Okta, Inc.

Townsend Street Labs, Inc.

Ankit Goyal

Pratyus Patnaik

Rajhans Samdani

Pierre Rappolt

Nathaniel Ackerman Rook

Suchit Agarwal

Emily Wang

William Stone Potter

Roopak Venkatakrishnan

A system creates and searches knowledge base (KB) articles inside an organization while supporting operation as a SaaS (Software as-a-Service) across a plurality of organizations. The system implements a real-time online learning-to-rank (L2R) algorithm for learning relevance scoring that is customized to each organization. This algorithm incorporates rich lexical features using a query similarity kernel. A scoring function includes a pairwise static module, which may be trained off-line using training data, and a lexical adaptive module, which is trained on-line based on user feedback. The scoring function makes the system easy to deploy, modify, and suitable for handling events that naturally happen over the lifecycle of any KB deployment, without manual training.

System for handling workplace queries using online learning to rank

A system stores resources such as text articles, videos, and so forth for an organization. During operation, the system receives a query and provides a response. During initial use of the system, there is little or no historical data available to help determine which resource is most relevant to a particular query. In this 'cold-start' situation, the system determines attributes associated with a user account of the user making the query. The query is used to search a data store and retrieve a set of resources based on a term match with the query and to find the resources which correspond to the attributes of the user account. This allows the system to provide simplified output that is more likely to be relevant to that particular user in the 'cold-start' situation.

Personalized search based on account attributes

A system dynamically routes queries from an organization for processing. An incoming query is classified using a base model to determine a canonical group. The query is routed based on the canonical group. Responsive to the query, response data is then provided to the user device. Event data about the query, the routing, and the response data is used to train a first model that is used to determine subsequent routing. When a threshold is reached with respect to the available event data, a second model is trained using the event data. Processing of subsequent queries using the second model results in reduced demand for computer resources and reduces response times. The second model may be retrained on a schedule, or upon occurrence of a trigger event, such as an addition or removal of a canonical group.

Dynamic query routing system

Growing community of inventors

Daly City, CA, United States of America

Ankit Goyal