, container=True, scale=20)
                     source3_page = gr.Number(label="Pagina", scale=1)
             with gr.Row():
-                msg = gr.Textbox(placeholder="Digita un messaggio (es. 'Di cosa parla questo documento?')", container=True)
+                msg = gr.Textbox(placeholder="Inserisci messaggio (es. 'Di cosa tratta questo documento?')", container=True)
             with gr.Row():
                 submit_btn = gr.Button("Invia messaggio")
-                clear_btn = gr.ClearButton([msg, chatbot], value="Pulisci conversazione")
-            with gr.Row():
-                language_selector = gr.Radio(choices=["italiano", "inglese"], value="italiano", label="Lingua")
-            
+                clear_btn = gr.ClearButton([msg, chatbot], value="Cancella conversazione")
+           
         # 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])
         qachain_btn.click(initialize_LLM, \
-            inputs=[llm_btn, slider_temperature, slider_maxtokens, slider_topk, vector_db, language], \
+            inputs=[llm_btn, slider_temperature, slider_maxtokens, slider_topk, vector_db], \
             outputs=[qa_chain, llm_progress]).then(lambda:[None,"",0,"",0,"",0], \
             inputs=None, \
             outputs=[chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], \
@@ -204,11 +328,11 @@ def demo():
 
         # Chatbot events
         msg.submit(conversation, \
-            inputs=[qa_chain, msg, chatbot, language], \
+            inputs=[qa_chain, msg, chatbot], \
             outputs=[qa_chain, msg, chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], \
             queue=False)
         submit_btn.click(conversation, \
-            inputs=[qa_chain, msg, chatbot, language], \
+            inputs=[qa_chain, msg, chatbot], \
             outputs=[qa_chain, msg, chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], \
             queue=False)
         clear_btn.click(lambda:[None,"",0,"",0,"",0], \
@@ -217,5 +341,7 @@ def demo():
             queue=False)
     demo.queue().launch(debug=True)
 
+
 if __name__ == "__main__":
     demo()
+