Researcher reverse-engineers Apple's personal knowledge graph: 181 entity classes, 278 predicates, and 144-dimensional kinship inference features
The researcher conducted a reverse analysis of the personal knowledge graph built by the two daemons knowledgeconstructiond and intelligenceplatformd in macOS 26, finding that its data is stored in 51 SQLite databases under ~/Library/IntelligencePlatform, with the core graph.db using a triple structure of 181 entity classes and 278 predicates. Kinship relationships are inferred by two XGBoost classifiers based on 144 features, of which 84 features come from co-occurring photos; the results are written to a separate entityTagging.db and recalculated daily. The author also open-sourced a local viewing tool, iKnow, and pointed out that deleted records enter the expired_stable_graph table, so conclusions may outlast the evidence.
Why read this
AI score 66
The author reverse-engineered the 181 entity classes and 278 predicates of the local knowledge graph on macOS and open-sourced a viewing tool, which can be used to understand how on-device personalized data is actually organized.