Update app.py

app.py

@@ -49,14 +49,7 @@ from llama_index.core.schema import BaseNode, TextNode
 api_token = os.getenv("HF_TOKEN")
 
 
-# default_persist_directory = './chroma_HF/'
-list_llm = ["mistralai/Mistral-7B-Instruct-v0.3", "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 = ["mistralai/Miceli", "mistralai/Mistral-7B-Instruct-v0.3"]  
 list_llm_simple = [os.path.basename(llm) for llm in list_llm]
 
 # Load PDF document and create doc splits
@@ -94,55 +87,18 @@ def initialize_llmchain(llm_model, temperature, max_tokens, top_k, vector_db, pr
     
     progress(0.5, desc="Initializing HF Hub...")
     
-    if llm_model == "mistralai/Mixtral-8x7B-Instruct-v0.3":
-        llm = HuggingFaceEndpoint(
-            repo_id=llm_model, 
-            api_key=api_token,
-            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, 
-            api_key=api_token,
-            temperature = temperature,
-            max_new_tokens = max_tokens,
-            top_k = top_k,
-        )
-    elif llm_model == "microsoft/phi-2":
-        llm = HuggingFaceEndpoint(
-            repo_id=llm_model, 
-            api_key=api_token,
-            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":
+    if llm_model == "mistralai/Mistral-7B-Instruct-v0.2":
         llm = HuggingFaceEndpoint(
             repo_id=llm_model,
-            api_key=api_token,
-            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, 
-            api_key=api_token,
+            huggingfacehub_api_token = api_token,
             temperature = temperature,
             max_new_tokens = max_tokens,
             top_k = top_k,
         )
     else:
         llm = HuggingFaceEndpoint(
+            huggingfacehub_api_token = api_token,
             repo_id=llm_model, 
-            api_key=api_token,
             temperature = temperature,
             max_new_tokens = max_tokens,
             top_k = top_k,
@@ -209,6 +165,7 @@ 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)
    
@@ -234,121 +191,118 @@ def upload_file(file_obj):
         list_file_path.append(file_path)
     return list_file_path
 
-# Initialize LlamaIndex parsing
-def initialize_llama_index(file_obj):
-    documents = LlamaParse(result_type="markdown", api_key=api_token).load_data(file_obj[0].name)
-    node_parser = MarkdownElementNodeParser(llm=None, num_workers=8)
-    nodes = node_parser.get_nodes_from_documents(documents)
-    base_nodes, objects = node_parser.get_nodes_and_objects(nodes)
-    index_with_obj = VectorStoreIndex(nodes=base_nodes + objects)
-    index_ret = index_with_obj.as_retriever(top_k=15)
-    recursive_query_engine = RetrieverQueryEngine.from_args(index_ret, node_postprocessors=[FlagEmbeddingReranker(
-        top_n=5,
-        model="BAAI/bge-reranker-large",
-    )], verbose=False)
-    return recursive_query_engine, "LlamaIndex parsing complete"
+list_llm = ["mistralai/Miceli", "mistralai/Mistral-7B-Instruct-v0.3"]  
+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 = []
+    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()
+    vectordb = Chroma.from_documents(
+        documents=splits,
+        embedding=embedding,
+        client=new_client,
+        collection_name=collection_name,
+    )
+    return vectordb
 
-def demo():
-    with gr.Blocks(theme="base") as demo:
-        vector_db = gr.State()
-        qa_chain = gr.State()
-        collection_name = gr.State()
-        llama_index_engine = gr.State()
-        
-        gr.Markdown(
-        """<center><h2>PDF-based chatbot</center></h2>
-        <h3>Ask any questions about your PDF documents</h3>""")
-        gr.Markdown(
-        """<b>Note:</b> Esta é a lucIAna, primeira Versão da IA para seus PDF documentos. \
-        Este chatbot leva em consideração perguntas anteriores ao gerar respostas (por meio de memória conversacional) e inclui referências a documentos para fins de clareza.
-        """)
-        
-        with gr.Tab("Step 1 - Upload PDF"):
-            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.Tab("Step 2 - Process document"):
-            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):
-                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 3 - Initialize QA chain"):
-            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):
-                with gr.Row():
-                    slider_temperature = gr.Slider(minimum=0.01, 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")
+# Load vector database
+def load_db():
+    embedding = HuggingFaceEmbeddings()
+    vectordb = Chroma(
+        embedding_function=embedding)
+    return vectordb
 
-        with gr.Tab("Step 4 - LlamaIndex parsing"):
-            with gr.Row():
-                llama_index_btn = gr.Button("Parse with LlamaIndex")
-            with gr.Row():
-                llama_index_progress = gr.Textbox(label="LlamaIndex parsing status", value="None")
+# 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...")
+    
+    if llm_model == "mistralai/Mistral-7B-Instruct-v0.2":
+        llm = HuggingFaceEndpoint(
+            repo_id=llm_model,
+            huggingfacehub_api_token = api_token,
+            temperature = temperature,
+            max_new_tokens = max_tokens,
+            top_k = top_k,
+        )
+    else:
+        llm = HuggingFaceEndpoint(
+            huggingfacehub_api_token = api_token,
+            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()
+    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
 
-        with gr.Tab("Step 5 - Chatbot"):
-            chatbot = gr.Chatbot(height=300)
-            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():
-                    doc_source3 = gr.Textbox(label="Reference 3", lines=2, container=True, scale=20)
-                    source3_page = gr.Number(label="Page", scale=1)
-            with gr.Row():
-                msg = gr.Textbox(placeholder="Type message (e.g. 'What is this document about?')", container=True)
-            with gr.Row():
-                submit_btn = gr.Button("Submit message")
-                clear_btn = gr.ClearButton([msg, chatbot], value="Clear conversation")
-            
-        # 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,"",0], 
-            inputs=None, 
-            outputs=[chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], 
-            queue=False)
-        llama_index_btn.click(initialize_llama_index, 
-            inputs=[document], 
-            outputs=[llama_index_engine, llama_index_progress])
+# Generate collection name for vector database
+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'
+    print('Filepath: ', filepath)
+    print('Collection name: ', collection_name)
+    return collection_name
 
-        # 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], 
-            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], 
-            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], 
-            queue=False)
-    demo.queue().launch(debug=True)
+# Initialize database
+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 name...")
+    collection_name = create_collection_name(list_file_path[0])
+    progress(0.25, desc="Loading document...")
+    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, "Complete!"
 
+def initialize_LLM(llm_option, llm_temperature, max_tokens, top_k, vector_db, progress=gr.Progress()):
+    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!"
 
-if __name__ == "__main__":
-    demo()
+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