lib/src/limits.js

"use strict";
Object.defineProperty(exports, "__esModule", { value: true });
exports.TokenLimits = void 0;
class TokenLimits {
    maxTokens;
    requestTokens;
    responseTokens;
    knowledgeCutOff;
    constructor(model = 'openai/gpt-oss-120b:free') {
        this.knowledgeCutOff = '2023-12-01';
        // Default for modern large context models
        this.maxTokens = 128000;
        this.responseTokens = 4000; // Standardized at 4k for safety across all models
        if (model.includes('kimi-k2')) {
            this.maxTokens = 256000;
        }
        else if (model.includes('llama-3.3')) {
            // Models with strict TPM limits
            // We must cap the context to avoid 'Request too large' 413 errors
            this.maxTokens = 12000;
            this.responseTokens = 3000; // Leave more room for input
        }
        else if (model.includes('qwen') || model.includes('llama-4')) {
            this.maxTokens = 128000;
        }
        else if (model.includes('gpt-oss-120b')) {
            this.maxTokens = 128000;
        }
        else if (model.includes('3.2-11b')) {
            this.maxTokens = 128000;
        }
        else if (model.includes('llama3') ||
            model.includes('llama-3') ||
            model.includes('mixtral')) {
            this.maxTokens = 128000;
        }
        // Provide some margin
        this.requestTokens = this.maxTokens - this.responseTokens - 500;
        // Ensure requestTokens is never negative
        if (this.requestTokens < 1000) {
            this.requestTokens = 1000;
        }
    }
    string() {
        return `max_tokens=${this.maxTokens}, request_tokens=${this.requestTokens}, response_tokens=${this.responseTokens}`;
    }
}
exports.TokenLimits = TokenLimits;
//# sourceMappingURL=limits.js.map