Upload 8 files

ChatErector.py

@@ -0,0 +1,30 @@
+from llm.utils import initialize_LLM, format_chat_history, postprocess
+from db.utils import initialize_database
+import gradio as gr
+import spaces
+
+def initializer(list_file_obj, llm_temperature, max_tokens, top_k, thold, progress=gr.Progress()):
+    vdb=initialize_database(list_file_obj)
+    qa_chain=initialize_LLM(llm_temperature, max_tokens, top_k, vdb, thold)#, progress)
+    return qa_chain, "Success."
+
+
+@spaces.GPU
+def conversation(qa_chain, message, history):
+    formatted_chat_history = format_chat_history(message, history)
+    # Generate response using QA chain
+    response = qa_chain.invoke({"question": message, "chat_history": formatted_chat_history})
+    response_answer = postprocess(response)#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()
+    # Langchain sources are zero-based
+    #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
+    # Append user message and response to chat history
+    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

CustomRetriever.py

@@ -0,0 +1,47 @@
+from langchain.schema.retriever import BaseRetriever
+from langchain_core.documents import Document
+from typing import List
+
+from langchain.callbacks.manager import CallbackManagerForRetrieverRun
+
+from langchain_core.documents import Document
+from langchain_core.runnables import chain
+
+class CustomRetriever():
+    def __init__(self, v_db, thold=0.7):
+        
+
+        #self.retriever=RetrieverWithScores()
+
+        class RetrieverWithScores(BaseRetriever):
+                #def __init__(self, vdb):
+                    #self.vdb=vdb
+                #def __init__(self, retriever: BaseRetriever):  # Add an __init__ to store the existing retriever
+                    #super().__init__(retriever=retriever)
+                def _get_relevant_documents(
+                    self, query: str, *, run_manager: CallbackManagerForRetrieverRun)-> List[Document]:
+
+                    @chain
+                    def retr_func(query: str)-> List[Document]: #(vdb, query: str)-> List[Document]:
+                        docs, scores = zip(*v_db.similarity_search_with_relevance_scores(query))#similarity_search_with_score(query))
+                        result=[]
+                        for doc, score in zip(docs, scores):
+                            if score>thold:
+                                doc.metadata["score"] = score
+                                result.append(doc)
+                        if len(result)==0:
+                            result.append(Document(metadata={}, page_content='No data'))
+
+                        return result #docs
+
+                    return retr_func.invoke(query) #(self.vdb, query)
+
+        self.retriever=RetrieverWithScores()
+        
+
+    
+    
+
+
+
+

README.md.txt

@@ -0,0 +1,12 @@
+---
+title: RAG Test1
+emoji: 🐢
+colorFrom: yellow
+colorTo: blue
+sdk: gradio
+sdk_version: 4.31.0
+app_file: app.py
+pinned: false
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

app.py

@@ -0,0 +1,3 @@
+from ui.gradio_ui import ui
+
+ui()

db/utils.py

@@ -0,0 +1,44 @@
+from langchain_community.document_loaders import PyPDFLoader
+from langchain.text_splitter import RecursiveCharacterTextSplitter
+from langchain_community.embeddings import HuggingFaceEmbeddings
+from langchain_community.vectorstores import FAISS
+import gradio as gr
+
+
+# Load and split PDF document
+def load_doc(list_file_path):
+    # Processing for one document only
+    # loader = PyPDFLoader(file_path)
+    # pages = loader.load()
+    loaders = [PyPDFLoader(x) for x in list_file_path]
+    pages = []
+    for loader in loaders:
+        pages.extend(loader.load())
+    text_splitter = RecursiveCharacterTextSplitter(
+        chunk_size = 1024,
+        chunk_overlap = 64
+    )
+    doc_splits = text_splitter.split_documents(pages)
+    return doc_splits
+
+
+
+def create_db(splits):
+    model_kwargs = {'device': 'cpu'}
+
+    embeddings = HuggingFaceEmbeddings(model_name="BAAI/bge-small-en", model_kwargs =model_kwargs)
+    vectordb = FAISS.from_documents(splits, embeddings)
+    return vectordb
+
+
+def initialize_database(list_file_obj, progress=gr.Progress()):
+    # Create a list of documents (when valid)
+    list_file_path = [x.name for x in list_file_obj if x is not None]
+    # Load document and create splits
+    doc_splits = load_doc(list_file_path)
+    # Create or load vector database
+    vector_db = create_db(doc_splits)
+    return vector_db #, "Database created!"
+
+
+

llm/utils.py

@@ -0,0 +1,74 @@
+from langchain_community.llms import HuggingFaceEndpoint
+from langchain.memory import ConversationBufferMemory
+from langchain.chains import ConversationalRetrievalChain
+import gradio as gr
+import os
+from CustomRetriever import CustomRetriever
+
+
+API_TOKEN=os.getenv("TOKEN")
+
+# Initialize langchain LLM chain
+def initialize_llmchain(llm_model, temperature, max_tokens, top_k, vdb, 
+                        thold=0.8, progress=gr.Progress()):
+    llm = HuggingFaceEndpoint(
+            huggingfacehub_api_token = API_TOKEN,
+            repo_id=llm_model,
+            temperature = temperature,
+            max_new_tokens = max_tokens,
+            top_k = top_k,
+        )
+
+    memory = ConversationBufferMemory(
+        memory_key="chat_history",
+        output_key='answer',
+        return_messages=True
+    )
+
+    retr=CustomRetriever(vdb, thold=thold)
+    retriever=retr.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
+
+
+
+# Initialize LLM
+def initialize_LLM(llm_temperature, max_tokens, top_k, vector_db, thold, progress=gr.Progress()):
+    # print("llm_option",llm_option)
+    llm_name = "mistralai/Mistral-7B-Instruct-v0.2" #list_llm[llm_option]
+    #print("llm_name: ",llm_name)
+    qa_chain = initialize_llmchain(llm_name, llm_temperature, max_tokens, top_k, vector_db, thold)
+    return qa_chain #, "QA chain initialized. Chatbot is ready!"
+
+
+
+def format_chat_history(chat_history):#message, chat_history): #no need message
+    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 postprocess(response):
+    try:
+        result=response["answer"]
+        for doc in response['source_documents']:
+            file_doc="\n\nFile: " + doc.metadata["source"]
+            page="\nPage: " + str(doc.metadata["page"])
+            content="\nFragment: " + doc.page_content.strip()
+            result+=file_doc+page+content
+        return result
+    except:
+        return response["answer"]
+
+
+

requirements.txt

@@ -0,0 +1,10 @@
+torch
+transformers
+sentence-transformers
+langchain
+langchain-community
+tqdm
+accelerate
+pypdf
+faiss-cpu
+#faiss-gpu

ui/gradio_ui.py

@@ -0,0 +1,88 @@
+import gradio as gr
+from ChatErector import conversation, initializer
+
+def ui():
+    # with gr.Blocks(theme=gr.themes.Default(primary_hue="sky")) as demo:
+    with gr.Blocks(theme=gr.themes.Default(primary_hue="red", secondary_hue="pink", neutral_hue = "sky")) as ui:
+        #vector_db = gr.State()
+        qa_chain = gr.State()
+        gr.HTML("<center><h1>RAG PDF chatbot</h1><center>")
+        gr.Markdown("""<b>Query your PDF documents!</b> This AI agent is designed to perform retrieval augmented generation (RAG) on PDF documents. The app is hosted on Hugging Face Hub for the sole purpose of demonstration. \
+        <b>Please do not upload confidential documents.</b>
+        """)
+        with gr.Row():
+            with gr.Column(scale = 86):
+                gr.Markdown("<b>Step 1 - Upload PDF documents and Initialize RAG pipeline</b>")
+                with gr.Row():
+                    document = gr.Files(height=300, file_count="multiple", file_types=["pdf"], interactive=True, label="Upload PDF documents")
+                #with gr.Row():
+                    #db_btn = gr.Button("Create vector database")
+                #with gr.Row():
+                        #db_progress = gr.Textbox(value="Not initialized", show_label=False) # label="Vector database status", 
+                #gr.Markdown("<style>body { font-size: 16px; }</style><b>Advanced settings</b>")
+                #with gr.Row():
+                    #llm_btn = gr.Radio(list_llm_simple, label="Available LLMs", value = list_llm_simple[0], type="index") # info="Select LLM", show_label=False
+                with gr.Row():
+                    with gr.Accordion("Advanced settings", open=False):
+                        with gr.Row():
+                            slider_temperature = gr.Slider(minimum = 0.01, maximum = 1.0, value=0.5, step=0.1, label="Temperature", info="Controls randomness in token generation", interactive=True)
+                        with gr.Row():
+                            slider_maxtokens = gr.Slider(minimum = 128, maximum = 9192, value=4096, step=128, label="Max New Tokens", info="Maximum number of tokens to be generated",interactive=True)
+                        with gr.Row():
+                            slider_topk = gr.Slider(minimum = 1, maximum = 10, value=3, step=1, label="top-k", info="Number of tokens to select the next token from", interactive=True)
+                        with gr.Row():
+                            thold = gr.Slider(minimum = 0.01, maximum = 1.0, value=0.8, step=0.1, label="Treshold", info="Retrieved information relevance level", interactive=True)
+                with gr.Row():
+                    qachain_btn = gr.Button("Initialize Question Answering Chatbot")
+                with gr.Row():
+                        llm_progress = gr.Textbox(value="Not initialized", show_label=False) # label="Chatbot status", 
+
+            with gr.Column(scale = 200):
+                gr.Markdown("<b>Step 2 - Chat with your Document</b>")
+                chatbot = gr.Chatbot(height=505)
+                #with gr.Accordion("Relevent context from the source document", 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="Ask a question", container=True)
+                with gr.Row():
+                    submit_btn = gr.Button("Submit")
+                    clear_btn = gr.ClearButton([msg, chatbot], value="Clear")
+            
+        # Preprocessing events
+        #db_btn.click(initialize_database, \
+            #inputs=[document], \
+            #outputs=[vector_db, db_progress])
+        qachain_btn.click(initializer, \
+            inputs=[document, slider_temperature, slider_maxtokens, slider_topk, thold], \
+            outputs=[qa_chain, llm_progress]).then(lambda:[None,"",0,"",0,"",0], \
+            inputs=None, \
+            outputs=[chatbot] \ 
+            queue=False)
+
+        # Chatbot events
+        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)
+    ui.queue().launch(debug=True)
+
+
+
+
+