@djangocfg/llm

LLM client with SDKRouter, OpenAI, and Anthropic support. Includes JSON translator with validation.

Installation

pnpm add @djangocfg/llm

Quick Start

import { createLLMClient, Model, createTranslator } from '@djangocfg/llm'

// Auto-detect from env (SDKRouter preferred)
const llm = createLLMClient()

// With model alias
const llm = createLLMClient({ model: Model.balanced({ code: true }) })

// Simple chat
const response = await llm.chat('Hello!')

// JSON response
const data = await llm.json<{ name: string }>('Return a user object')

// Translate JSON
const translator = createTranslator(llm)
const result = await translator.translate({ title: 'Hello' }, 'ru')
console.log(result.data) // { title: 'Привет' }

Environment Variables

# SDKRouter (recommended)
SDKROUTER_API_KEY=sk-...

# Or OpenAI
OPENAI_API_KEY=sk-...

# Or Anthropic
ANTHROPIC_API_KEY=sk-ant-...

Priority: SDKROUTER_API_KEY > OPENAI_API_KEY > ANTHROPIC_API_KEY

Model Aliases (SDKRouter)

SDKRouter uses smart model aliases to select the best model for your task:

import { Model, ModelPresets } from '@djangocfg/llm'

// Tier methods
Model.cheap()                    // '@cheap' - Cheapest
Model.balanced()                 // '@balanced' - Best value (default)
Model.smart()                    // '@smart' - Highest quality
Model.fast()                     // '@fast' - Lowest latency

// With capabilities
Model.cheap({ vision: true })    // '@cheap+vision'
Model.balanced({ tools: true })  // '@balanced+tools'
Model.smart({ json: true })      // '@smart+json'

// With categories
Model.balanced({ code: true })   // '@balanced+code'
Model.smart({ reasoning: true }) // '@smart+reasoning'

// Pre-built presets
ModelPresets.translation   // '@cheap+json' - For translation tasks
ModelPresets.code          // '@balanced+code' - For code generation
ModelPresets.reasoning     // '@smart+reasoning' - For complex reasoning

Available Tiers

• cheap - Cheapest available model
• budget - Budget-friendly with decent quality
• standard - Standard tier
• balanced - Best quality/price ratio (recommended)
• smart - Highest quality model
• fast - Lowest latency model
• premium - Top-tier premium model

Available Capabilities

• vision - Image understanding
• tools - Function/tool calling
• agents - Agent tool calling (verified)
• json - JSON mode
• streaming - Streaming support
• long - Long context (128k+)
• image - Image generation

Available Categories

• code - Code generation
• reasoning - Reasoning & math
• creative - Creative writing
• chat - Conversational
• analysis - Analysis & extraction

API

createLLMClient(config?)

const llm = createLLMClient({
  provider: 'sdkrouter',   // 'sdkrouter' | 'openai' | 'anthropic'
  apiKey: 'sk-...',        // Optional, uses env
  model: '@balanced',      // Default model
  temperature: 0.1,
  maxTokens: 4096,
})

llm.chat(prompt, options?)

const response = await llm.chat('Translate to Russian: Hello', {
  temperature: 0,
  system: 'You are a professional translator',
})

console.log(response.content)  // 'Привет'
console.log(response.usage)    // { promptTokens, completionTokens, totalTokens }

llm.json(prompt, options?)

interface User {
  name: string
  email: string
}

const user = await llm.json<User>('Return a sample user object')

createTranslator(llm)

const translator = createTranslator(llm)

const result = await translator.translate(
  { title: 'Hello', items: ['One', 'Two'] },
  'ru',
  {
    sourceLanguage: 'en',
    maxRetries: 2
  }
)

if (result.valid) {
  console.log(result.data) // { title: 'Привет', items: ['Один', 'Два'] }
} else {
  console.error('Errors:', result.errors)
}

Providers

SDKRouter (Default)

• Base URL: https://llm.sdkrouter.com/v1
• Uses smart model aliases (@cheap, @balanced, @smart)
• Automatically selects best model for capabilities

OpenAI

Models: gpt-4o, gpt-4o-mini, gpt-4-turbo, gpt-3.5-turbo

Anthropic

Models: claude-3-5-sonnet-latest, claude-3-5-haiku-latest, claude-3-opus-latest

License

MIT