Details
- Perplexity introduced Fast Search, a new lower-latency mode for its Search API aimed at day-to-day agentic tasks.
- The service runs on Photon, a Rust-based retrieval and ranking system developed by a small engineering team with support from hundreds of agents.
- Fast Search returns 95% of results within 230 milliseconds; its reported median latency is 160 milliseconds.
- Across six agent benchmarks, Fast Search reduced cost per task by 68% versus Perplexity’s default preset while delivering comparable quality.
- Internal testing found a 0.24-point relevance decline and roughly a three-percentage-point reduction in answer availability on long-tail and broad-coverage queries.
- Photon has replaced Perplexity’s previous open-source retrieval engine across the company’s search systems.
- Its compact index formats reduce data read per query, while asynchronous batched disk reads help prevent storage access from delaying responses.
- Photon separates index building from live serving and rolls out updates to one serving group at a time, warming caches and checking readiness before routing traffic.
- Perplexity reports that Photon reduced p99 response time from about 800 milliseconds to 65 milliseconds, uses about 20% fewer serving machines, and stores 2.5 times more data per document.
Impact
Fast Search targets the main trade-off in AI search infrastructure: lower latency and cost at a modest reduction in recall and answer availability. Its reported 160-millisecond median places it in the competitive range of specialized agent-search APIs, although independent benchmarks vary by workload and configuration. The larger significance is operational: Photon gives Perplexity more control over indexing, capacity, and rollout speed while reducing infrastructure requirements, potentially supporting broader developer adoption of real-time search in agent workflows.
