Create app.py

app.py

@@ -0,0 +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 HuggingFacePipeline
+from langchain.chains import ConversationChain
+from langchain.memory import ConversationBufferMemory
+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 tqdm
+import accelerate
+import re
+
+# Static PDF file link
+static_pdf_link = "https://huggingface.co/spaces/CCCDev/PDFChat/resolve/main/Data-privacy-policy.pdf"
+
+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]
+
+
+# Load PDF document and create doc splits
+def load_doc(file_path, chunk_size, chunk_overlap):
+    loader = PyPDFLoader(file_path)
+    pages = 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
+
+
+# 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/Mixtral-8x7B-Instruct-v0.1":
+        llm = HuggingFaceEndpoint(
+            repo_id=llm_model,
+            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")
+    elif llm_model == "microsoft/phi-2":
+        llm = HuggingFaceEndpoint(
+            repo_id=llm_model,
+            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,
+            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...")
+    else:
+        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()
+    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
+
+
+# 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
+
+
+# Initialize database
+def initialize_database(chunk_size, chunk_overlap, progress=gr.Progress()):
+    file_path = static_pdf_link
+    progress(0.1, desc="Creating collection name...")
+    collection_name = create_collection_name(file_path)
+    progress(0.25, desc="Loading document...")
+    doc_splits = load_doc(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!"
+
+
+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"]
+    response_source1 = response_sources[0].page_content.strip()
+    response_source2 = response_sources[1].page_content.strip()
+    response_source3 = response_sources[2].page_content.strip()
+    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
+
+    new_history = history + [(message, response_answer)]
+    return qa_chain, gr.update(
+        value=""), new_history, response_source1, response_source1_page, response_source2, response_source2_page, response_source3, response_source3_page
+
+
+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>PDF-based chatbot</center></h2>
+            <h3>Ask any questions about your PDF documents</h3>""")
+        gr.Markdown(
+            """<b>Note:</b> This AI assistant, using Langchain and open-source LLMs, performs retrieval-augmented generation (RAG) from your PDF documents. \
+            The user interface explicitely shows multiple steps to help understand the RAG workflow. 
+            This chatbot takes past questions into account when generating answers (via conversational memory), and includes document references for clarity purposes.<br>
+            <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 a reply.
+            """)
+
+        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():
+                    chunk_size = gr.Slider(64, 4096, value=512, step=32, label="Text chunk size",
+                                           info="Text length of each document chunk being embedded into the vector database. Default is 512.")
+                    chunk_overlap = gr.Slider(0, 1024, value=24, step=8, label="Text chunk overlap",
+                                              info="Text overlap between each document chunk being embedded into the vector database. Default is 24.")
+
+            initialize_db = gr.Button("Process document")
+
+            with gr.Row():
+                output_db = gr.Textbox(label="Database initialization steps", placeholder="", show_label=False)
+                with gr.Accordion("Vector database collection details", open=False):
+                    collection = gr.Textbox(label="Collection name", placeholder="", show_label=False)
+
+        with gr.Tab("Step 3 - Initialize LLM"):
+            with gr.Row():
+                llm_options = gr.Dropdown(list_llm_simple, label="Choose open-source LLM",
+                                          value="Mistral-7B-Instruct-v0.2",
+                                          info="Choose among the proposed open-source LLMs")
+            with gr.Accordion("Advanced LLM options", open=False):
+                with gr.Row():
+                    llm_temperature = gr.Slider(0.01, 1.0, value=0.1, step=0.01, label="LLM temperature",
+                                                info="LLM sampling temperature, in [0.01,1.0] range. Default is 0.1")
+                    llm_max_tokens = gr.Slider(32, 1024, value=512, step=16, label="Max tokens",
+                                               info="Maximum number of new tokens to be generated, in [32,1024] range. Default is 512")
+                    llm_top_k = gr.Slider(1, 40, value=20, step=1, label="Top K",
+                                          info="The number of highest probability vocabulary tokens to keep for top-k-filtering. Default is 20.")
+
+            initialize_llm = gr.Button("Initialize LLM")
+
+            with gr.Row():
+                output_llm = gr.Textbox(label="LLM initialization steps", placeholder="", show_label=False)
+
+        with gr.Tab("Step 4 - Start chatting"):
+            chatbot = gr.Chatbot(label="PDF chatbot", height=500)
+            msg = gr.Textbox(label="Your question", placeholder="Type your question here...", show_label=False)
+            clear = gr.Button("Clear chat")
+
+            with gr.Accordion("Document sources (3)", open=False):
+                gr.Markdown("Source 1")
+                response_src1 = gr.Textbox(label="Source 1", placeholder="", show_label=False)
+                response_src1_page = gr.Number(label="Page number", value=0, precision=0, interactive=False)
+                gr.Markdown("Source 2")
+                response_src2 = gr.Textbox(label="Source 2", placeholder="", show_label=False)
+                response_src2_page = gr.Number(label="Page number", value=0, precision=0, interactive=False)
+                gr.Markdown("Source 3")
+                response_src3 = gr.Textbox(label="Source 3", placeholder="", show_label=False)
+                response_src3_page = gr.Number(label="Page number", value=0, precision=0, interactive=False)
+
+        initialize_db.click(initialize_database,
+                            inputs=[chunk_size, chunk_overlap],
+                            outputs=[vector_db, collection_name, output_db])
+        initialize_llm.click(initialize_LLM,
+                             inputs=[llm_options, llm_temperature, llm_max_tokens, llm_top_k, vector_db],
+                             outputs=[qa_chain, output_llm])
+        msg.submit(conversation,
+                   inputs=[qa_chain, msg, chatbot],
+                   outputs=[chatbot, msg, chatbot, response_src1, response_src1_page, response_src2, response_src2_page,
+                            response_src3, response_src3_page])
+        clear.click(lambda: None, None, chatbot, queue=False)
+        clear.click(lambda: None, None, msg, queue=False)
+
+    return demo.queue().launch(debug=True)
+
+
+# demo().launch(server_name="0.0.0.0")
+if __name__ == "__main__":
+    demo()