Abstract: Implementations are described herein for updating user-specific generative model conditioning data at users' requests. In various implementations, a first input prompt is assembled to include user-specific conditioning data (USCD) built from past user interactions and more recent dialog turns between the user and a generative model application. The first input prompt is processed by one or more generative models to generate output identifying out-of-date portions of the USCD. Based on these identified portions, the USCD is updated and stored to reflect the new dialog turns. This updated USCD is then used for subsequent generative model queries.

Abstract: Implementations are described herein for selectively using retrieval augmented generation (RAG) for generative model prompting. In various implementations, a generative model query of a user may be analyzed to determine whether retrieval augmented generation (RAG) should be used to generate a response. If RAG should be used, a generative model input prompt may be formed with data indicative of the generative model query and data indicative of: (i) user-specific conditioning data (USCD) associated with the user, or (ii) personal RAG data of the user. The user-specific conditioning data may have been built over time based at least in part on the personal RAG data of the user. The prompt may be processed using generative model(s) to generate generative model output, conditioned on one or both of the USCD or the personal RAG data of the user, that includes a response to the generative model query.

Abstract: Implementations described herein personalize a user's interactions with generative model(s), including through an automated assistant, by conditioning its queries and responses on user interactions with various computing devices and appliances over time. A first input prompt is assembled, including user-specific conditioning data built over time and new user interactions. This prompt is processed using generative models to identify outdated portions of the conditioning data. A second prompt is then assembled, including the identified outdated portions and the new interactions. This prompt is processed to generate new versions of the outdated portions. Finally, an updated user-specific conditioning data is stored for subsequent use, incorporating the new versions and unaltered portions of the user-specific conditioning data.

Abstract: Implementations described herein relate to updating user-specific conditioning data (USCD) using mappings to user interactions. In various implementations, Data indicative of new user interactions between a user and computing device(s) may be stored. Based on the new user interactions, portions of USCD representing user attributes may be updated. The USCD may be operable to condition a generative model to generate model output tailored to the user. Mappings between the updated portions of the USCD and the data indicative of the new user interactions may also be stored. When it is determined that user interaction(s) have been altered, the mappings may be used to update corresponding portions of the USCD to reflect the alteration of the user interactions.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for using conversation histories to augment a software agent. In particular, the software agent can selectively and adaptively be provided with data from conversation history data that comprises respective segment data for each of one or more past segments of each of a plurality of past conversations between a user and a software agent, the respective segment data comprising a respective segment embedding for the past segment.