Update app.py

app.py

@@ -1,5 +1,274 @@
 import gradio as gr
 import os
+from langchain_community.document_loaders import PyPDFLoader
+from langchain.text_splitter import RecursiveCharacterTextSplitter
+from langchain_community.vectorstores import Chroma
+from langchain.chains import ConversationalRetrievalChain
+from langchain_community.embeddings import HuggingFaceEmbeddings 
+from langchain_community.llms import HuggingFaceEndpoint
+from langchain.memory import ConversationBufferMemory
+from pathlib import Path
+import chromadb
+from unidecode import unidecode
+import re
+
+# List of available LLM models
+list_llm = [
+    "mistralai/Mistral-7B-Instruct-v0.2", "mistralai/Mixtral-8x7B-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 load_doc(list_file_path, chunk_size, chunk_overlap):
+    loaders = [PyPDFLoader(x) for x in list_file_path]
+    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
+
+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
+    )
+    return vectordb
+
+def initialize_llmchain(llm_model, temperature, max_tokens, top_k, vector_db, progress=gr.Progress()):
+    progress(0.5, desc="Initializing HF Hub...")
+    llm = HuggingFaceEndpoint(
+        repo_id=llm_model,
+        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_kwargs={'k': 5})  # Increased from 3 to 5
+    progress(0.8, desc="Defining retrieval chain...")
+    qa_chain = ConversationalRetrievalChain.from_llm(
+        llm,
+        retriever=retriever,
+        chain_type="stuff", 
+        memory=memory,
+        return_source_documents=True,
+        verbose=False,
+    )
+    progress(0.9, desc="Done!")
+    return qa_chain
+
+def create_collection_name(filepath):
+    collection_name = Path(filepath).stem
+    collection_name = collection_name.replace(" ", "-")
+    collection_name = unidecode(collection_name)
+    collection_name = re.sub('[^A-Za-z0-9]+', '-', collection_name)
+    collection_name = collection_name[:50]
+    if len(collection_name) < 3:
+        collection_name = collection_name + 'xyz'
+    if not collection_name[0].isalnum():
+        collection_name = 'A' + collection_name[1:]
+    if not collection_name[-1].isalnum():
+        collection_name = collection_name[:-1] + 'Z'
+    return collection_name
+
+def initialize_database(list_file_obj, chunk_size, chunk_overlap, progress=gr.Progress()):
+    list_file_path = [x.name for x in list_file_obj if x is not None]
+    progress(0.1, desc="Creating collection...")
+    collection_name = create_collection_name(list_file_path[0])
+    progress(0.25, desc="Loading documents...")
+    doc_splits = load_doc(list_file_path, chunk_size, chunk_overlap)
+    
+    progress(0.5, desc="Generating vector database...")
+    vector_db = create_db(doc_splits, collection_name)
+    progress(0.9, desc="Done!")
+
+    return vector_db, collection_name, "Completed!"
+
+def initialize_LLM(llm_option, llm_temperature, max_tokens, top_k, vector_db, progress=gr.Progress()):
+    llm_name = list_llm[llm_option]
+    qa_chain = initialize_llmchain(llm_name, llm_temperature, max_tokens, top_k, vector_db, progress)
+    return qa_chain, "Completed!"
+
+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):
+    formatted_chat_history = format_chat_history(message, history)
+    response = qa_chain({"question": message, "chat_history": formatted_chat_history})
+    response_answer = response["answer"]
+    if response_answer.find("Helpful Answer:") != -1:
+        response_answer = response_answer.split("Helpful Answer:")[-1]
+    response_sources = response["source_documents"]
+    
+    source_info = []
+    for i in range(min(5, len(response_sources))):  # Increased from 3 to 5
+        source = response_sources[i]
+        source_info.append({
+            'content': source.page_content.strip(),
+            'page': source.metadata["page"] + 1
+        })
+    
+    new_history = history + [(message, response_answer)]
+    return qa_chain, gr.update(value=""), new_history, *[info['content'] for info in source_info], *[info['page'] for info in source_info]
+
+# The rest of the code (demo function and UI setup) remains largely the same,
+# but update the outputs of the conversation function to handle 5 sources instead of 3.
+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()
+        
+        gr.Markdown(
+        """<center><h2>Creatore di chatbot basato su PDF</center></h2>
+        <h3>Potete fare domande su i vostri documenti PDF</h3>""")
+        
+        gr.Markdown(
+        """<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 PDFs"):
+            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.Tab("Step 2 - Processa i documenti"):
+            with gr.Row():
+                db_btn = gr.Radio(["ChromaDB"], label="Vector database type", value = "ChromaDB", type="index", info="Choose your vector database")
+            with gr.Accordion("Opzioni Avanzate - Document text splitter", open=False):
+                with gr.Row():
+                    slider_chunk_size = gr.Slider(minimum = 100, maximum = 1000, value=1000, step=20, label="Chunk size", info="Chunk size", interactive=True)
+                with gr.Row():
+                    slider_chunk_overlap = gr.Slider(minimum = 10, maximum = 200, value=100, 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("Genera vector database")
+            
+        with gr.Tab("Step 3 - Inizializza QA chain"):
+            with gr.Row():
+                llm_btn = gr.Radio(list_llm_simple, \
+                    label="LLM models", value = list_llm_simple[5], type="index", info="Scegli il tuo modello LLM")
+            with gr.Accordion("Advanced options - LLM model", open=False):
+                with gr.Row():
+                    slider_temperature = gr.Slider(minimum = 0.01, maximum = 1.0, value=0.3, 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():
+                language_btn = gr.Radio(["Italian", "English"], label="Linua", value="Italian", type="index", info="Seleziona la lingua per il chatbot")
+            with gr.Row():
+                llm_progress = gr.Textbox(value="None",label="QA chain initialization")
+            with gr.Row():
+                qachain_btn = gr.Button("Inizializza Question Answering chain")
+        
+        
+        with gr.Tab("Passo 4 - Chatbot"):
+            chatbot = gr.Chatbot(height=300)
+            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)
+                with gr.Row():
+                    doc_source2 = gr.Textbox(label="Riferimento 2", lines=2, container=True, scale=20)
+                    source2_page = gr.Number(label="Pagina", scale=1)
+                with gr.Row():
+                    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="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="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], \
+            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], \
+            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, 
+                     doc_source3, source3_page,
+                     doc_source4, source4_page,
+                     doc_source5, source5_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, 
+                     doc_source3, source3_page,
+                     doc_source4, source4_page,
+                     doc_source5, source5_page], \ 
+            queue=False)
+        clear_btn.click(lambda:[None,"",0,"",0,"",0], \
+            inputs=None, \
+            outputs=[chatbot, \ 
+                     doc_source1, source1_page, 
+                     doc_source2, source2_page, 
+                     doc_source3, source3_page,
+                     doc_source4, source4_page,
+                     doc_source5, source5_page], \
+            queue=False)
+    demo.queue().launch(debug=True)
+
+
+if __name__ == "__main__":
+    demo()
+
+
+
+
+
+#####################################################
+import gradio as gr
+import os
 
 from langchain_community.document_loaders import PyPDFLoader
 from langchain.text_splitter import RecursiveCharacterTextSplitter