Classify, validate and filter

Zai provides methods for classifying, validating, and organizing data using an LLM.

Check a condition

zai.check() asks the LLM a yes/no question about some input and returns true or false:

import { adk } from '@botpress/runtime'

const emailBody = 'Click here for a FREE iPhone! Limited time only!!!'
const isSpam = await adk.zai.check(emailBody, 'is this spam?')
// true

Call .result() for the Boolean plus the LLM’s reasoning:

const { output } = await adk.zai.check(emailBody, 'is this spam?').result()
// output.value       → true
// output.explanation → "The message uses urgency tactics and promises free rewards..."

Label content

zai.label() applies multiple Boolean labels to content:

const result = await adk.zai.label(emailBody, {
  spam: 'is this email spam?',
  urgent: 'does this require immediate attention?',
  promotional: 'is this promotional content?',
})
// { spam: true, urgent: false, promotional: true }

Filter an array

zai.filter() keeps items that match a condition:

const comments = ['Great product, I love it!', 'This is terrible spam content', 'Very helpful review, thank you']

const clean = await adk.zai.filter(comments, 'is not spam or inappropriate')
// ["Great product, I love it!", "Very helpful review, thank you"]

Sort items

zai.sort() orders items using natural language criteria:

const tasks = [
  'Update documentation',
  'Fix critical security bug',
  'Add new feature',
  'System is down - all users affected',
]

const prioritized = await adk.zai.sort(tasks, 'by urgency and impact, most urgent first')
// ["System is down...", "Fix critical security bug", "Add new feature", "Update documentation"]

Rate items

zai.rate() scores items on a 1-5 scale. Pass a string for a single criterion:

const reviews = [
  'Amazing product! Best purchase ever!',
  "It's okay, nothing special",
  'Terrible quality, broke immediately',
]

const ratings = await adk.zai.rate(reviews, 'Rate the sentiment')
// [5, 3, 1]

Pass an object to score each item across multiple criteria. Each result includes a total:

const essays = ['... first essay ...', '... second essay ...']

const ratings = await adk.zai.rate(essays, {
  grammar: 'Rate the grammar and spelling',
  clarity: 'Rate how clear and well-organized the writing is',
  argumentation: 'Rate the strength of arguments and evidence',
})
// [
//   { grammar: 4, clarity: 5, argumentation: 3, total: 12 },
//   { grammar: 3, clarity: 4, argumentation: 4, total: 11 },
// ]

Group items

zai.group() categorizes items into groups. With just instructions, it discovers groups on its own:

const messages = [
  "I can't log in to my account",
  'How do I reset my password?',
  'When will my order arrive?',
  'The app keeps crashing',
]

const groups = await adk.zai.group(messages, {
  instructions: 'Group by type of customer issue',
})
// { "Login Issues": [...], "Shipping Questions": [...], "Technical Errors": [...] }

To force items into predefined categories, pass initialGroups:

const articles = ['How to build a React app', 'Python machine learning tutorial', 'Understanding Docker containers']

const groups = await adk.zai.group(articles, {
  instructions: 'Categorize by technology',
  initialGroups: [
    { id: 'frontend', label: 'Frontend Development' },
    { id: 'backend', label: 'Backend Development' },
    { id: 'ml', label: 'Machine Learning' },
  ],
})

Option	Type	Description
instructions	string	How to group the items
initialGroups	Array<{id, label}>	Predefined categories to use instead of discovering them
maxGroups	number	Upper limit on groups. Smaller groups merge at the end until within the limit
minElements	number	Minimum items per group. Groups below the threshold have their elements redistributed
tokensPerElement	number	Max tokens per item
chunkLength	number	Max tokens per chunk for large inputs