Update app.py

app.py

@@ -21,22 +21,22 @@ import tqdm
 import accelerate
 
 
-#Set parameters
 
-llm_model = 'mistralai/Mixtral-8x7B-Instruct-v0.1'
-list_file_obj = '/home/user/app/pdfs/'
-chunk_size = 1024
-chunk_overlap = 128
-temperature = 0.1
-max_tokens = 6000
-top_k = 3
+# default_persist_directory = './chroma_HF/'
+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", "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]
 
-
-def load_doc(list_file_path):
+# Load PDF document and create doc splits
+def load_doc(list_file_path, chunk_size, chunk_overlap):
     # Processing for one document only
     # loader = PyPDFLoader(file_path)
     # pages = loader.load()
-    loaders = [PyPDFLoader(list_file_obj+x) for x in list_file_path]
+    loaders = [PyPDFLoader(x) for x in list_file_path]
     pages = []
     for loader in loaders:
         pages.extend(loader.load())
@@ -48,7 +48,6 @@ def load_doc(list_file_path):
     return doc_splits
 
 
-
 # Create vector database
 def create_db(splits, collection_name):
     embedding = HuggingFaceEmbeddings()
@@ -62,6 +61,7 @@ def create_db(splits, collection_name):
     )
     return vectordb
 
+
 # Load vector database
 def load_db():
     embedding = HuggingFaceEmbeddings()
@@ -70,20 +70,99 @@ def load_db():
         embedding_function=embedding)
     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...")
+    # HuggingFacePipeline uses local model
+    # Note: it will download model locally...
+    # tokenizer=AutoTokenizer.from_pretrained(llm_model)
+    # progress(0.5, desc="Initializing HF pipeline...")
+    # pipeline=transformers.pipeline(
+    #     "text-generation",
+    #     model=llm_model,
+    #     tokenizer=tokenizer,
+    #     torch_dtype=torch.bfloat16,
+    #     trust_remote_code=True,
+    #     device_map="auto",
+    #     # max_length=1024,
+    #     max_new_tokens=max_tokens,
+    #     do_sample=True,
+    #     top_k=top_k,
+    #     num_return_sequences=1,
+    #     eos_token_id=tokenizer.eos_token_id
+    #     )
+    # llm = HuggingFacePipeline(pipeline=pipeline, model_kwargs={'temperature': temperature})
     
+    # 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
+    if llm_model == "mistralai/Mixtral-8x7B-Instruct-v0.1":
+        llm = HuggingFaceHub(
+            repo_id=llm_model, 
+            model_kwargs={"temperature": temperature, "max_new_tokens": max_tokens, "top_k": top_k, "load_in_8bit": True}
+        )
+    elif llm_model == "microsoft/phi-2":
+        raise gr.Error("phi-2 model requires 'trust_remote_code=True', currently not supported by langchain HuggingFaceHub...")
+        llm = HuggingFaceHub(
+            repo_id=llm_model, 
+            model_kwargs={"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 = HuggingFaceHub(
+            repo_id=llm_model, 
+            model_kwargs={"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...")
+        llm = HuggingFaceHub(
+            repo_id=llm_model, 
+            model_kwargs={"temperature": temperature, "max_new_tokens": max_tokens, "top_k": top_k}
+        )
+    else:
+        llm = HuggingFaceHub(
+            repo_id=llm_model, 
+            # model_kwargs={"temperature": temperature, "max_new_tokens": max_tokens, "top_k": top_k, "trust_remote_code": True, "torch_dtype": "auto"}
+            model_kwargs={"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(search_type="similarity", search_kwargs={'k': 3})
+    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,
+        # combine_docs_chain_kwargs={"prompt": your_prompt})
+        return_source_documents=True,
+        #return_generated_question=False,
+        verbose=False,
+    )
+    progress(0.9, desc="Done!")
+    return qa_chain
+
+
 # Initialize database
-def initialize_database(list_file_obj):
+def initialize_database(list_file_obj, chunk_size, chunk_overlap, progress=gr.Progress()):
     # Create list of documents (when valid)
-    #list_file_path = [x.name for x in list_file_obj if x is not None]
-    list_file_path = os.listdir(list_file_obj)
+    list_file_path = [x.name for x in list_file_obj if x is not None]
     # Create collection_name for vector database
+    progress(0.1, desc="Creating collection name...")
     collection_name = Path(list_file_path[0]).stem
     # Fix potential issues from naming convention
     ## Remove space
     collection_name = collection_name.replace(" ","-") 
     ## Limit lenght to 50 characters
     collection_name = collection_name[:50]
-    print(collection_name)
     ## Enforce start and end as alphanumeric character
     if not collection_name[0].isalnum():
         collection_name[0] = 'A'
@@ -91,32 +170,24 @@ def initialize_database(list_file_obj):
         collection_name[-1] = 'Z'
     # print('list_file_path: ', list_file_path)
     print('Collection name: ', collection_name)
+    progress(0.25, desc="Loading document...")
     # Load document and create splits
-    doc_splits = load_doc(list_file_path)
+    doc_splits = load_doc(list_file_path, chunk_size, chunk_overlap)
     # Create or load vector database
+    progress(0.5, desc="Generating vector database...")
     # global vector_db
     vector_db = create_db(doc_splits, collection_name)
-    return vector_db, collection_name
+    progress(0.9, desc="Done!")
+    return vector_db, collection_name, "Complete!"
 
 
-def initialize_llmchain(vector_db):
-    # Initialize langchain LLM chain
-    llm = HuggingFaceHub(repo_id = llm_model,model_kwargs={"temperature": temperature, 
-                                                           "max_new_tokens": max_tokens, 
-                                                           "top_k": top_k, 
-                                                           "load_in_8bit": True})
-    retriever=vector_db.as_retriever()
-    memory = ConversationBufferMemory(memory_key="chat_history", output_key='answer', return_messages=True)
-    qa_chain = ConversationalRetrievalChain.from_llm(llm,retriever=retriever,chain_type="stuff", 
-                                                         memory=memory,return_source_documents=True,verbose=False)
-        
-    return qa_chain
-
-def initialize_LLM(vector_db):
+def initialize_LLM(llm_option, llm_temperature, max_tokens, top_k, vector_db, progress=gr.Progress()):
     # print("llm_option",llm_option)
-    llm_name = llm_model
-    qa_chain = initialize_llmchain(vector_db)
-    return qa_chain
+    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 = []
@@ -124,6 +195,7 @@ def format_chat_history(message, 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)
@@ -148,57 +220,60 @@ def conversation(qa_chain, message, history):
     # Append user message and response to chat history
     new_history = history + [(message, response_answer)]
     # return gr.update(value=""), new_history, response_sources[0], response_sources[1] 
-    return qa_chain, new_history, response_source1, response_source1_page, response_source2, response_source2_page, response_source3, response_source3_page
+    return qa_chain, gr.update(value=""), new_history, response_source1, response_source1_page, response_source2, response_source2_page, response_source3, response_source3_page
+    
+
+def upload_file(file_obj):
+    list_file_path = []
+    for idx, file in enumerate(file_obj):
+        file_path = file_obj.name
+        list_file_path.append(file_path)
+    # print(file_path)
+    # initialize_database(file_path, progress)
+    return list_file_path
+
 
 def demo():
-    with gr.Blocks() as demo:
+    with gr.Blocks(theme="base") as demo:
         vector_db = gr.State()
         qa_chain = gr.State()
         collection_name = gr.State()
 
+
+        document = gr.Files(value = '/home/user/app/pdfs/Annual-Report-2022-2023-English_1.pdf',height=100, 
+                            file_count="multiple", file_types=["pdf"], interactive=True, visible=False,
+                            label="Upload your PDF documents (single or multiple)")
         chatbot = gr.Chatbot(height=300)
-        with gr.Accordion("References", open=True):
+        with gr.Accordion("Advanced - Document references", open=False):
             with gr.Row():
-                doc_source1 = gr.Textbox(label="Reference 1", lines=5, container=True, scale=20)
+                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=5, container=True, scale=20)
+                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=5, container=True, scale=20)
+                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", container=True)
-            with gr.Row():
-                #db_btn = gr.Button('Initialize database')
-                #qachain_btn = gr.Button('Start chatbot')
-                submit_btn = gr.Button("Submit")
-                clear_btn = gr.ClearButton([msg, chatbot])
-
-        # document = list_file_obj
-        vector_db, collection_name = initialize_database(list_file_obj)
-        print(collection_name)
-        # #upload_btn.upload(upload_file, inputs=[upload_btn], outputs=[document])
-        # db_btn.click(initialize_database, \
-        #     inputs=[document], \
-        #     outputs=[vector_db, collection_name])
-        
-        # qachain_btn.click(initialize_LLM, \
-        #     inputs=[vector_db], \
-        #     outputs=[qa_chain]).then(lambda:[0], \
-        #     inputs=None, \
-        #     outputs=[chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], \
-        #     queue=False)    
-
-        #qachain = initialize_LLM(vector_db)
-        llm = HuggingFaceHub(repo_id = llm_model,model_kwargs={"temperature": temperature, 
-                                                           "max_new_tokens": max_tokens, 
-                                                           "top_k": top_k, 
-                                                           "load_in_8bit": True})
-        retriever=vector_db.as_retriever()
-        memory = ConversationBufferMemory(memory_key="chat_history", output_key='answer', return_messages=True)
-        qa_chain = ConversationalRetrievalChain.from_llm(llm,retriever=retriever,chain_type="stuff", 
-                                                         memory=memory,return_source_documents=True,verbose=False)
+        with gr.Row():
+            msg = gr.Textbox(placeholder="Type message", container=True)
+        with gr.Row():
+            db_btn = gr.Button("Generate vector database...")
+            qachain_btn = gr.Button("Initialize question-answering chain...")
+            submit_btn = gr.Button("Submit")
+            clear_btn = gr.ClearButton([msg, chatbot])
+            
+        # Preprocessing events
+        #upload_btn.upload(upload_file, inputs=[upload_btn], outputs=[document])
+        db_btn.click(initialize_database, \
+            inputs=[document, slider_chunk_size, slider_chunk_overlap], \
+            outputs=[vector_db, collection_name, db_progress])
+        qachain_btn.click(initialize_LLM, \
+            inputs=[llm_btn, slider_temperature, slider_maxtokens, slider_topk, vector_db], \
+            outputs=[qa_chain, llm_progress]).then(lambda:[None,"",0,"",0,"",0], \
+            inputs=None, \
+            outputs=[chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], \
+            queue=False)
+
         # Chatbot events
         msg.submit(conversation, \
             inputs=[qa_chain, msg, chatbot], \
@@ -214,5 +289,6 @@ def demo():
             queue=False)
     demo.queue().launch(debug=True)
 
+
 if __name__ == "__main__":
     demo()