US 12,373,506 B1
Personalized retrieval-augmented generation system
Anthony Penta, Bellevue, WA (US); Ashok Pancily Poothiyot, San Francisco, CA (US); Geoff Hulten, Lynnwood, WA (US); Ameya Bhatawdekar, Issaquah, WA (US); Tim Gasser, Austin, TX (US); Sateesh Srinivasan, Redwood City, CA (US); and Vasanth Krishna Namasivayam, Danville, CA (US)
Assigned to Dropbox, Inc., San Francisco, CA (US)
Filed by Dropbox, Inc., San Francisco, CA (US)
Filed on Jun. 14, 2024, as Appl. No. 18/744,393.
Claims priority of provisional application 63/624,191, filed on Jan. 23, 2024.
Int. Cl. G06F 16/9535 (2019.01); G06F 16/907 (2019.01)
CPC G06F 16/9535 (2019.01) [G06F 16/907 (2019.01)] 20 Claims
OG exemplary drawing
 
1. A computer-implemented method comprising:
generating, from a query generated by an entity account associated with a content management system hosting a plurality of personalized retrieval-augmented generation systems (RAGs) personalized in part by data contexts of content items stored in the content management system for respective entity accounts, a query embedding comprising a vector representation of the query by utilizing an embedding model that is part of a personalized RAG specific to the entity account from among the plurality of personalized RAGs;
determining a plurality of vectorized segments according to a size of a context window of the personalized RAG specific to the entity account;
determining, utilizing the personalized RAG to perform a retrieval augmentation specific to the entity account by accessing a database storing the plurality of vectorized segments of content items indicated by the size of the context window associated with the entity account, a data context specific to the entity account by comparing the query embedding and the plurality of vectorized segments specific to the entity account;
generating a response from the retrieval augmentation by providing the data context specific to the entity account and the query to a large language model within the personalized RAG, prompting the large language model to generate a personalized response informed by the data context specific to the entity account; and
providing the personalized response specific to the entity account for display on a client device associated with the entity account.