Personalized recommendations used in many applications and websites are generated using techniques such as collaborative filtering, content-based filtering, reinforcement learning, etc. These are task-specific approaches. Large language models (LLMs) can generate predictions based on priming with specific input without the need for task-specific model tuning. However, LLMs have not been applied for making personalized recommendations because their maximum input size is smaller than the typical size of user histories used to personalize recommendations. This disclosure describes techniques to obtain personalized recommendations via LLMs by automatically augmenting a user command or query with relevant text phrases about the user. The set of relevant phrases that fit within the input limits of the LLM are extracted from a collection of phrases obtained from relevant historical and contextual information sources based on the embeddings generated based on the user command or query. Implementation of the techniques can improve the relevance and utility of personalized recommendations and can lead to increased user engagement with the recommended content.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.