Upload 6 files

.huggingface/spaces.yaml

@@ -0,0 +1,3 @@
+sdk: gradio
+sdk_version: 4.26.0
+app_file: app.py

app.py

@@ -0,0 +1,106 @@
+import os
+import gradio as gr
+from dotenv import load_dotenv
+
+import indexing
+import retrieval
+
+list_llm = [
+    "mistralai/Mistral-7B-Instruct-v0.3",
+    "mistralai/Mixtral-8x7B-Instruct-v0.1",
+    "HuggingFaceH4/zephyr-7b-beta"
+]
+list_llm_simple = [os.path.basename(llm) for llm in list_llm]
+
+def retrieve_api():
+    load_dotenv()
+    global huggingfacehub_api_token
+    huggingfacehub_api_token = os.environ.get("HUGGINGFACE_API_KEY")
+
+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 = indexing.create_collection_name(list_file_path[0])
+    progress(0.25, desc="Loading document...")
+    doc_splits = indexing.load_doc(list_file_path, chunk_size, chunk_overlap)
+    progress(0.5, desc="Generating vector database...")
+    vector_db = indexing.create_db(doc_splits, collection_name)
+    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]
+    qa_chain = retrieval.initialize_llmchain(
+        llm_name, huggingfacehub_api_token, llm_temperature, max_tokens, top_k, vector_db, progress
+    )
+    return qa_chain, "Complete!"
+
+def conversation(qa_chain, message, history):
+    qa_chain, new_history, response_sources = retrieval.invoke_qa_chain(qa_chain, message, history)
+    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
+    return (
+        qa_chain,
+        gr.update(value=""),
+        new_history,
+        response_source1,
+        response_source1_page,
+        response_source2,
+        response_source2_page,
+        response_source3,
+        response_source3_page,
+    )
+
+with gr.Blocks(theme="base") as demo:
+    vector_db = gr.State()
+    qa_chain = gr.State()
+    collection_name = gr.State()
+    gr.Markdown("""# PDF Chatbot with RAG""")
+    with gr.Row():
+        document = gr.File(file_types=[".pdf"], file_count="multiple", label="Upload PDF(s)")
+    with gr.Accordion("Document Settings", open=False):
+        slider_chunk_size = gr.Slider(100, 1000, value=600, step=20, label="Chunk size")
+        slider_chunk_overlap = gr.Slider(10, 200, value=40, step=10, label="Chunk overlap")
+    db_progress = gr.Textbox(label="DB Init Status", value="None")
+    db_btn = gr.Button("Generate vector database")
+
+    db_btn.click(
+        initialize_database,
+        inputs=[document, slider_chunk_size, slider_chunk_overlap],
+        outputs=[vector_db, collection_name, db_progress],
+    )
+
+    with gr.Row():
+        llm_btn = gr.Radio(list_llm_simple, label="Choose LLM", type="index")
+    with gr.Accordion("LLM Parameters", open=False):
+        slider_temperature = gr.Slider(0.01, 1.0, value=0.7, step=0.1, label="Temperature")
+        slider_maxtokens = gr.Slider(224, 4096, value=1024, step=32, label="Max Tokens")
+        slider_topk = gr.Slider(1, 10, value=3, step=1, label="Top-k")
+    llm_progress = gr.Textbox(label="LLM Status", value="None")
+    qachain_btn = gr.Button("Initialize QA Chain")
+
+    qachain_btn.click(
+        initialize_llm,
+        inputs=[llm_btn, slider_temperature, slider_maxtokens, slider_topk, vector_db],
+        outputs=[qa_chain, llm_progress],
+    )
+
+    chatbot = gr.Chatbot(height=300)
+    msg = gr.Textbox(placeholder="Ask something...", label="Message")
+    submit_btn = gr.Button("Submit")
+    clear_btn = gr.ClearButton([msg, chatbot])
+
+    doc_source1 = gr.Textbox(label="Reference 1")
+    source1_page = gr.Number(label="Page 1")
+    doc_source2 = gr.Textbox(label="Reference 2")
+    source2_page = gr.Number(label="Page 2")
+    doc_source3 = gr.Textbox(label="Reference 3")
+    source3_page = gr.Number(label="Page 3")
+
+    msg.submit(conversation, [qa_chain, msg, chatbot], [qa_chain, msg, chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page])
+    submit_btn.click(conversation, [qa_chain, msg, chatbot], [qa_chain, msg, chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page])
+
+demo.queue().launch()

indexing.py

@@ -0,0 +1,83 @@
+"""
+Indexing with vector database
+"""
+
+from pathlib import Path
+import re
+
+import chromadb
+
+from unidecode import unidecode
+
+from langchain_community.document_loaders import PyPDFLoader
+from langchain.text_splitter import RecursiveCharacterTextSplitter
+from langchain_chroma import Chroma
+from langchain_huggingface import HuggingFaceEmbeddings
+
+
+
+# Load PDF document and create doc splits
+def load_doc(list_file_path, chunk_size, chunk_overlap):
+    """Load PDF document and create doc splits"""
+
+    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
+
+
+# Generate collection name for vector database
+#  - Use filepath as input, ensuring unicode text
+#  - Handle multiple languages (arabic, chinese)
+def create_collection_name(filepath):
+    """Create collection name for vector database"""
+
+    # Extract filename without extension
+    collection_name = Path(filepath).stem
+    # Fix potential issues from naming convention
+    ## Remove space
+    collection_name = collection_name.replace(" ", "-")
+    ## ASCII transliterations of Unicode text
+    collection_name = unidecode(collection_name)
+    ## Remove special characters
+    collection_name = re.sub("[^A-Za-z0-9]+", "-", collection_name)
+    ## Limit length to 50 characters
+    collection_name = collection_name[:50]
+    ## Minimum length of 3 characters
+    if len(collection_name) < 3:
+        collection_name = collection_name + "xyz"
+    ## Enforce start and end as alphanumeric character
+    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("\n\nFilepath: ", filepath)
+    print("Collection name: ", collection_name)
+    return collection_name
+
+
+# Create vector database
+def create_db(splits, collection_name):
+    """Create embeddings and vector database"""
+
+    embedding = HuggingFaceEmbeddings(
+        model_name="sentence-transformers/paraphrase-multilingual-mpnet-base-v2",
+        # model_name="sentence-transformers/all-MiniLM-L6-v2",
+        # model_kwargs={"device": "cpu"},
+        # encode_kwargs={'normalize_embeddings': False}
+    )
+    chromadb.api.client.SharedSystemClient.clear_system_cache()
+    new_client = chromadb.EphemeralClient()
+    vectordb = Chroma.from_documents(
+        documents=splits,
+        embedding=embedding,
+        client=new_client,
+        collection_name=collection_name,
+        # persist_directory=default_persist_directory
+    )
+    return vectordb

prompt_template.json

@@ -0,0 +1,5 @@
+{
+	"title": "System prompt",
+	"prompt": "You are an assistant for question-answering tasks. Use the following pieces of context to answer the question at the end. If you don't know the answer, just say that you don't know, don't try to make up an answer. Keep the answer concise. Question: {question}  \\n Context: {context} \\n Helpful Answer:"
+}
+ 

requirements.txt

@@ -0,0 +1,11 @@
+gradio==4.26.0
+langchain
+chromadb
+pypdf
+python-dotenv
+huggingface_hub
+unidecode
+sentence-transformers
+langchain-community
+langchain-core
+langchain-chroma

retrieval.py

@@ -0,0 +1,114 @@
+"""
+LLM chain retrieval
+"""
+
+import json
+import gradio as gr
+
+from langchain.chains.conversational_retrieval.base import ConversationalRetrievalChain
+from langchain.memory import ConversationBufferMemory
+from langchain_huggingface import HuggingFaceEndpoint
+from langchain_core.prompts import PromptTemplate
+
+
+# Add system template for RAG application
+PROMPT_TEMPLATE = """
+You are an assistant for question-answering tasks. Use the following pieces of context to answer the question at the end. 
+If you don't know the answer, just say that you don't know, don't try to make up an answer. Keep the answer concise.
+Question: {question} 
+Context: {context} 
+Helpful Answer:
+"""
+
+
+# Initialize langchain LLM chain
+def initialize_llmchain(
+    llm_model,
+    huggingfacehub_api_token,
+    temperature,
+    max_tokens,
+    top_k,
+    vector_db,
+    progress=gr.Progress(),
+):
+    """Initialize Langchain LLM chain"""
+
+    progress(0.1, desc="Initializing HF tokenizer...")
+    # HuggingFaceHub uses HF inference endpoints
+    progress(0.5, desc="Initializing HF Hub...")
+    # Use of trust_remote_code as model_kwargs
+    # Warning: langchain issue
+    # URL: https://github.com/langchain-ai/langchain/issues/6080
+
+    llm = HuggingFaceEndpoint(
+        repo_id=llm_model,
+        task="text-generation",
+        temperature=temperature,
+        max_new_tokens=max_tokens,
+        top_k=top_k,
+        huggingfacehub_api_token=huggingfacehub_api_token,
+    )
+
+    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_type="similarity", search_kwargs={'k': 3})
+    retriever = vector_db.as_retriever()
+
+    progress(0.8, desc="Defining retrieval chain...")
+    with open('prompt_template.json', 'r') as file:
+    	system_prompt = json.load(file)
+    prompt_template = system_prompt["prompt"]
+    rag_prompt = PromptTemplate(
+        template=prompt_template, input_variables=["context", "question"]
+    )
+    qa_chain = ConversationalRetrievalChain.from_llm(
+        llm,
+        retriever=retriever,
+        chain_type="stuff",
+        memory=memory,
+        combine_docs_chain_kwargs={"prompt": rag_prompt},
+        return_source_documents=True,
+        # return_generated_question=False,
+        verbose=False,
+    )
+    progress(0.9, desc="Done!")
+
+    return qa_chain
+
+
+def format_chat_history(message, chat_history):
+    """Format chat history for llm chain"""
+
+    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 invoke_qa_chain(qa_chain, message, history):
+    """Invoke question-answering chain"""
+
+    formatted_chat_history = format_chat_history(message, history)
+    # print("formatted_chat_history",formatted_chat_history)
+
+    # Generate response using QA chain
+    response = qa_chain.invoke(
+        {"question": message, "chat_history": formatted_chat_history}
+    )
+
+    response_sources = response["source_documents"]
+
+    response_answer = response["answer"]
+    if response_answer.find("Helpful Answer:") != -1:
+        response_answer = response_answer.split("Helpful Answer:")[-1]
+
+    # Append user message and response to chat history
+    new_history = history + [(message, response_answer)]
+
+    # print ('chat response: ', response_answer)
+    # print('DB source', response_sources)
+
+    return qa_chain, new_history, response_sources