app.py

from functools import cache
import os
import gradio as gr
from langchain.llms import OpenAI
from langchain.chains.summarize import load_summarize_chain
from langchain.document_loaders import TextLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.vectorstores import Chroma
from langchain.embeddings.openai import OpenAIEmbeddings
from langchain.chains.question_answering import load_qa_chain
import tiktoken


@cache
def tiktoken_len_builder(model_name):
    tokenizer = tiktoken.encoding_for_model(model_name)

    def token_len(text):
        tokens = tokenizer.encode(text, disallowed_special=())
        return len(tokens)

    return token_len


def split_documents(docs, length_function, chunk_size=400):
    text_splitter = RecursiveCharacterTextSplitter(
        chunk_size=chunk_size,
        chunk_overlap=20,
        length_function=length_function,
    )
    return text_splitter.split_documents(docs)


def summarize_docs(llm, docs):
    chain = load_summarize_chain(llm, chain_type="map_reduce")
    return chain.run(docs)


class MdnaQA:
    def __init__(self, llm, docs):
        self.docs = docs
        self.chain = load_qa_chain(llm, chain_type="stuff")
        embeddings = OpenAIEmbeddings(openai_api_key=llm.openai_api_key)
        self.docsearch = Chroma.from_documents(docs, embeddings)

    def ask(self, question):
        input_documents = self.docsearch.similarity_search(question)
        return self.chain.run(input_documents=input_documents, question=question)


filename = "2023-05-12_2023_q1_goog_mdna.txt"
loader = TextLoader(filename)
documents = loader.load()
model_name = "text-davinci-003"
tiktoken_len = tiktoken_len_builder(model_name)
docs = split_documents(documents, tiktoken_len)
tokens_sum = sum(tiktoken_len(d.page_content) for d in docs)

title = "Alphabet's Q1 2023 10-Q MD&A"
video = '<iframe width="560" height="315" src="https://www.youtube.com/embed/LuXtsWQfmFg" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" allowfullscreen></iframe>'

with gr.Blocks(title=title) as demo:
    gr.Markdown(f"# {title}")
    gr.HTML(video)
    gr.Markdown("Blog post https://blog.experienced.dev/earnings-report-insights-programmer-decodes-alphabets-q1-2023-10-q-form/")
    gr.Markdown(
        "You can get an API key [from OpenAI](https://platform.openai.com/account/api-keys)"
    )
    openai_api_key = gr.Text(
        value=os.getenv("OPENAI_API_KEY"),
        type="password",
        label="OpenAI API key",
    )
    temperature = gr.Slider(
        0,
        2,
        value=0,
        step=0.1,
        label="Temperature",
        info="adjusts a model's output from predictable to random",
    )
    mdna = gr.State(docs)
    tokens_total = gr.Textbox(
        label="Total input tokens",
        value=tokens_sum,
        info="how many tokens will be spent on input / embeddings",
    )
    with gr.Tabs(visible=True) as tabs:
        with gr.TabItem("Summary"):
            summarize = gr.Button(
                "Summarize MD&A",
                variant="primary",
                info="On click you spent tokens on input, instructions and output",
            )
            summary = gr.TextArea(label="Summary")

            def summarize_mdna(docs, api_key, temp):
                llm = OpenAI(temperature=temp, openai_api_key=api_key)
                mdna_summary = summarize_docs(llm, docs)
                return mdna_summary

            summarize.click(
                summarize_mdna,
                inputs=[mdna, openai_api_key, temperature],
                outputs=[summary],
            )
        with gr.TabItem("QA with MD&A"):
            start_qa = gr.Button("Start QA with MD&A", variant="primary")
            chatbot = gr.Chatbot(label="QA with MD&A", visible=False)
            question = gr.Textbox(
                label="Your question", interactive=True, visible=False
            )
            qa_chat = gr.State()
            send = gr.Button("Ask question", variant="primary", visible=False)

            def start_chat(docs, api_key, temp):
                llm = OpenAI(temperature=temp, openai_api_key=api_key)
                qa_chat = MdnaQA(llm, docs)
                return (
                    qa_chat,
                    gr.Textbox.update(visible=True),
                    gr.Textbox.update(visible=True),
                    gr.Button.update(visible=True),
                )

            start_qa.click(
                start_chat,
                [mdna, openai_api_key, temperature],
                [qa_chat, chatbot, question, send],
            )

            def respond(qa_chat, question, chat_history):
                answer = qa_chat.ask(question)
                chat_history.append((question, answer))
                return "", chat_history

            send.click(respond, [qa_chat, question, chatbot], [question, chatbot])
            question.submit(respond, [qa_chat, question, chatbot], [question, chatbot])


demo.launch()