Perplexity's Photon Search Engine for AI Agents

Perplexity has introduced its own search engine, Photon, designed to serve AI agents. The new system replaced an open-source engine that could not handle the growing load.

Development was carried out by a small team of engineers alongside a swarm of coding agents. The infrastructure became more efficient: servers store 2.5 times more data per document while using 20% fewer machines.

System performance has increased significantly. Latency within the engine dropped from 800 to 65 milliseconds, ensuring high-speed query processing.

Developers now have access to a faster Search API mode. It offers a median latency of 160 ms and processes 95% of queries within 230 ms.

• Cost reduced to $1 per 1,000 queries (previously $5)
• Spending on models and search decreased by 68%
• Result quality remains stable: 64.3% success rate versus 64.0% for the old version

There are optimization trade-offs. Relevance of results has slightly decreased, and direct answers occur less frequently—56.7% of cases compared to the previous 59.6%.