Dynamic AI system for context-aware, domain-specific tasks
The described system provides a novel, efficient, and scalable approach to using LLMs for domain-specific applications. By combining a pre-trained LLM model with a semantic search engine and a range of specialized modules, the system overcomes the limitations of traditional LLM-based systems. It offers accurate, context-aware responses to domain-specific queries without the need for costly retraining, making it an ideal solution for dynamic and complex business environments.
1 . A computer-implemented method comprising:
receiving a query from a client;
initiating a classification process of the query that generates an action attribute for the query;
initiating an extraction process of the query that generates a parameter attribute for the action attribute of the query;
initiating a plurality of parameter functions based on the parameter attribute that generate at least one or more outputs; and
generating, using a large language model (LLM), a response to the query using the action attribute generated by the classification process, the parameter attribute generated by the extraction process, and the one or more outputs generated by the plurality of parameter functions;
wherein the outputs generated by the plurality of parameter functions comprise at least generating a sentiment classification, generating semantically similar content, and generating a dynamically constructed prompt, wherein receiving the query comprises simultaneously transmitting the query to (i) a question-and-response endpoint application programming interface (API) and (ii) a sentiment predictor endpoint API,
wherein generating the sentiment classification comprises the sentiment predictor endpoint API forwarding the query to a sentiment predictor engine that invokes a pre-trained LLM in a sentiment prediction mode to output the sentiment classification,
wherein generating the semantically similar content comprises initiating an augmenting content retriever engine that obtains semantically similar content to the query from an augmenting content store using a semantic search engine,
wherein the augmenting content store is maintained by a content updater that adds new domain-specific content and removes content items that are superseded by newer versions of corresponding content items stored in the augmenting content store,
wherein generating the dynamically constructed prompt comprises a dynamic augmented prompt builder constructing an enriched prompt by combining (a) the query, (b) the action attribute, (c) the parameter attribute, (d) the sentiment classification, and (e) the semantically similar content obtained from the augmenting content store, and
wherein generating the response comprises invoking the pre-trained LLM using the enriched prompt to generate the response.
2 . The computer-implemented method of claim 1 , wherein generating the sentiment classification associated with outputs generated by the plurality of parameter functions comprises initiating a sentiment classification process that uses the LLM to generate a sentiment for the query.
3 . The computer-implemented method of claim 1 , wherein generating semantically similar content comprises initiating a semantic search engine that obtains semantically similar content to the query from a content specific data store.
4 . The computer-implemented method of claim 1 , wherein generating the dynamically constructed prompt comprises using the semantically similar content generated by initiating the semantic search engine.
5 . The computer-implemented method of claim 1 , wherein the classification process and the extraction process of the query each comprise an artificial intelligence (AI) model, wherein the AI model is trained using labeled domain-specific data.
6 . The computer-implemented method of claim 1 , wherein the LLM comprises a pre-trained LLM and a Generative Pretrained Transformer (GPT).
7 . The computer-implemented method of claim 1 , wherein the action attribute of the query specifies an intent of a user, wherein the user is associated with the client.
8 . The computer-implemented method of claim 7 , wherein the parameter attribute of the query specifies a set of parameters associated with the intent of the user.
9 . The computer-implemented method of claim 1 , wherein the response comprises information requested in the query.
10 . A system comprising:
one or more computing processors; and
a non-transitory machine-readable storage medium storing instructions that, when executed by the one or more processors, cause the system to:
receive a query from a client;
initiate a classification process of the query that generates an action attribute for the query;
initiate an extraction process of the query that generates a parameter attribute for the action attribute of the query;
initiate a plurality of parameter functions based on the parameter attribute that generate at least one or more outputs; and
generate, using a large language model (LLM), a response to the query using the action attribute generated by the classification process, the parameter attribute generated by the extraction process, and the one or more outputs generated by the plurality of parameter functions;
wherein the outputs generated by the plurality of parameter functions comprise at least a sentiment classification, a semantically similar content and a dynamically constructed prompt, wherein receiving the query comprises simultaneously transmitting the query to (i) a question-and-response endpoint application programming interface (API) and (ii) a sentiment predictor endpoint API,
wherein generating the sentiment classification comprises the sentiment predictor endpoint API forwarding the query to a sentiment predictor engine that invokes a pre-trained LLM in a sentiment prediction mode to output the sentiment classification,
wherein generating the semantically similar content comprises initiating an augmenting content retriever engine that obtains semantically similar content to the query from an augmenting content store using a semantic search engine,
wherein the augmenting content store is maintained by a content updater that adds new domain-specific content and removes content items that are superseded by newer versions of corresponding content items stored in the augmenting content store,
wherein generating the dynamically constructed prompt comprises a dynamic augmented prompt builder constructing an enriched prompt by combining (a) the query, (b) the action attribute, (c) the parameter attribute, (d) the sentiment classification, and (e) the semantically similar content obtained from the augmenting content store, and
wherein generating the response comprises invoking the pre-trained LLM using the enriched prompt to generate the response.
11 . The system of claim 10 , wherein the sentiment classification associated with the outputs generated by the plurality of parameter functions comprises initiating a sentiment classification process that uses the LLM to generate a sentiment for the query.
12 . The system of claim 10 , wherein the semantically similar content comprises initiating a semantic search engine that obtains semantically similar content to the query from a content specific data store.
13 . The system of claim 10 , wherein the dynamically constructed prompt comprises the semantically similar content generated by initiating the semantic search engine.
14 . The system of claim 10 , wherein the classification process and the extraction process of the query each comprise an artificial intelligence (AI) model, wherein the AI model is trained using labeled domain-specific data.
15 . The system of claim 10 , wherein the LLM comprises a pre-trained LLM and a Generative Pretrained Transformer (GPT).
16 . The system of claim 10 , wherein the action attribute of the query specifies an intent of a user, wherein the user is associated with the client.
17 . The system of claim 16 , wherein the parameter attribute of the query specifies a set of parameters associated with the intent of the user.
18 . The system of claim 10 , wherein the response comprises information requested in the query.