# US 12,762,576 B2 · Systems and methods for generating actionable regulatory insights for predictive analytics > Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt - Number: US 12,762,576 B2 - Document ID: US12762576B2 - Status: Granted - Kind: B2 - Application number: 19/005,134 - Filed: 2024-12-30 - Published: 2026-07-02 - Granted: 2026-10-13 - Earliest claimed: 2024-12-30 - Assignees: RegASK Pte. Ltd. - Applicants: RegASK Pte. Ltd. - Inventors: Caroline Shleifer; Amenallah Reghimi - Web page: https://patentgrep.com/patent/US12762576B2 ## CPC - G06Q 30/018: Certifying business or products - G06F 16/953: Querying, e.g. by the use of web search engines - G06F 18/241: relating to the classification model, e.g. parametric or non-parametric approaches - G06Q: INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR - G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N) ## Abstract Systems and methods for generating actionable regulatory insights for predictive analytics are disclosed. A method comprises receiving, by an artificial intelligence (AI) model, a regulatory query from a client device. A determination is made whether the query requires new data retrieval or if existing data can answer the query. When the query requires new data retrieval, a Web search is activated, which includes prompting an external large language model (LLM) to provide a regulatory document that answers the regulatory query. The regulatory document is retrieved in answer to the regulatory query. The answer to the query is transmitted to the client computing device. The generated answer, regulatory content, and the regulatory documents are provided in a feedback loop to the AI model, to improve training data. The client computing device is provided an actionable regulatory insight, including a recommended workflow of tasks based on the recommendation regarding the regulatory content. ## Claims ### Claim 1 A method for AI-powered document retrieval with hallucination prevention, the method comprising: receiving, by a generative AI (GEN-AI) agent executing on one or more processors coupled to a memory of an artificial intelligence (AI) model, a regulatory query from a client computing device associated with an organization, the regulatory query comprising a question related to a regulation relevant to an industry of the organization, the GEN-AI agent comprising specialized components including an index evaluator, a document retriever, a document grader, a web searcher, and an answer generator; determining, by the index evaluator, whether the regulatory query requires new data retrieval from a web search or whether existing internal data accessible to the AI model can answer the regulatory query; when the new data retrieval is required, activating, by the web searcher, the web search utilizing an external large language model (LLM) to identify and retrieve a regulatory document that answers the regulatory query, the utilizing comprising transmitting a structured query request from the one or more processors to the external LLM via a network interface; retrieving, by the document retriever, the regulatory document in answer to the regulatory query from at least one of the existing internal data or from the external LLM; evaluating, by the document grader, the retrieved regulatory document for regulatory relevance to the regulatory query, the evaluating comprising: (i) automatically cross-referencing the retrieved regulatory document against multiple independent online sources to confirm factual accuracy and identify inconsistencies indicative of hallucinated content, the cross-referencing comprising executing comparison algorithms stored in the memory to detect textual inconsistencies between the retrieved regulatory document and the multiple independent online sources, thereby computationally identifying hallucinated content through automated discrepancy detection and thereby improving accuracy of computer-generated regulatory responses; and (ii) running a supplemental web search to verify that the retrieved regulatory document is relevant and accurate and is not a hallucination; generating, by the answer generator, a generated answer to the regulatory query based on the retrieved regulatory document, the generating comprising implementing a multi-tiered validation framework comprising: (i) a first AI agent using a first model generating a candidate answer and producing a first output; (ii) a second AI agent using a second model architecturally distinct from the first model independently validating the first output by performing parallel analysis of the retrieved regulatory document to produce a second output; (iii) automatically comparing the first output and the second output to identify semantic discrepancies, the identification of semantic discrepancies exceeding a predetermined threshold stored in the memory indicating hallucinated content, the predetermined threshold comprising a numerical similarity score value and discrepancies exceeding the predetermined threshold triggering automated re-processing of the regulatory query, thereby reducing hallucinations in computer-generated regulatory answers through computational model disagreement analysis; (iv) prompting the external LLM with follow-up verification questions to check veracity of the generated answer and reduce chances of hallucination by leveraging generative capabilities of the external LLM; and (v) utilizing a vertically-trained artificial intelligence model, specialized in the organization's specific regulatory domain, to establish context of the retrieved regulatory document and reduce chances that hallucinated content from the external LLM will be incorporated into the generated answer; and transmitting from the one or more processors to the client computing device via the network interface, the generated answer including regulatory content from the retrieved regulatory document and a reference to the retrieved regulatory document. ### Claim 2 (depends on 1) The method of claim 1, further comprising requesting the organization to verify whether the GEN-AI agent is to perform the web search online regarding the regulatory query or answer the regulatory query using only the existing internal data. ### Claim 3 (depends on 2) The method of claim 2, wherein the existing internal data comprises at least one of existing internal data stored in a database or a server associated with the AI model, data that is uploaded by the organization and stored by the AI model, data in the organization's private LLM that is accessible to the AI model, or training data of the AI model. ### Claim 4 (depends on 1) The method of claim 1, further comprising evaluating and grading the retrieved regulatory document. ### Claim 5 (depends on 1) The method of claim 1, wherein the generated answer includes an indication whether it was derived from the organization's own data collection, from the organization's private LLM, or from a search online. ### Claim 6 (depends on 1) The method of claim 1, wherein activating the web search further includes sourcing supplementary information from the Web to enhance an answer to the regulatory query, when the existing internal data is insufficient. ### Claim 7 (depends on 1) The method of claim 1, further comprising: prompting the external large language model (LLM) to provide a plurality of recommended tasks to be implemented for regulatory compliance based on the regulatory document; requesting the external large language model (LLM) to test its plurality of recommended tasks in parallel tests; based on outcomes of the parallel tests, based on outcomes of the parallel tests, asking the external large language model (LLM) to prioritize its plurality of recommended tasks in order based on optimization; and obtaining the plurality of recommended tasks and prioritization from the external large language model (LLM). ### Claim 8 A system for AI-powered document retrieval with hallucination prevention, the system comprising: one or more processors; a memory coupled to the one or more processors; a network interface coupled to the one or more processors and to a network, the network interface configured to communicate with at least one client computing device via the network; and a generative AI (GEN-AI) agent executing on the one or more processors and communicatively coupled to the memory and the network interface, the GEN-AI agent comprising specialized components including an index evaluator, a document retriever, a document grader, a web searcher, and an answer generator; the one or more processors being configured to execute instructions stored in the memory to: receive, by the GEN-AI agent, a regulatory query from a client computing device associated with an organization, the regulatory query comprising a question related to a regulation relevant to an industry of the organization; determine, by the index evaluator, whether the regulatory query requires new data retrieval from a web search or whether existing internal data accessible to an artificial intelligence (AI) model can answer the regulatory query; when the new data retrieval is required, activate, by the web searcher, the web search utilizing an external large language model (LLM) to identify and retrieve a regulatory document that answers the regulatory query, the utilizing comprising transmitting a structured query request from the one or more processors to the external LLM via the network interface; retrieve, by the document retriever, the regulatory document in answer to the regulatory query from at least one of the existing internal data or from the external LLM; evaluate, by the document grader, the retrieved regulatory document for regulatory relevance to the regulatory query, the evaluating comprising: (i) automatically cross-referencing the retrieved regulatory document against multiple independent online sources to confirm factual accuracy and identify inconsistencies indicative of hallucinated content, the cross-referencing comprising executing comparison algorithms stored in the memory to detect textual inconsistencies between the retrieved regulatory document and the multiple independent online sources, to computationally identify hallucinated content through automated discrepancy detection and improve accuracy of computer-generated regulatory responses; and (ii) running a supplemental web search to verify that the retrieved regulatory document is relevant and accurate and is not a hallucination; generate, by the answer generator, a generated answer to the regulatory query based on the retrieved regulatory document, the generating comprising implementing a multi-tiered validation framework comprising: (i) a first AI agent using a first model, to generate a candidate answer and produce a first output; (ii) a second AI agent using a second model architecturally distinct from the first model, to independently validate the first output by performing parallel analysis of the retrieved regulatory document to produce a second output; (iii) automatically comparing the first output and the second output to identify semantic discrepancies, the identification of semantic discrepancies exceeding a predetermined threshold stored in the memory indicating hallucinated content, the predetermined threshold comprising a numerical similarity score value and discrepancies exceeding the predetermined threshold triggering automated re-processing of the regulatory query, to reduce hallucinations in computer-generated regulatory answers through computational model disagreement analysis; (iv) prompting the external LLM with follow-up verification questions to check veracity of the generated answer and reduce chances of hallucination by leveraging generative capabilities of the external LLM; and (v) utilizing vertical artificial intelligence to establish context of the retrieved regulatory document and reduce chances that hallucinated content from the external LLM will be incorporated into the generated answer; and transmit, from the one or more processors to the client computing device via the network interface, the generated answer including regulatory content from the retrieved regulatory document, and a reference to the retrieved regulatory document. ### Claim 9 (depends on 8) The system of claim 8, wherein the one or more processors are further configured to evaluate and grade the retrieved regulatory document. ### Claim 10 (depends on 8) The system of claim 8, wherein the one or more processors are further configured to request the organization to verify whether the GEN-AI agent is to perform the web search online regarding the regulatory query or answer the regulatory query using only the existing internal data. ### Claim 11 (depends on 8) The system of claim 8, wherein the generated answer includes an indication whether it was derived from the organization's own data collection, from the organization's private LLM, or from a search online. ### Claim 12 (depends on 8) The system of claim 8, wherein activating the web search further includes sourcing supplementary information from the Web to enhance an answer to the regulatory query, when the existing internal data is insufficient. ### Claim 13 (depends on 8) The system of claim 8, wherein the one or more processors are further configured to: prompt the external large language model (LLM) to provide a plurality of recommended tasks to be implemented for regulatory compliance based on the regulatory document; request the external large language model (LLM) to test its plurality of recommended tasks in parallel tests, based on outcomes of the parallel tests; based on outcomes of the parallel tests, ask the external large language model (LLM) to prioritize its plurality of recommended tasks in order based on optimization; and obtain the plurality of recommended tasks and prioritization from the external large language model (LLM).