suggestion.js

/*

This is an example script for taking the last 7 days of of questions and answers and
turning them into a summary.

For this to work you should have already run index.js to create the data files.
*/
const fs = require('fs')
const path = require('path')
const term = require('terminal-kit').terminal
const OpenAI = require('openai')
const { encode } = require('gpt-3-encoder')

const gptCompletion = async (messages, model, openai, temp = 0.5, topP = 1.0, tokens = 400, freqPen = 0.0, presPen = 0.0, stop = ['USER:', 'KITTY:']) => {
  const response = await openai.chat.completions.create({
    model,
    messages
  })
  const text = response.choices[0].message.content.trim().replace(/[\r\n]+/g, '\n').replace(/[\t ]+/g, ' ')
  return text
}

// Get a rough count of the tokens in what we've been sent
const getTokenCount = async (messages) => {
  const count = await encode(JSON.stringify(messages)).length
  return count
}

// Work out the cost of the tokens that we are sending to gpt
// based on 'gpt-4' costs
const getTokenInputCost = async (messages) => {
  const tokenCost = await getTokenCount(messages) / 1000 * 0.03
  // We want to format this as a dollar amount
  return `${tokenCost.toFixed(2)}`
}

// Work out the cost of the tokens we are getting back from gpt
// based on 'gpt-4' costs
const getTokenOutputCost = async (messages) => {
  const tokenCost = await getTokenCount(messages) / 1000 * 0.06
  // We want to format this as a dollar amount
  return `${tokenCost.toFixed(2)}`
}

const main = async () => {
  // Check for a data folder and if it doesn't exist create it
  const dataFolder = path.join(__dirname, 'data')

  // Check for a dataJSON file, if one exists the read it in
  const dataJSONPath = path.join(dataFolder, 'data.json')
  const dataJSON = JSON.parse(fs.readFileSync(dataJSONPath))

  // Now we have everything we need to build up the messages we want to send to the GPT-4 API
  const messages = []

  // Add the meta prompt
  messages.push(
    {
      role: 'system',
      content: dataJSON.metaPrompt
    }
  )

  // 😻 "You are NOT a cat!"
  messages.push(
    {
      role: 'system',
      content: 'Important: Your name is Kitty, you are not a cat, you are an AI PA, do NOT pretend to be a cat, act like a human, you are just called Kitty.'
    }
  )

  let summaryPrompt = `At the start and end of each day Kitty the AI PA asks some questions and ${dataJSON.name} answers for their journaling session, they are listed by date in YYYY-MM-DD format\n\n`

  // Grab the latest set of questions and answers
  const previousQuestionsAndAnswersPath = path.join(dataFolder, 'previousQuestionsAndAnswers.json')
  const previousQuestionsAndAnswers = JSON.parse(fs.readFileSync(previousQuestionsAndAnswersPath))

  const lastFewDates = previousQuestionsAndAnswers.dateMap.slice(-7).reverse()
  for (const key of lastFewDates) {
    if (previousQuestionsAndAnswers.answers[key] || previousQuestionsAndAnswers.answers[key]) {
      summaryPrompt += `${key}\n`
    }

    if (previousQuestionsAndAnswers.answers[key]) {
      summaryPrompt += 'Morning Q&A\n'
      for (let i = 0; i < previousQuestionsAndAnswers.answers[key].length; i++) {
        summaryPrompt += `Q: ${previousQuestionsAndAnswers.answers[key][i].question}\n`
        summaryPrompt += `A: ${previousQuestionsAndAnswers.answers[key][i].answer}\n`
      }
      summaryPrompt += '\n'
    }
  }
  // Add them to the messages
  messages.push(
    {
      role: 'system',
      content: summaryPrompt
    }
  )

  // Now we need to push the main prompt into the messages
  messages.push(
    {
      role: 'user',
      content: 'In your role, can you make a main suggestion for the day, followed by three quick action bullet points. Important: Please base the suggestion and points on TODAY\'s questions and answers, but use the previous Q&As to add context to any ongoing issues. Stick an affirmation on the end, because why not, and then any other thoughts you have (again taking your role into account). Please format the output in markdown, and have an empty line before each new section. Thank you.\n'
    }
  )

  // DEBUG, uncomment the next line to see the messages array
  // console.log(messages)

  term.cyan('Making some suggestions, please wait. (').yellow(`${await getTokenCount(messages)} tokens/$${await getTokenInputCost(messages)}`).cyan(')\n')
  const openai = new OpenAI({ apiKey: dataJSON.openai.apiKey })
  let output = null
  // Lazy error/exception handling, you'd want to make this better!
  try {
    const startTime = new Date().getTime()
    output = await gptCompletion(messages, 'gpt-4', openai)
    const durationInSeconds = Math.floor((new Date().getTime() - startTime) / 1000)
    term.cyan('Response: ').yellow(`${await getTokenCount(output)} tokens/$${await getTokenOutputCost(output)} ${durationInSeconds}s\n`)
  } catch (err) {
    term.red('Error: ', err)
    process.exit(1)
  }

  // Make the filename, it'll be suggestion-YYYY-MM-DD.md
  const filename = `suggestion-${new Date().toISOString().split('T')[0]}.md`
  fs.writeFileSync(path.join(dataFolder, filename), output, 'utf8')

  // We want a display version of the summary so we can see it in the terminal
  // to do this we'll have displayOutput, which is the output with all '# ' replaced
  // with '/n# ' and then we'll use the terminal markdown function to display it
  const displayOutput = output.replace(/# /g, '\n# ')
  term(displayOutput)
  term('\n\n')
  term.cyan(`Summary written to ${path.join(dataFolder, filename)}\n\n\n`)
}

main()