| Platforms | iOS 18 and later |
| Engine models | About 375 MB, downloaded once while indexing starts |
| Sources | Photos and screenshots. Videos are not indexed yet. |
| Facets | Up to 64, with up to 32 sample photos each |
| Results per page | Up to 100 |
| Cloud assist | 8 captures per naming request, 30 requests a minute per install. Indexing uses it only with namesWithCloudAssist on a device without Apple’s model, so a device with Apple Intelligence spends no units on indexing. |
| Rules | Up to 500 characters. Cloud assist judges at most 48 captures per call, 24 per request. |
| Ranking | Up to 500 candidates per call. Taste vectors: 1 to 12 centres, none below 20 saved captures. |
| Scores | Classify probabilities are fitted to the cloud judge on a small test library; values between 0.2 and 0.8 are approximate. Rank scores are not fitted. |
| Trips | Cities of 15,000 people or more. None until 30 photos on 10 days carry a location. Up to 20 photos per trip. |
| Places | Nothing until home is known. A visit: within 90 minutes and 400 m. Venues within the visit’s radius plus 50 m. At most one place lookup every 1.5 s and 60 per run, never on a metered network. |
| Agent MCP server | 5 agent tokens per account. 60 calls at once and 300 an hour. |
| Answers | Questions up to 1,000 characters, read from up to 40 things and excerpts |
| Profiles | Schemas up to 32 KB, 100 fields and 5 levels, read from up to 80 things and excerpts |
| Apps per account | 10 |
| Active keys per app | 10 |
| Bundle IDs per app | 5 |
