Perplexity has introduced Photon, a new retrieval and ranking engine written in Rust. It replaces an open-source engine previously used for its AI-native search stack. Photon now manages retrieval and ranking for all production traffic, including a new Fast Search mode.
Photon is available as a hosted API. Users can set search_type: "fast" on POST /search requests. The cost is $1 per 1,000 requests. It is not open source, so it cannot be self-hosted.
The old engine faced several limits as the dataset grew. Its p99 latency was near 800 ms, often caused by cold reads and page faults. During disk index fusion, p99 latency climbed to about 1.2 seconds. Recovery from updates could take over a week and caused partial responses.
Perplexity built Photon from scratch because it was simpler and cheaper than maintaining its fork. Photon uses a load balancer to route requests to shards. Each shard performs retrieval, ranking, and candidate scoring. The system merges candidates and fetches document fields.
Photon uses several technical improvements. It has adaptive posting lists that split long lists into blocks. It employs a budgeted traversal algorithm to bound candidate scores. It stores documents with compact records using Elias-Fano encoding. Disk reads are batched with io_uring, and cache management avoids locks.
Photon runs on about 20% fewer servers than the old system. It stores 2.5 times more data per document, improving ranking quality. Index version switches no longer cause latency spikes. Latency for p99 retrieval dropped from 800 ms to about 65 ms.
Perplexity tested Fast Search on six benchmarks. It scored 64.3% at $59.73 in estimated total cost. The default preset scored 64.0% at $187.60, making Fast about 68% cheaper. Relevance on internal benchmarks slightly decreased, so keep the default for hard queries.
Developers can try the API with a simple POST request. The system offers faster responses for agent workflows and day-to-day search tasks. The technical details and documentation are available for further exploration.