src/limits.ts

export class TokenLimits {
  maxTokens: number
  requestTokens: number
  responseTokens: number
  knowledgeCutOff: string

  constructor(model = 'llama-3.1-8b-instant') {
    this.knowledgeCutOff = '2023-12-01'

    // Adjust for Groq TPM/RPM limits on Free/On-Demand tier
    if (model.includes('llama-4')) {
      this.maxTokens = 29500
      this.responseTokens = 4000
    } else if (
      model.includes('llama-3.1-8b') ||
      model.includes('llama-3.2-1') ||
      model.includes('llama-3.1-70b') ||
      model.includes('llama-3.3-70b')
    ) {
      // Groq limits are ~6k TPM for 8b and ~12k TPM for 70b
      this.maxTokens = model.includes('8b') ? 5500 : 11500
      this.responseTokens = 2000 // Reserve less for response to allow more context
    } else {
      this.maxTokens = 128000
      this.responseTokens = 16000
    }

    // Provide some margin
    this.requestTokens = this.maxTokens - this.responseTokens - 500
  }

  string(): string {
    return `max_tokens=${this.maxTokens}, request_tokens=${this.requestTokens}, response_tokens=${this.responseTokens}`
  }
}