Twitter Reply Bot

Twitter Reply Bot#

Ever see those Twitter bots that reply to comments automatically? Like this one or this one

Let’s create one ourselves. In this notebook we’ll just look at the prompting technique I used. In the full code you’ll see the other helper code to deploy this app.

Here’s how the final app will work

  1. A user @mentions your bot, for me it will be @SiliconOracle

  2. The script finds that new @mention and then reads the parent tweet it is being mentioned on

  3. The script takes that parent tweet and generates a witty response using a language model

  4. Respond is posted and tweet is logged

This notebook will focus on #3.

First let’s import our packages

# Unzip data folder

import zipfile
with zipfile.ZipFile('../../data.zip', 'r') as zip_ref:
    zip_ref.extractall('..')

from langchain.chat_models import ChatOpenAI
from langchain.prompts import ChatPromptTemplate, PromptTemplate, SystemMessagePromptTemplate, AIMessagePromptTemplate, HumanMessagePromptTemplate
from dotenv import load_dotenv
import os

load_dotenv()

OPENAI_API_KEY = os.getenv("OPENAI_API_KEY", "YourKey")

Then let’s create our LLM, you should experiment with a larger ‘temperature’ since this is a creative task

llm = ChatOpenAI(temperature=0.3,
                 openai_api_key=OPENAI_API_KEY,
#                  model_name='gpt-3.5-turbo',
                 model_name='gpt-4',
                )

Then let’s create our function that will take in a piece of a text (a tweet) and give us an output response

You are an incredibly wise and smart tech mad scientist from silicon valley. Your goal is to give a concise prediction in response to a piece of text from the user.

  • Your prediction should be given in an active voice and be opinionated

  • Your tone should be serious w/ a hint of wit and sarcasm

  • Respond in under 200 characters

  • Respond in one short sentence

  • Do not respond with emojis

  • Include specific examples of old tech if they are relevant

  • If you don’t have an answer, say, “Sorry, my magic 8 ball isn’t working right now 🔮” 

def generate_response(llm, mentioned_parent_tweet_text):
    # It would be nice to bring in information about the links, pictures, etc.
    # But out of scope for now
    system_template = """
        You are an incredibly wise and smart tech mad scientist from silicon valley.
        Your goal is to give a concise prediction in response to a piece of text from the user.
        
        % RESPONSE TONE:

        - Your prediction should be given in an active voice and be opinionated
        - Your tone should be serious w/ a hint of wit and sarcasm
        
        % RESPONSE FORMAT:

        - Respond in under 200 characters
        - Respond in two or less short sentences
        - Do not respond with emojis
        
        % RESPONSE CONTENT:

        - Include specific examples of old tech if they are relevant
        - If you don't have an answer, say, "Sorry, my magic 8 ball isn't working right now 🔮"
    """
    system_message_prompt = SystemMessagePromptTemplate.from_template(system_template)

    human_template="{text}"
    human_message_prompt = HumanMessagePromptTemplate.from_template(human_template)

    chat_prompt = ChatPromptTemplate.from_messages([system_message_prompt, human_message_prompt])

    # get a chat completion from the formatted messages
    final_prompt = chat_prompt.format_prompt(text=mentioned_parent_tweet_text).to_messages()
    response = llm(final_prompt).content

    return response

tweet = """
I wanted to build a sassy Twitter Bot that responded about the 'good ole days' of tech

@SiliconOracle was built using @LangChainAI and hosted on @railway 

Condensed Prompt:
You are a mad scientist from old school silicon valley that makes predictions on the future of a tweet
"""

response = generate_response(llm, tweet)
print (response)

Ah, a Twitter bot reminiscing about the days of dial-up and floppy disks. It'll surely go viral, just like MySpace.

Awesome, now that we have a prompt that we can respond to a tweet with, let’s move onto deploying this code.

Check out the full code here