ChatGPT, Forced Moderator For Do_Not_Reply Inbox

LLMs can be introduced into a web application in a variety of ways, and before I began implementing it as a core solution for any large problem, I decided first that it would be entertaining to integrate it in an entirely useless way, as a moderator for a Do_Not_Reply Inbox.

I was in the process of developing my web application for gravel-focused cycling route creation, https://sherpa-map.com, and I discovered the need for a login, something I had never programmed before. While adding all of the typical, useful, items expected in a sign-up, and heavily leveraging ChatGPT in the development process, it’s so useful I didn’t even bother installing Postman, I was struck with a hilarious idea, why not integrate ChatGPT API as a service that checks the inbox for the do_not_reply@sherpa-map.com and replies in a snarky, sarcastic manner?

In order to implement this in a Node JS backend that interfaces with Microsoft Graph to send/reply/receive emails a few key pieces were necessary:

1. A function that listens and sends replies.

// Function to listen for replies to the "do_not_reply" mailbox
async function listenForReplies() {
  const client = graph.Client.init({
    defaultVersion: 'v1.0',
    authProvider: (done) => {
      credential.getToken('https://graph.microsoft.com/.default')
        .then((tokenResponse) => {
          const accessToken = tokenResponse.token;
          done(null, accessToken);
        })
        .catch((error) => {
          done(error, null);
        });
    },
  });
  try {
    // Fetch unread messages in the "do_not_reply" mailbox
    const messages = await client.api(`/users/${userID}/messages`).filter("isRead eq false").get();
    if (messages.value && messages.value.length > 0) {
      for (const message of messages.value) {
        // Send a snarky reply to the sender
        await sendSnarkyReply(client, message, message.from.emailAddress.address);

        // Mark the message as read
        await client.api(`/users/${userID}/messages/${message.id}`).patch({ isRead: true });
      }
    }
  } catch (error) {
    console.error('Error listening for replies:', error);
  }

2. A main function with a never-ending loop that checks the inbox every minute by activating the “listenForReplies()” function.

async function main() {

  while (true) {
    await listenForReplies();
    // Check for new emails every 60 seconds
    await new Promise(resolve => setTimeout(resolve, 60 * 1000));
  }
}

main();

3. The “sendSnarkyReply()” function which formats and sends the response to the user.

// Helper function to send a snarky reply
async function sendSnarkyReply(client, originalMessage, recipientEmail) {
  const snarkyResponse = await generateSnarkyResponse(originalMessage);

  const message = {
    subject: 'RE: ' + originalMessage.subject,
    body: {
      contentType: 'HTML',
      content: snarkyResponse,
    },
    toRecipients: [
      {
        emailAddress: {
          address: recipientEmail,
        },
      },
    ],
  };

  try {
    await client.api(`/users/${userID}/sendMail`).post({ message });
    console.log('Snarky email sent successfully!');
  } catch (error) {
    console.error('Error sending snarky email:', error);
  }
}

4. Lastly, the good stuff, the “snarkyResponse()” function which interprets the unread email from the loop process, sends the request to OpenAI, and returns the response.

async function generateSnarkyResponse(message) {

  const apiUrl = 'https://api.openai.com/v1/chat/completions';
  
const emailBody = message.body.content; // or message.body.content, depending on what you want to use

 
 try {
   const response = await axios.post(apiUrl, {
  model: 'gpt-3.5-turbo', // Specify the model to use
  messages: [
    {
   "role": "system",
   "content": "You are a witty, sarcastic, and snarky AI trained to humorously reply to emails sent to a do_not_reply inbox. Use your humor to create a playful response while reminding the sender that they shouldn't have emailed this inbox in the first place."
    },
    {
   "role": "user",
   "content": emailBody
    }
  ],

  max_tokens: 250, // Adjust the desired length of the generated response
  temperature: 0.7, // Adjust the level of creativity in the generated response
  // n: 3, // Number of responses to generate for each prompt
   }, {
  headers: {
    'Content-Type': 'application/json',
    'Authorization': `Bearer ${apiKey}`,
  },
   });


    const { choices } = response.data;
 
    const reply = response.data.choices[0].message.content;
 

 // Add the disclaimer text with <br> tags for line breaks and <strong> tags for bold text
 const disclaimer = '<br><br><strong>Disclaimer: This is an OpenAI bot, instructed to act a bit snarky as people don\'t normally email a do-not-reply inbox. This reply was generated by an AI, not a human being.</strong>';
 // Add the bot's signature
 const signature = '<br><br>--<br><strong>Snarky Bot</strong><br>Do-Not-Reply Representative<br>Sherpa LLC<br><a href="https://www.sherpa-map.com">www.sherpa-map.com</a>';

 const replyWithDisclaimerAndSignature = reply + disclaimer + signature;

  
   return replyWithDisclaimerAndSignature;
  } catch (error) {
    console.error('Error generating snarky response:', error);
 console.log('Error response data:', error.response.data); // Add this line
    return 'Oops, something went wrong!';
  }
}

There may be a lot going on here, but it basically boils down to forwarding the email contents to OpenAI, utilizing the role prompt “You are a witty, sarcastic, and snarky AI trained to humorously reply to emails sent to a do_not_reply inbox. Use your humor to create a playful response while reminding the sender that they shouldn’t have emailed this inbox in the first place.” Appending a few additional details, such as a disclaimer and its own signature.

Here’s an example of a test verification email I am replying to:

Here is an example of a response:

What tops off the entertainment value of this is the fact it’s not limited to a one-time response, each reply to an email sends the entire email string, so it’s entirely possible to have a conversation!