app.py

import re
from typing import List
import gradio as gr
import openai
import os
from dotenv import load_dotenv
import phoenix as px
import llama_index
from llama_index import OpenAIEmbedding, Prompt, ServiceContext, VectorStoreIndex, SimpleDirectoryReader
from llama_index.chat_engine.types import ChatMode
from llama_index.llms import ChatMessage, MessageRole, OpenAI
from llama_index.vector_stores.qdrant import QdrantVectorStore
from llama_index.text_splitter import SentenceSplitter
from llama_index.extractors import TitleExtractor
from llama_index.ingestion import IngestionPipeline
from chat_template import CHAT_TEXT_QA_PROMPT
from chatbot import Chatbot, ChatbotVersion
from custom_io import UnstructuredReader, default_file_metadata_func
from qdrant import client as qdrantClient

load_dotenv()
openai.api_key = os.getenv("OPENAI_API_KEY")


class AwesumCareChatbot(Chatbot):
    DENIED_ANSWER_PROMPT = ""
    SYSTEM_PROMPT = ""
    CHAT_EXAMPLES = [
        "什麼是安心三寶？",
        "點樣立平安紙？"
    ]

    def _load_doucments(self):
        dir_reader = SimpleDirectoryReader('./awesumcare_data', file_extractor={
            ".pdf": UnstructuredReader(),
            ".docx": UnstructuredReader(),
            ".pptx": UnstructuredReader(),
        },
            recursive=True,
            exclude=["*.png", "*.pptx"],
            file_metadata=default_file_metadata_func)

        self.documents = dir_reader.load_data()
        super()._load_doucments()

    def _setup_service_context(self):
        self.service_context = ServiceContext.from_defaults(
            chunk_size=self.chunk_size,
            llm=self.llm,
            embed_model=self.embed_model
        )
        super()._setup_service_context()

    def _setup_vector_store(self):
        self.vector_store = QdrantVectorStore(
            client=qdrantClient, collection_name=self.vdb_collection_name)
        super()._setup_vector_store()

    def _setup_index(self):
        if self.vdb_collection_name in [col.name for col in qdrantClient.get_collections().collections] and qdrantClient.get_collection(self.vdb_collection_name).vectors_count > 0:
            self.index = VectorStoreIndex.from_vector_store(
                self.vector_store, service_context=self.service_context)
            print("set up index from vector store")
            return
        pipeline = IngestionPipeline(
            transformations=[
                SentenceSplitter(),
                OpenAIEmbedding(),
            ],
            vector_store=self.vector_store,
        )
        pipeline.run(documents=self.documents)
        self.index = VectorStoreIndex.from_vector_store(
            self.vector_store, service_context=self.service_context)
        super()._setup_index()

    # def _setup_index(self):
    #     self.index = VectorStoreIndex.from_documents(
    #         self.documents,
    #         service_context=self.service_context
    #     )
    #     super()._setup_index()

    def _setup_chat_engine(self):
        # testing #
        from llama_index.agent import OpenAIAgent
        from llama_index.tools.query_engine import QueryEngineTool

        query_engine = self.index.as_query_engine(
            text_qa_template=CHAT_TEXT_QA_PROMPT)
        query_engine_tool = QueryEngineTool.from_defaults(
            query_engine=query_engine)
        self.chat_engine = OpenAIAgent.from_tools(
            tools=[query_engine_tool],
            llm=self.service_context.llm,
            similarity_top_k=1,
            verbose=True
        )
        print("set up agent as chat engine")
        # testing #
        # self.chat_engine = self.index.as_chat_engine(
        #     chat_mode=ChatMode.BEST,
        #     similarity_top_k=5,
        #     text_qa_template=CHAT_TEXT_QA_PROMPT)
        super()._setup_chat_engine()


# gpt-3.5-turbo-1106, gpt-4-1106-preview
awesum_chatbot = AwesumCareChatbot(ChatbotVersion.CHATGPT_35.value,
                                   chunk_size=2048,
                                   vdb_collection_name="v2")


def vote(data: gr.LikeData):
    if data.liked:
        gr.Info("You up-voted this response: " + data.value)
    else:
        gr.Info("You down-voted this response: " + data.value)


chatbot = gr.Chatbot()

with gr.Blocks() as demo:
    gr.Markdown("# Awesum Care demo")

    with gr.Tab("With awesum care data prepared"):
        gr.ChatInterface(
            awesum_chatbot.stream_chat,
            chatbot=chatbot,
            examples=awesum_chatbot.CHAT_EXAMPLES,
        )
        chatbot.like(vote, None, None)

    with gr.Tab("With Initial System Prompt (a.k.a. prompt wrapper)"):
        gr.ChatInterface(
            awesum_chatbot.predict_with_prompt_wrapper, examples=awesum_chatbot.CHAT_EXAMPLES)

    with gr.Tab("Vanilla ChatGPT without modification"):
        gr.ChatInterface(awesum_chatbot.predict_vanilla_chatgpt,
                         examples=awesum_chatbot.CHAT_EXAMPLES)

demo.queue()
demo.launch()