US 12,596,710 B1Granted

AI-assisted natural language request transformation with reduced cognitive load

AssigneeSap SE

InventorsKarl Fuerst, Daniel Steinberg, Erika Fonseca de Lima, Muamer Rovcanin, Pranadharthi Haran Sathiyaseelan, Allan Felipe Brochier dos Santos da Silva

Life of this patent

Estimating the term…

patentgrep

Loading patent…

Abstract

Strategic actions can be performed before and during runtime to reduce the cognitive load of an LLM during transformation of a natural language request into a refined structured request. During a development stage, object-type-specific examples are collected; a given example includes an example natural language request and a corresponding example structured request representing a desired LLM output for the example natural language request. The examples can include function calls to train the LLM to include function calls in its output to delegate certain sub-tasks. At runtime, a prompt generation module generates a prompt for a natural language request which includes the natural language request, one or more of the examples, and a data model. The prompt is input to the LLM which produces a preliminary structured output including one or more function calls. The post-processing module executes the function(s) and generates a refined structured output based on the execution results.

Claims20

Dependent claims

A computer-implemented method comprising:

receiving, by a prompt generation module, a natural language request;
generating, by the prompt generation module, a prompt for a Large Language Model (LLM) based on the natural language request, wherein the prompt comprises a data model defined in a data processing engine, one or more examples defined in a prompt generation database, and the natural language request, and wherein a given example of the one or more examples comprises an example natural language request and a corresponding example structured request, and wherein the example structured request comprises a call to a function corresponding to a sub-task to be performed programmatically;
submitting, by the prompt generation module, the prompt to the LLM;
receiving, by a post-processing module, a preliminary structured request from the LLM generated in response to the prompt, wherein the preliminary structured request comprises the call to the function;
post-processing the preliminary structured request, by the post-processing module, wherein post-processing the preliminary structured request comprises:
retrieving computer-executable code of the function from the prompt generation database;
executing the computer-executable code of the function;
and
generating a refined structured request, wherein generating the refined structured request comprises modifying the preliminary structured request based on an execution result of the function, wherein the modifying comprises inserting the execution result of the function into the preliminary structured request;
outputting the refined structured request to the data processing engine;
and
generating, by the data processing engine, a response to the natural language request using the refined structured request.

9 more elements

A computing system comprising:

at least one hardware processor;
at least one memory coupled to the at least one hardware processor;
a prompt generation database comprising a plurality of stored prompt templates, wherein a given prompt template of the plurality of stored prompt templates comprises a data model defined in a data processing engine and one or more examples defined in the prompt generation database, wherein a given example of the one or more examples comprises an example natural language request and a corresponding example structured request, and wherein the example structured request comprises a call to a function corresponding to a sub-task to be performed programmatically;
and
one or more non-transitory computer-readable media having stored therein computer-executable instructions that, when executed by the computing system, cause the computing system to perform:
receiving, by a prompt generation module, a natural language request;
generating, by the prompt generation module, a prompt for a Large Language Model (LLM) based on the natural language request, wherein generating the prompt comprises:
determining an object type relevant to the natural language request;
selecting a prompt template from the plurality of stored prompt templates based on the object type;
and
populating the prompt template with the natural language request;
submitting, by the prompt generation module, the prompt to the LLM;
receiving, by a post-processing module, a preliminary structured request from the LLM generated in response to the prompt, wherein the preliminary structured request comprises the call to the function;
post-processing the preliminary structured request, by the post-processing module, wherein post-processing the preliminary structured request comprises:
retrieving computer-executable code of the function from the prompt generation database;
executing the computer-executable code of the function;
and
generating a refined structured request, wherein generating the refined structured request comprises modifying the preliminary structured request based on an execution result of the function, wherein the modifying comprises inserting the execution result of the function into the preliminary structured request;
outputting the refined structured request to the data processing engine;
and
generating, by the data processing engine, a response to the natural language request using the refined structured request.

18 more elements

One or more non-transitory computer-readable media comprising computer-executable instructions that, when executed by a computing system, cause the computing system to perform operations comprising:

receiving, by a prompt generation module, a natural language request;
generating, by the prompt generation module, a prompt for a Large Language Model (LLM) based on the natural language request, wherein generating the prompt comprises:
selecting, from a plurality of prompt templates stored in a prompt generation database, a prompt template comprising a data model defined in a data processing engine and one or more examples defined in the prompt generation database, wherein a given example of the one or more examples comprises an example natural language request and a corresponding example structured request comprising content expressed in a first query language;
and
populating the prompt template with the natural language request;
submitting, by the prompt generation module, the prompt to the LLM;
receiving, by a post-processing module, a preliminary structured request from the LLM generated in response to the prompt, wherein the preliminary structured request comprises content expressed in the first query language and a call to a function corresponding to a sub-task to be performed programmatically;
post-processing the preliminary structured request, wherein post-processing the preliminary structured request comprises:
retrieving computer-executable code of the function from the prompt generation database;
executing the computer-executable code of the function, wherein the executing comprises translating the content expressed in the first query language into content expressed in a second query language;
and
generating a refined structured request, wherein generating the refined structured request comprises modifying the preliminary structured request, comprising:
replacing the content expressed in the first query language with the content expressed in the second query language, and
inserting an execution result of the function into the preliminary structured request;
outputting the refined structured request to the data processing engine;
and
generating, by the data processing engine, a response to the natural language request using the refined structured request.

14 more elements

Dependencies are inferred from the claim text.