Create app.py

app.py

@@ -0,0 +1,188 @@
+import gradio as gr
+from langchain_community.document_loaders import PyPDFLoader
+from langchain.text_splitter import RecursiveCharacterTextSplitter
+from langchain_community.vectorstores import Chroma
+from langchain_huggingface import HuggingFacePipeline
+from langchain_huggingface import HuggingFaceEmbeddings
+from langchain.chains import ConversationalRetrievalChain
+from langchain.memory import ConversationBufferMemory
+
+from pathlib import Path
+import chromadb
+from unidecode import unidecode
+from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, pipeline
+import re
+
+# Constants
+LLM_MODEL = "t5-large"  # Using a larger model for better performance and longer responses
+LLM_MAX_TOKEN = 1024
+DB_CHUNK_SIZE = 512
+CHUNK_OVERLAP = 24
+TEMPERATURE = 0.1
+MAX_TOKENS = 1024
+TOP_K = 20
+pdf_url = "https://huggingface.co/spaces/CCCDev/PDFChat/resolve/main/Privacy-Policy%20(1).pdf"  # Replace with your static PDF URL or path
+
+# Load PDF document and create doc splits
+def load_doc(pdf_url, chunk_size, chunk_overlap):
+    loader = PyPDFLoader(pdf_url)
+    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(model_name="sentence-transformers/all-MiniLM-L6-v2")
+    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(0.5, desc="Initializing HF Hub...")
+    
+    tokenizer = AutoTokenizer.from_pretrained(llm_model)
+    model = AutoModelForSeq2SeqLM.from_pretrained(llm_model)
+    summarization_pipeline = pipeline("summarization", model=model, tokenizer=tokenizer)
+    pipe = HuggingFacePipeline(pipeline=summarization_pipeline)
+
+    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=pipe,
+        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'
+    return collection_name
+
+# Initialize database
+def initialize_database(pdf_url, chunk_size, chunk_overlap, progress=gr.Progress()):
+    collection_name = create_collection_name(pdf_url)
+    progress(0.25, desc="Loading document...")
+    doc_splits = load_doc(pdf_url, 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_temperature, max_tokens, top_k, vector_db, progress=gr.Progress()):
+    qa_chain = initialize_llmchain(LLM_MODEL, 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 "Helpful Answer:" in response_answer:
+        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 explicitly 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 4 - 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")
+
+        # Automatic preprocessing
+        db_progress = gr.Textbox(label="Vector database initialization", value="Initializing...")
+        db_btn = gr.Button("Generate vector database", visible=False)
+        qachain_btn = gr.Button("Initialize Question Answering chain", visible=False)
+        llm_progress = gr.Textbox(value="None", label="QA chain initialization")
+
+        def auto_initialize():
+            vector_db, collection_name, db_status = initialize_database(pdf_url, DB_CHUNK_SIZE, CHUNK_OVERLAP)
+            qa_chain, llm_status = initialize_LLM(TEMPERATURE, LLM_MAX_TOKEN, 20, vector_db)
+            return vector_db, collection_name, db_status, qa_chain, llm_status, "Initialization complete."
+
+        demo.load(auto_initialize, [], [vector_db, collection_name, db_progress, qa_chain, llm_progress])
+
+        # 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)
+    return demo.queue().launch(debug=True)
+
+if __name__ == "__main__":
+    demo()