US 12,705,269 B1Granted

Computationally efficient language model

AssigneeIntuit Inc.

InventorEilon Sheetrit

Life of this patent

Estimating the term…

Loading patent…

Abstract

A method including receiving a query to a retrieval augmented generation (RAG) language model having electronic accesses to a number of potential contexts. A desired selected context is unknown when the query is received. A number of ranking models is executed on the query. Each of the number of ranking models outputs a corresponding candidate context from among the number of potential contexts. A number of vector data structures are generated by transforming, for each output of the number of ranking models, a combination of the query and the corresponding candidate context into a corresponding vector data structure. A predictor model is executed on the number of vector data structures to generate the selected context from among the number of potential contexts. The RAG language model is enhanced by applying the selected context to the RAG language model to generate an enhanced RAG language model.

Claims19

Dependent claims

A method of increasing a computational efficiency of a computer executing a retrieval augmented generation (RAG) language model having electronic access to a plurality of potential contexts, the method comprising:

receiving a query to the RAG language model having electronic accesses to the plurality of potential contexts, wherein:
the plurality of potential contexts further include a selected context, applicable to the query, selected from among the plurality of potential contexts, and
the selected context is unknown when the query is received;
executing a plurality of ranking models on the query, wherein each of the plurality of ranking models outputs a corresponding candidate context from among the plurality of potential contexts,
wherein each of the plurality of ranking models are different from each other, and
wherein a plurality of candidate contexts are identified;
generating a plurality of vector data structures by transforming, for each output of the plurality of ranking models, a combination of the query and the corresponding candidate context into a corresponding vector data structure,
wherein each of the plurality of vector data structures corresponds to a different output of the plurality of ranking models, and
wherein transforming contextually relates each of the plurality of candidate contexts to the query;
executing a predictor model on the plurality of vector data structures to generate a plurality of probabilities, wherein each of the plurality of probabilities indicates a corresponding likelihood that the corresponding candidate context is the selected context;
selecting, according to the plurality of probabilities, the selected context from among the plurality of candidate contexts;
and
enhancing the RAG language model by applying only the selected context to the RAG language model to generate an enhanced RAG language model.

10 more elements

A system for increasing a computational efficiency of a retrieval augmented generation (RAG) language model having electronic access to a plurality of potential contexts, the system comprising:

a computer processor and the RAG language model, executable by the computer processor, wherein the computer processor comprises a hardware computer processor;
a data repository in communication with the computer processor and storing:
a query,
the plurality of potential contexts,
a selected context applicable to the query,
a corresponding candidate context, and
a plurality of vector data structures,
a plurality of ranking models which, when executed by the computer processor on the query, outputs, for each of the plurality of ranking models, the corresponding candidate context from among the plurality of potential contexts,
wherein each of the plurality of ranking models are different from each other, and
wherein a plurality of candidate contexts are identified;
a vector model which, when executed by the computer processor, transforms a combination of the query and the corresponding candidate context into a corresponding vector data structure for each output of the plurality of ranking models, wherein:
the plurality of vector data structures are generated by transforming,
each of the plurality of vector data structures corresponds to a different output of the plurality of ranking models, and
transforming contextually relates each of the plurality of candidate contexts to the query;
a predictor model which, when executed by the computer processor on the plurality of vector data structures, generates a plurality of probabilities, wherein each of the plurality of probabilities indicates a corresponding likelihood that the corresponding candidate context is the selected context;
and
a server controller which, when executed by the computer processor, enhances the RAG language model by:
selecting, according to the plurality of probabilities, the selected context from among the plurality of candidate contexts, and
applying only the selected context to the RAG language model to generate an enhanced RAG language model.

16 more elements

Dependencies are inferred from the claim text.