A new retrieval and ranking service called Photon now delivers 95% of search results in 230 milliseconds or less for users of the Perplexity Search API. The update introduces a Fast Search option optimized for agentic workloads, which features a median latency of 160 milliseconds. Across six agent benchmarks, the new system reduces the cost per task by 68% compared to the default preset while maintaining comparable quality.

Photon replaces the open source engine previously used by Perplexity for all retrieval and ranking tasks. Built with the Rust language, the system was developed by a small engineering team using $300k in tokens and hundreds of auto research loops. The engine operates on 20% fewer machines and stores 2.5x more data per document, though internal tests on broad queries show it scores 0.24 points lower on relevance and 3 percentage points lower on answer availability.