Perplexity、モデルに検索コードを自ら書かせることでトークンコストを85%削減
Perplexityは「search as code」を導入しました。これは、固定APIを呼び出す代わりにモデルがカスタムのPython検索パイプラインを書くアーキテクチャです。これにより、複雑なリサーチエージェントにおいてトークン使用量が85%削減され、データ精度の向上が期待されます。
perplexity detailed its new search as code architecture, which replaces standard api calls with model-generated python scripts that execute in a secure sandbox. by using an agentic search sdk, models can perform custom retrieval, filtering, and reranking to keep context windows lean and improve result accuracy. the company demonstrated this on a cybersecurity task tracking 200 vulnerabilities, where the agent used 85% fewer tokens and outperformed standard pipelines. this approach challenges the dominant fixed api paradigm, suggesting that giving models programmatic control over search logic may be more efficient than universal, black-box tools.