Loading patent…
AssigneeDataRobot, Inc.
InventorsStefan Hackmann, Alexander Jason Conway, Mikalai Valynets, Michael Schmidt, Debadeepta Dey, Mark L. Steadman
Estimating the term…
Constructing AI flows involves navigating a large number of choices of components including vector databases, embedding models, text splitters, retrievers, and synthesizing LLMs. The complexity further escalates with the introduction of advanced agentic and non-agentic paradigms that involve additional components including verifiers, rewriters, and re-rankers, each with intricate hyperparameter options. These choices pose significant challenges in optimizing for tradeoffs between objectives such as latency, accuracy, and cost. To address these challenges, the systems and methods of the present disclosure conduct an efficient multi-objective search across an extensive space of agentic and non-agentic AI flows (e.g., RAG-based AI flows). Leveraging advances in Bayesian Optimization, the multi-objective search can, in some examples, identify high-performing (e.g., Pareto-optimal) flows that by definition cannot be improved upon in selected objectives (e.g., task accuracy and/or latency) without sacrificing on one of the objectives.
A computer-implemented method, comprising:
A computing device, comprising:
A non-transitory computer-readable medium storing instructions that, when executed by a processor, causes the processor to perform operations comprising:
Dependencies are inferred from the claim text.