app.py

import datetime
from google.protobuf import message
import torch
import time
import threading
import streamlit as st
import random
from typing import Iterable
from unsloth import FastLanguageModel
from transformers import AutoTokenizer, AutoModelForCausalLM, TextIteratorStreamer, PreTrainedTokenizerFast
from datetime import datetime
from threading import Thread

fine_tuned_model_name = "jed-tiotuico/twitter-llama"
sota_model_name = "unsloth/mistral-7b-instruct-v0.2-bnb-4bit"
# fine_tuned_model_name = "MBZUAI/LaMini-GPT-124M"
# sota_model_name = "MBZUAI/LaMini-GPT-124M"

alpaca_input_text_format = "### Instruction:\n{}\n\n### Response:\n"
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# if device is cpu try mps?
if device == "cpu":
    # check if mps is available
    device = torch.device("mps" if torch.backends.mps.is_available() else "cpu")

def get_model_tokenizer(sota_model_name):
    model, tokenizer = FastLanguageModel.from_pretrained(
        model_name = "jed-tiotuico/twitter-llama",
        max_seq_length = 200,
        dtype = None,
        load_in_4bit = True,
        cache_dir = "/data/.cache/hf-models",
        token=st.secrets["HF_TOKEN"]
    )
    FastLanguageModel.for_inference(model)

    return model, tokenizer

def write_user_chat_message(user_chat, customer_msg):
    if customer_msg:
        if user_chat == None:
            user_chat = st.chat_message("user")

        user_chat.write(customer_msg)

def write_stream_user_chat_message(user_chat, model, token, prompt):
    if prompt:
        if user_chat == None:
            user_chat = st.chat_message("user")

        new_customer_msg = user_chat.write_stream(
            stream_generation(
                prompt,
                show_prompt=False,
                tokenizer=tokenizer,
                model=model,
            )
        )

        return new_customer_msg

def get_mistral_model_tokenizer(sota_model_name):
    max_seq_length = 2048
    dtype = torch.float16
    load_in_4bit = True # Use 4bit quantization to reduce memory usage. Can be False.
    model, tokenizer = FastLanguageModel.from_pretrained(
        model_name = "unsloth/mistral-7b-instruct-v0.2-bnb-4bit",
        max_seq_length = max_seq_length,
        dtype = dtype,
        load_in_4bit = load_in_4bit,
        cache_dir = "/data/.cache/hf-models",
    )
    FastLanguageModel.for_inference(model)

    return model, tokenizer

class DeckPicker:
    def __init__(self, items):
        self.items = items[:]  # Make a copy of the items to shuffle
        self.original_items = items[:]  # Keep the original order
        random.shuffle(self.items)  # Shuffle the items
        self.index = -1  # Initialize the index

    def pick(self):
        """Pick the next item from the deck. If all items have been picked, reshuffle."""
        self.index += 1
        if self.index >= len(self.items):
            self.index = 0
            random.shuffle(self.items)  # Reshuffle if at the end
        return self.items[self.index]

    def get_state(self):
        """Return the current state of the deck and the last picked index."""
        return self.items, self.index

# Example of usage
nouns = [
    "service", "issue", "account", "support", "problem", "help", "team",
    "request", "response", "email", "ticket", "update", "error", "system",
    "connection", "downtime", "billing", "charge", "refund", "password",
    "outage", "agent", "feature", "access", "status", "interface", "network",
    "subscription", "upgrade", "notification", "data", "server", "log", "message",
    "renewal", "setup", "security", "feedback", "confirmation", "printer"
]

verbs = [
    "have", "print", "need", "help", "update", "resolve", "access", "contact",
    "receive", "reset", "support", "experience", "report", "request", "process",
    "check", "confirm", "explain", "manage", "handle", "disconnect", "renew",
    "change", "fix", "cancel", "complete", "notify", "respond", "fail", "restore",
    "review", "escalate", "submit", "configure", "troubleshoot", "log", "operate",
    "suspend", "pay", "adjust"
]

adjectives = [
    "quick", "immediate", "urgent", "unable", "detailed", "frequent", "technical",
    "possible", "slow", "helpful", "unresponsive", "secure", "successful", "necessary",
    "available", "scheduled", "regular", "interrupted", "automatic", "manual", "last",
    "online", "offline", "new", "current", "prior", "due", "related", "temporary",
    "permanent", "next", "previous", "complicated", "easy", "difficult", "major",
    "minor", "alternative", "additional", "expired"
]

def create_few_shots(noun_picker, verb_picker, adjective_picker):
  noun = noun_picker.pick()
  verb = verb_picker.pick()
  adjective = adjective_picker.pick()

  context = f"""
Write a short realistic customer support tweet message by a customer for another company.
Avoid adding hashtags or mentions in the message.
Ensure that the sentiment is negative.
Ensure that the word count is around 15 to 25 words.
Ensure the message contains the noun: {noun}, verb: {verb}, and adjective: {adjective}.

Example of return messages 5/5:

1/5: your website is straight up garbage. how do you sell high end technology but you cant get a website right?
2/5: my phone is all static during calls and when i plug in headphones any audio still comes thru the speaks wtf
3/5: hi, i'm having trouble logging into my groceries account it keeps refreshing back to the log in page, any ideas?
4/5: please check you dms asap if you're really about customer service. 2 weeks since my accident and nothing.
5/5: I'm extremely disappointed with your service. You charged me for a temporary solution, and there's no adjustment in sight.

Now it's your turn, ensure to only generate one message
1/1:
"""
  return context

st.header("ReplyCaddy")
st.write("AI-powered customer support assistant. Reduces anxiety when responding to customer support on social media.")
# image https://github.com/unslothai/unsloth/blob/main/images/made%20with%20unsloth.png?raw=true
# st.write("Made with [Unsloth](https://github.com/unslothai/unsloth/blob/main/images/made%20with%20unsloth.png?raw=true")

def stream_generation(
    prompt: str,
    tokenizer: PreTrainedTokenizerFast,
    model: AutoModelForCausalLM,
    max_new_tokens: int = 2048,
    temperature: float = 0.7,
    top_p: float = 0.9,
    top_k: int = 100,
    repetition_penalty: float = 1.1,
    penalty_alpha: float = 0.25,
    no_repeat_ngram_size: int = 3,
    show_prompt: bool = False,
) -> Iterable[str]:
    """
    Stream the generation of a prompt.

    Args:
        prompt (str): the prompt
        max_new_tokens (int, optional): the maximum number of tokens to generate. Defaults to 32.
        temperature (float, optional): the temperature of the generation. Defaults to 0.7.
        top_p (float, optional): the top-p value of the generation. Defaults to 0.9.
        top_k (int, optional): the top-k value of the generation. Defaults to 100.
        repetition_penalty (float, optional): the repetition penalty of the generation. Defaults to 1.1.
        penalty_alpha (float, optional): the penalty alpha of the generation. Defaults to 0.25.
        no_repeat_ngram_size (int, optional): the no repeat ngram size of the generation. Defaults to 3.
        show_prompt (bool, optional): whether to show the prompt or not. Defaults to False.
        tokenizer (PreTrainedTokenizerFast): the tokenizer
        model (AutoModelForCausalLM): the model

    Yields:
        str: the generated text
    """
    # init the streaming object with tokenizer
    # skip_prompt = not show_prompt, skip_special_tokens = True
    streamer = TextIteratorStreamer(tokenizer, skip_prompt=not show_prompt, skip_special_tokens=True)  # type: ignore

    # setup kwargs for generation
    generation_kwargs = dict(
        input_ids = tokenizer(prompt, return_tensors="pt")["input_ids"].to(device),
        streamer=streamer,
        do_sample=True,
        temperature=temperature,
        top_p=top_p,
        top_k=top_k,
        repetition_penalty=repetition_penalty,
        penalty_alpha=penalty_alpha,
        no_repeat_ngram_size=no_repeat_ngram_size,
        max_new_tokens=max_new_tokens,
    )

    # start the generation in a separate thread
    generation_thread = threading.Thread(
        target=model.generate, kwargs=generation_kwargs  # type: ignore
    )
    generation_thread.start()

    blacklisted_tokens = ["<|url|>"]
    for new_text in streamer:
        # filter out blacklisted tokens
        if any(token in new_text for token in blacklisted_tokens):
            continue

        yield new_text

    # wait for the generation to finish
    generation_thread.join()

twitter_llama_model = None
twitter_llama_tokenizer = None
streamer = None

# define state and the chat messages
def init_session_states(assistant_chat, user_chat):
    if "user_msg_as_prompt" not in st.session_state:
        st.session_state["user_msg_as_prompt"] = ""

user_chat = None
if "user_msg_as_prompt" in st.session_state:
    user_chat = st.chat_message("user")

assistant_chat = st.chat_message("assistant")
if "greet" not in st.session_state:
    st.session_state["greet"] = False
    greeting_text = "Hello! I'm here to help. Copy and paste your customer's message, or generate using AI."
    assistant_chat.write(greeting_text)

init_session_states(assistant_chat, user_chat)

# Generate Response Tweet
if user_chat:
    if st.button("Generate Polite and Friendly Response"):
        if "user_msg_as_prompt" in st.session_state:
            customer_msg = st.session_state["user_msg_as_prompt"]
            if customer_msg:
                write_user_chat_message(user_chat, customer_msg)

                model, tokenizer = get_model_tokenizer(sota_model_name)

                input_text = alpaca_input_text_format.format(customer_msg)
                st.markdown(f"""```\n{input_text}```""", unsafe_allow_html=True)
                response_tweet = assistant_chat.write_stream(
                    stream_generation(
                        input_text,
                        show_prompt=False,
                        tokenizer=tokenizer,
                        model=model,
                    )
                )
            else:
                st.error("Please enter a customer message, or generate one for the ai to respond")

# below ui prompt
# - examples
# st.markdown("<b>Example:</b>", unsafe_allow_html=True)
if st.button("your website is straight up garbage. how do you sell high end technology but you cant get a website right?"):
    customer_msg = "your website is straight up garbage. how do you sell high end technology but you cant get a website right?"
    st.session_state["user_msg_as_prompt"] = customer_msg
    write_user_chat_message(user_chat, customer_msg)
    model, tokenizer = get_model_tokenizer(sota_model_name)
    input_text = alpaca_input_text_format.format(customer_msg)
    st.write(f"```\n{input_text}```")
    assistant_chat.write_stream(
        stream_generation(
            input_text,
            show_prompt=False,
            tokenizer=tokenizer,
            model=model,
        )
    )

# - Generate Customer Tweet
if st.button("Generate Customer Message using Few Shots"):
    model, tokenizer = get_mistral_model_tokenizer(sota_model_name)

    noun_picker = DeckPicker(nouns)
    verb_picker = DeckPicker(verbs)
    adjective_picker = DeckPicker(adjectives)
    few_shots = create_few_shots(noun_picker, verb_picker, adjective_picker)
    few_shot_prompt = f"<s>[INST]{few_shots}[/INST]\n"
    st.markdown("Prompt:")
    st.markdown(f"""```\n{few_shot_prompt}```""", unsafe_allow_html=True)

    new_customer_msg = write_stream_user_chat_message(user_chat, model, tokenizer, few_shot_prompt)
    st.session_state["user_msg_as_prompt"] = new_customer_msg


st.markdown("------------")
st.markdown("<p>Thanks to:</p>", unsafe_allow_html=True)
st.markdown("""Unsloth https://github.com/unslothai check out the [wiki](https://github.com/unslothai/unsloth/wiki)""")
st.markdown("""Georgi Gerganov's ggml https://github.com/ggerganov/ggml""")
st.markdown("""Meta's Llama https://github.com/meta-llama""")
st.markdown("""Mistral AI  - https://github.com/mistralai""")
st.markdown("""Zhang Peiyuan's TinyLlama https://github.com/jzhang38/TinyLlama""")
st.markdown("""Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois,
    Xuechen Li, Carlos Guestrin, Percy Liang, Tatsunori B. Hashimoto
    - [Alpaca: A Strong, Replicable Instruction-Following Model](https://crfm.stanford.edu/2023/03/13/alpaca.html)""")

# main ui prompt
# - text box
# - submit
with st.form(key="my_form"):
    customer_msg = st.text_area("Customer Message")
    write_user_chat_message(user_chat, customer_msg)
    if st.form_submit_button("Submit"):
        st.session_state["user_msg_as_prompt"] = customer_msg
        write_user_chat_message(user_chat, customer_msg)
        model, tokenizer = get_model_tokenizer(sota_model_name)
        input_text = alpaca_input_text_format.format(customer_msg)
        st.write(f"```\n{input_text}```")
        assistant_chat.write_stream(
            stream_generation(
                input_text,
                show_prompt=False,
                tokenizer=tokenizer,
                model=model,
            )
        )

if True:
    gpu_stats = torch.cuda.get_device_properties(0)
    max_memory = gpu_stats.total_memory / 1024 ** 3
    start_gpu_memory = torch.cuda.memory_reserved(0) / 1024 ** 3
    st.write(f"GPU = {gpu_stats.name}. Max memory = {max_memory} GB.")
    st.write(f"{start_gpu_memory} GB of memory reserved.")

st.write("Packages:")
st.write(f"pytorch: {torch.__version__}")

st.markdown("""
## ReplyCaddy

Small business owners, app developers, and freelance talent professionals need more time and money to compete for social media presence. Social media apps provide excellent reach on X, FaceBook, and Instagram, so customers sometimes request initial support from those. The correct response has always been to redirect them to the proper support funnels, such as account validation, issue classification, and prioritization.

However, composing the right tone for the first response message has been challenging and time-consuming, especially when the business owner has yet to experience customer support in public. One has to match the right tone and clarity and instruct the customer to redirect to the proper support channels.

By providing diverse responses, we can ensure each support message feels unique and tailored to the customer's message, avoiding the impersonal feel of a canned response.

## Problems with existing solutions

We prompted ChatGPT to respond an irate, emotionally charged, and informal tone to match generation criteria which we will provide below.

"You are a customer support representative. compose a customer response to this tweet:
Your website is straight up garbage. how do you sell high end technology but you cant get a website right?"

[GenerationCriteria]

"Thank you for reaching out and sharing your feedback. We apologize for the trouble you're experiencing with our website. Please DM us or contact our support team at [support link] so we can help resolve any issues you're encountering."

First, the response is wordily composed and does not contain a request to DM; second, it has an apology and its context. Third, from our experience, ChatGPT values politeness more than a human would.

As a business, we want our brands to hold a high standard for these responses.

We present ReplyCaddy, an AI-powered first-response text message generator that will help us compose the right first-response message that composes personal messages and matches the customer's tone.

We tested the prompt above to ReplyCaddy, and it generated these examples:

"hi! let's talk about it."

"we'd love to help. we're here to help!"

"we understand that you are not happy with the website. please send us an email at <|url|>"
""")