Create app.py

app.py

@@ -0,0 +1,134 @@
+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.chains import ConversationalRetrievalChain
+from langchain_community.embeddings import HuggingFaceEmbeddings 
+from langchain_community.llms import HuggingFacePipeline, HuggingFaceEndpoint
+from langchain.memory import ConversationBufferMemory
+from pathlib import Path
+import chromadb
+import re
+
+def load_doc(list_file_path, chunk_size=600, chunk_overlap=40):
+    loaders = [PyPDFLoader(x) for x in list_file_path]
+    pages = [page for loader in loaders for page in 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()
+    client = chromadb.EphemeralClient()
+    vectordb = Chroma.from_documents(
+        documents=splits,
+        embedding=embedding,
+        client=client,
+        collection_name=collection_name,
+    )
+    return vectordb
+
+def initialize_llmchain(llm_model, vector_db, progress=gr.Progress()):
+    llm = HuggingFaceEndpoint(
+        repo_id=llm_model,
+        temperature=0.7,
+        max_new_tokens=1024,
+        top_k=3,
+    )
+    memory = ConversationBufferMemory(memory_key="chat_history", output_key='answer', return_messages=True)
+    retriever = vector_db.as_retriever()
+    qa_chain = ConversationalRetrievalChain.from_llm(
+        llm,
+        retriever=retriever,
+        chain_type="stuff",
+        memory=memory,
+        return_source_documents=True,
+        verbose=False,
+    )
+    return qa_chain
+
+def create_collection_name(filepath):
+    collection_name = Path(filepath).stem
+    collection_name = re.sub('[^A-Za-z0-9]+', '-', collection_name)[:50]
+    if len(collection_name) < 3:
+        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, progress=gr.Progress()):
+    list_file_path = [x.name for x in list_file_obj if x is not None]
+    collection_name = create_collection_name(list_file_path[0])
+    doc_splits = load_doc(list_file_path)
+    vector_db = create_db(doc_splits, collection_name)
+    return vector_db, collection_name, "Complete!"
+
+def initialize_LLM(llm_model, vector_db, progress=gr.Progress()):
+    qa_chain = initialize_llmchain(llm_model, vector_db, progress)
+    return qa_chain, "Complete!"
+
+def conversation(qa_chain, message, history):
+    formatted_chat_history = [(f"User: {user_message}", f"Assistant: {bot_message}") for user_message, bot_message in 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 (powered by LangChain and open-source LLMs)</center></h2>
+            <h3>Ask any questions about your PDF documents, along with follow-ups</h3>
+            <b>Note:</b> This AI assistant performs retrieval-augmented generation from your PDF documents.
+            When generating answers, it takes past questions into account (via conversational memory), and includes document references for clarity purposes.
+            <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 an output.<br>
+            """)
+
+        with gr.Tab("Step 1 - Document pre-processing"):
+            document = gr.Files(height=100, file_count="multiple", file_types=["pdf"], interactive=True, label="Upload your PDF documents (single or multiple)")
+            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 2 - QA chain initialization"):
+            llm_btn = gr.Radio(["mistralai/Mistral-7B-Instruct-v0.2"], label="LLM models", value="mistralai/Mistral-7B-Instruct-v0.2", type="index", info="Choose your LLM model")
+            with gr.Row():
+                llm_progress = gr.Textbox(value="None", label="QA chain initialization")
+            with gr.Row():
+                qachain_btn = gr.Button("Initialize question-answering chain...")
+
+        with gr.Tab("Step 3 - Conversation with chatbot"):
+            chatbot = gr.Chatbot(height=300)
+            with gr.Row():
+                msg = gr.Textbox(placeholder="Type message", container=True)
+            with gr.Row():
+                submit_btn = gr.Button("Submit")
+                clear_btn = gr.ClearButton([msg, chatbot])
+
+        db_btn.click(initialize_database, inputs=[document], outputs=[vector_db, collection_name, db_progress])
+        qachain_btn.click(initialize_LLM, inputs=[llm_btn, vector_db], outputs=[qa_chain, llm_progress]).then(lambda:[None,"",0,"",0,"",0], inputs=None, outputs=[chatbot], queue=False)
+
+        msg.submit(conversation, inputs=[qa_chain, msg, chatbot], outputs=[qa_chain, msg, chatbot], queue=False)
+        submit_btn.click(conversation, inputs=[qa_chain, msg, chatbot], outputs=[qa_chain, msg, chatbot], queue=False)
+        clear_btn.click(lambda:[None,"",0,"",0,"",0], inputs=None, outputs=[chatbot], queue=False)
+
+    demo.queue().launch(debug=True)
+
+if __name__ == "__main__":
+    demo()