The user’s personal data is turned into a highly contextualized timeline of activity they did during the day. This includes turning location traces into places visited, parsing chat messages to extract people and places, mining emails for hotel and restaurant reservations, and much more! This runs fully on-device, with no user data being sent to our servers, ensuring privacy by design.
Personal Knowledge Graph
The user’s activity timeline and all other contextual data are linked to create his Personal Knowledge Graph.
Similarly to how the human brain does, our assistants use this as their memory, enabling contextual disambiguation in natural language interactions. This also runs fully on-device, keeping the user’s data safe.
Intent & Entity Recognition
Understanding Natural Language requires two key technologies: detecting the intention of the user, and extracting entities – “things” – they are talking about.
Our algorithms leverage both classical NLP methods and Deep Neural Networks, offering a good tradeoff between precision and performance.
And since we also wanted to do that on device, we optimized our models so that they could fit on a smartphone with no impact on battery!
Deep Natural Language Queries
Deep Natural Language Queries (which involves solving multiple sub-queries) are binded to the user’s Personal Knowledge Graph, enabling arbitrarily complex queries that require contextual understanding. Thanks to our technology, assistants can now answer queries like “show me pictures of the food I ate at that french place I went to for dinner last Thursday in New York”. And yes, this ALSO runs 100% on device, making Snips the first end-to-end private AI assistant 🙂