Multi-model routing

Beamdesk routes AI requests across 13 production models from four families: Claude, GPT, Gemini, and GLM-based. Each persona can specify preferred model family and fallback chain. Routing balances latency, cost, and capability per request.

Default routing uses a tiered fallback chain: Haiku for quick responses, Sonnet for standard requests, and Opus for complex tasks requiring deep reasoning. Personas can override this per-channel or per-intent.

Persona model configuration:
Routing tier: auto  // haiku → sonnet → opus
Fallback family: claude
Timeout ms: 30000
Max output tokens: 8192

Per-channel overrides:
  email: { tier: 'opus', fallback: 'claude' }
  chat_widget: { tier: 'sonnet', fallback: 'gpt' }
  voice: { tier: 'sonnet', fallback: 'claude' }

Per-intent routing:
  'pricing_question': { model: 'gpt-5.5' }
  'technical_debug': { model: 'claude-opus' }
  'refund_request': { model: 'claude-opus', consensus: true }

When a model times out or errors, routing falls back to the next model in the chain. If fallback chain is exhausted, the request fails with a clear error. You can configure per-channel timeouts and max output tokens.

Each persona can specify model family preference, routing tier, and consensus mode. This lets you use fast models for low-risk queries and expensive models for high-risk decisions.

Persona: Billing
System prompt:
  You handle refunds, invoices, and subscription changes.
  Always verify policy before approving exceptions.

Model settings:
  primary_family: claude
  fallback_family: gpt
  tier: opus
  consensus_mode: auto

Consensus triggers:
  - refund_amount >= 50
  - policy_exception
  - customer_risk_level: high

When consensus is enabled, routing runs:
  1. Primary model generates draft
  2. Secondary model verifies draft
  3. If disagreement: third model breaks tie

consensus_mode: auto enables consensus for high-risk decisions detected by guardrails. Manual mode forces consensus for every request. Disabled mode skips consensus entirely.

Beamdesk adds zero markup to model costs. You pay the provider rate directly: Claude, GPT, Gemini, GLM, Kimi, and MiniMax are all billed at their published per-token rates. Usage breakdown is available in workspace billing.

Consensus mode runs multiple models on the same request and compares outputs. Use consensus for high-stakes decisions: refunds, policy exceptions, security questions, and account changes. The tradeoff is higher cost and latency for higher reliability.

Consensus configuration:
  enabled: true
  quorum: 2 // 2/3 models must agree
  timeout_ms: 60000

Model selection for consensus:
  - primary: persona preferred model
  - verifier: different family (avoid echo chamber)
  - tiebreaker: opus or gpt-5.5

Example refund flow:
  1. Draft model: Claude Opus suggests refund
  2. Verifier: GPT-5.5 checks against policy
  3. Tiebreaker: Gemini searches for similar cases
  4. Decision: block (policy violation found)

Result:
  - Action: abstain
  - Citations: [policy_kb_123, case_456]
  - Reason: "Refund exceeds 30-day window without override"

Consensus adds 2-3x cost and latency for consensus-enabled requests. Use it selectively on high-risk queries identified by guardrails, not for every interaction.

import { Beamdesk } from '@beamdesk/sdk';

const beam = new Beamdesk({
  apiKey: process.env.BEAMDESK_API_KEY!,
  baseUrl: 'https://beamdesk.preview.softblaze.net',
});

// Create persona with model preferences
await beam.personas.create({
  name: 'Support',
  systemPrompt: 'You resolve...',
  modelConfig: {
    primaryFamily: 'claude',
    fallbackFamily: 'gpt',
    tier: 'auto', // haiku → sonnet → opus
    consensusMode: 'auto',
    consensusTriggers: [
      { type: 'refund_amount', threshold: 50 },
      { type: 'policy_exception' },
      { type: 'customer_risk', threshold: 'high' },
    ],
    perChannel: {
      email: { tier: 'opus', fallback: 'claude' },
      chat_widget: { tier: 'sonnet', fallback: 'gpt' },
    },
  },
});