Confidence Tier Routing

import { createIntentForm } from '@intentform/core'
import { openaiProvider } from '@intentform/provider-openai'
import { anthropicProvider } from '@intentform/provider-anthropic'

const engine = createIntentForm({
  tiers: [
    {
      id: 'fast',
      provider: openaiProvider({ model: 'gpt-4o-mini' }),
      threshold: 0.72,
    },
    {
      id: 'smart',
      provider: openaiProvider({ model: 'gpt-4o' }),
      threshold: 0.88,
    },
    {
      id: 'premium',
      provider: anthropicProvider({ model: 'claude-opus-4-7' }),
      // no threshold = final fallback
    },
  ],
  models: [myModel],
})

How it works

Each tier has a threshold. When the AI returns a confidence score below the threshold, the engine discards the result and retries with the next tier’s provider.

intent → fast tier (gpt-4o-mini)
  confidence = 0.65 → below 0.72 → escalate

intent → smart tier (gpt-4o)
  confidence = 0.91 → above 0.88 → accept

Escalation also triggers on:

Parse failure (Zod validation error on AI output)
Missing required fields in the AI response
Rule conflicts that cannot be resolved

Single provider (no tiers)

If you pass provider instead of tiers, all requests use that single provider:

const engine = createIntentForm({
  provider: openaiProvider({ model: 'gpt-4o' }),
  models: [myModel],
})

Cost optimization

Start cheap. The fast tier handles most well-formed intents. Only ambiguous or complex inputs escalate. Typical distribution: ~80% resolved at fast tier.

tiers: [
  { id: 'fast', provider: ollamaProvider({ model: 'llama3.2' }), threshold: 0.70 },
  { id: 'smart', provider: openaiProvider({ model: 'gpt-4o-mini' }), threshold: 0.88 },
  { id: 'premium', provider: openaiProvider({ model: 'gpt-4o' }) },
]