A smart rule engine for AI usage routing and cost control logic. This TypeScript library allows you to define declarative rules that match request contexts and determine model routing, rate limiting, and cost enforcement policies.
- Declarative Rules: Define conditions and actions in a simple, readable format
- Smart Routing: Route requests to different AI models based on user tier, usage, or custom criteria
- Cost Control: Enforce usage limits by requests, tokens, cost, or concurrent connections
- Flexible Conditions: Support for comparison operators, logical operators, and nested conditions
- Window-based Limits: Track usage across different time windows (minute, hour, day, month, year)
- Graceful Degradation: Downgrade to cheaper models when limits are exceeded
- Type Safety: Full TypeScript support with Zod schema validation
npm install @tokenlay/rules
import { evaluateRules, type Rule, type EvaluationContext } from '@tokenlay/rules';
// Define your rules
const rules: Rule[] = [
{
id: 'free-tier',
if: { tier: 'free' },
then: {
model: 'gpt-3.5-turbo',
limits: [{
type: 'requests',
max: 100,
window: 'day',
per: 'user'
}]
},
priority: 1
},
{
id: 'pro-tier',
if: { tier: 'pro' },
then: {
model: 'gpt-4',
limits: [{
type: 'cost',
max: 50,
window: 'month',
per: 'user'
}],
on_limit_exceeded: {
action: 'downgrade',
downgrade_to: 'gpt-3.5-turbo'
}
},
priority: 2
}
];
// Evaluate against context
const context: EvaluationContext = {
tier: 'pro',
user: 'user123',
usage: {
requests: 50,
tokens: 10000,
cost: 45,
concurrent: 2
},
windows: {}
};
const result = evaluateRules(rules, context);
console.log(result);
// { matched: true, rule_id: 'pro-tier', action: 'allow', model: 'gpt-4', ... }
Support various matching patterns:
// Simple equality
{ tier: 'free' }
// Comparison operators
{ 'usage.cost': { gte: 50 } }
{ requests: { lt: 100 } }
{ feature: { in: ['chat', 'completion'] } }
// Logical operators
{
and: [
{ tier: 'enterprise' },
{ 'usage.cost': { lt: 1000 } }
]
}
{
or: [
{ admin: true },
{ tier: { in: ['pro', 'enterprise'] } }
]
}
// Negation
{ not: { blocked: true } }
gt
,gte
,lt
,lte
: Numeric comparisonseq
,neq
: Equality checksin
,nin
: Array membershipcontains
,startsWith
,endsWith
: String operations
Define what happens when a rule matches:
{
model: 'gpt-4', // Model to use
model_params: { temperature: 0.7 }, // Model parameters
action: 'allow', // allow | block | warn | queue
limits: [{ // Usage limits
type: 'tokens',
max: 100000,
window: 'day',
per: 'user'
}],
on_limit_exceeded: { // What to do when limit hit
action: 'downgrade',
downgrade_to: 'gpt-3.5-turbo',
error_message: 'Daily token limit exceeded'
}
}
requests
: Number of API requeststokens
: Token consumptioncost
: Monetary costconcurrent
: Concurrent connections
minute
,hour
,day
,month
,year
,total
Group usage by custom fields:
{
limits: [{
type: 'cost',
max: 100,
window: 'month',
per: 'project',
tracking_key: 'project_id' // Use context.project_id for grouping
}]
}
Rules are evaluated by priority (higher first), then by order:
const rules: Rule[] = [
{ if: { tier: 'free' }, then: { model: 'gpt-3.5' }, priority: 1 },
{ if: { tier: 'free', vip: true }, then: { model: 'gpt-4' }, priority: 10 }
];
// VIP free users get GPT-4
Track usage across time windows:
const context: EvaluationContext = {
usage: { requests: 50, tokens: 10000, cost: 5, concurrent: 1 },
windows: {
'requests_hour_user_123': {
current: 45,
limit: 50,
resets_at: '2025-07-27T15:00:00Z'
}
},
// ... other fields
};
Main evaluation function that processes rules against context.
interface Rule {
id?: string;
description?: string;
priority?: number;
if: Condition;
then: Action;
}
interface EvaluationContext {
usage: {
requests: number;
tokens: number;
cost: number;
concurrent: number;
};
windows: Record<string, WindowData>;
[key: string]: any; // Custom fields
}
interface EvaluationResult {
matched: boolean;
rule_id?: string;
action: 'allow' | 'block' | 'warn' | 'queue';
model?: string;
limit_exceeded?: boolean;
error_code?: string;
error_message?: string;
// ... other fields
}
# Install dependencies
npm install
# Run tests
npm test
# Build
npm run build
# Type check
npm run typecheck
MIT