DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Status of Claims
Claims 1-12, and 15-20 are pending of which claims 1, 9 and 17 are in independent form.
Claims 1-12, and 15-20 are rejected under 35 U.S.C. 101.
Claims 1-12, and 15-20 are rejected under 35 U.S.C. 103.
Response to Arguments
Applicant’s arguments with respect to claim(s) 1-12, and 15-20 have been considered but are moot because the new ground of rejection does not rely on any reference applied in the prior rejection of record for any teaching or matter specifically challenged in the argument.
Regarding 35 USC 101 (Abstract Idea)
Applicant’s arguments have been fully considered but are not persuasive. While Applicant contends that the claimed LLM based embedding generation, query expansion, and cluster fine-tuning cannot practically be performed in the human mind, the Examiner notes that claim 1 recites the abstract concept of evaluating, categorizing, and updating information based on preferred criteria. The examiner acknowledges that the claimed ML operations are computer implemented and does not characterize those operations themselves as mental process. However, under Step 2A, Prong Two, the additional limitations merely recite these operations at a high level of function generality without specifying a particular technical mechanism that improves LLM functionality, query expansion technology, of computer operation. Furthermore, Applicant’s assertion that the invention addresses chatbot safety vulnerabilities does not, by itself, establish a technological improvement, as the claim does not sufficiently specify how the alleged improvement is technically achieved beyond the recited processing results. Accordingly, the claimed additional elements do not integrate the identified abstract idea into a practical application, and the rejection under 35 USC 101 is maintained.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1-12, and 15-20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to a judicial exception (i.e., a law of nature, a natural phenomenon, or an abstract idea) without significantly more.
The claim(s) recite(s) system and method for generating guardrail data structure in a LLM chatbot system.
With respect to step 1 of the patent subject matter eligibility analysis, the claims are directed to a process, machine, manufacture, or composition of matter.
Independent claims 1, and 17 are directed to a system, which comprises, one or more processors.
Independent claims 9 are directed to a method, which is a process.
Independent All other claims depend on claims 1, 9, and 17. As such, claims 1-20 are directed to a statutory category.
Regarding claims 1, 9 and 17:
With respect to step 2A, prong one (Judicial Exception), it is noted that the independent claims recite an abstract idea falling within the Mental Health grouping of abstract ideas. Specifically, the following limitations recite mathematical concepts and/or mental processes and/or certain methods of organizing human activity.
The claim recites the following limitations directed to an abstract idea:
“instructing a remote computing device to identify, using the guardrail data structure, flagged data based on the chat data” the limitation as drafted recites a mental process involving evaluation information against predefined conditions or criteria;
“fine-tuning, using the LLM, at least one cluster membership of a plurality of content clusters based on the expanded query dataset” the limitation as drafted recites a mental process involving determining which information is belonging to which category (classification and judgement).
“instructing the remote computing device to update the guardrail data structure based on the update to the at least one cluster membership” the limitation as drafted recites a mental process of applying decision (i.e. updating rules/classification based on evaluation).
These limitations correspond to concept that can be performed in the human mind (Mental Process) and therefore fall within the mental process category of abstract idea (see MPEP 2016.04(a)(2))
The claims fall within:
Thus, the claims recite an abstract idea (mental process/mathematical concepts/information processing).
With respect to step 2A, Prong Two (Particular Application), the claims do not recite additional elements that integrate the judicial exception into a practical application. The following limitations are considered “additional elements” and explanation will be given as to why these “additional elements” do not integrate the judicial exception into a practical application.
The claims recite the use of:
“one or more processors; or more transitory or non-transitory computer-readable media storing instructions that are executable to cause the one or more processors to perform operations” as drafted recites generic computer performing generic computer functions (see MPEP 2106.05(f)(2)); and generic storage and execution of instructions, which is conventional and amounts to mere instructions to apply the abstract idea using a computer (see MPEP 2106.05(f)).
“receiving, from a first user of a plurality of users of a guardrail data structure, interaction data associated with chat data” recites data gathering, which is insignificant extra-solution activity (see MPEP 2106.05(g)).
“instructing a remote computing device to identify, using the guardrail data structure, flagged data based on the chat data” as drafted recites directing another computing device to evaluate chat information using a guardrail data structure and identify information satisfying particular criteria. Accordingly, the limitation merely uses a remote computing device to implement the generic underlaying evaluation and identification process (see MPEP 2106.05(f)).
“generating, using a large language model (LLM), a plurality of embeddings based on the flagged data” as drafted recites using an LLM to generate embedding representation of the flagged data. The limitation, considered in the context of the limitation as a whole, does not sufficiently establish that the manner of generating embeddings provides a technological improvement to the functioning of the LLM of another computer technology (see MPEP 2106.04(d)(1)).
“generating, using a query expansion model, an expanded query dataset based on the plurality of embeddings” as drafted recites using a query expansion model to generate additional query information based on generated embeddings. The limitation functionally identifies the information supplied to the query expansion model and the resulting dataset without specifying how the claimed expansion improves the technological operations of the model, or the computer technology. Therefore, the additional elements, considered in the context of the claimed sequence, does not sufficiently establish a technological improvement integrating the judicial exception into a practical application (see MPEP 2106.04(d)(1)).
“fine-tuning, using the LLM, at least one cluster membership of a plurality of content clusters based on the expanded query dataset” recites using the LLM to refine the association of information with content clusters based on the expanded query dataset. The limitation, broadly recites the desired result of refining cluster membership based on additional information. The limitation does not sufficiently specify how the fine-tuning operation produces an improvement to the functioning of the LLM or guardrail technology beyond the stated classification result (see MPEP 2106.04(d)(1)).
“transmitting the update of the at least one cluster membership to a remote computing device” recites routine data transmission, which is conventional computer function and amounts to insignificant post-solution activity.
“instructing the remote computing device to update the guardrail data structure based on the update to the at least one cluster membership” recites application of the abstract idea, reflecting a result oriented functions limitations (update based on decisions), without specifying how the update is technically performed.
The LLM is used as a tool to perform the abstract classification and analysis. The guardrail data structure is simply a data representation of rule/policies. The computing device performs generic processing. The transmission step is routine data communication.
The claims do not:
Improve LLM architecture or training mechanism
Improve clustering algorithm technically,
Improve computer performance or memory
Provide a new data structure with technical functionality
Solve problem rooted in computer technology.
There are no improvements to computer functionality or any specific technical solution to a computer centric problem. Instead, the computer and semiconductor environment are used as tools to execute abstract mathematical encoding, data analysis, and decision making, with the result merely being applied in a generic manner.
There is no recitation of, a new data structure that changes computer operation, improved network functioning, an unconventional indexing technique, a specific hardware solution.
Instead, the claims recite conventional and generic computer functions performed in a routine manner, which does not amount to a practical application.
With respect to Step 2B. The claims do not include additional elements that are sufficient to amount to significantly more than the judicial exception. The recited components are merely generic computer/database elements performing their routine, well-understood, and conventional functions. See Alive, MPEP 2016.05(d).
The steps mentioned in the independent claims are merely generic computing device, LLM performing conventional process, data structure representing rules, routine transmission. Courts have consistently helped such high-level information management operations are conventional.
The claims recite only functional, result oriented language (“detecting”, “propagating”, “transferring”,…), without specifying any technical mechanism for performing these operations in a non-conventional manner.
Considering claims as a whole, the ordered combination of elements also reflects nothing more than the typical workflow of distributed systems, and therefore DOES NOT add “significantly more” than the abstract idea.
Such generic, high‐level, and nominal involvement of a computer or computer‐based elements for carrying out the invention merely serves to tie the abstract idea to a particular technological environment, which is not enough to render the claims patent‐eligible, as noted at pg.74624 of Federal Register/Vol. 79, No. 241, citing Alice, which in turn cites Mayo. Further, See, e.g., Alice Corp. Pty. Ltd. v. CLS Bank Int'l, 134 S. Ct. 2347, 2359‐60, 110 USPQ2d 1976, 1984 (2014). See also OIP Techs. v. Amazon.com, 788 F.3d 1359, 1364, 115 USPQ2d 1090, 1093‐94 (Fed. Cir. 2015) ("Just as Diehr could not save the claims in Alice, which were directed to 'implement[ing] the abstract idea of intermediated settlement on a generic computer', it cannot save O/P's claims directed to implementing the abstract idea of price optimization on a generic computer.") (citations omitted). See also, Affinity Labs of Texas LLC v. DirecTV LLC, 838 F.3d 1253, 1257‐1258 (Fed. Cir. 2016) (mere recitation of a GUI does not make a claimpatent‐eligible); Intellectual Ventures I LLC v. Capital One Bank, 792 F.3d 1363, 1370 (Fed. Cir. 2015) ("the interactive interface limitation is a generic computer element".).
The additional elements are broadly applied to the abstract idea at a high level of generality ("similar to how the recitation of the computer in the claims in Alice amounted to mere instructions to apply the abstract idea of intermediated settlement on a generic computer,") as explained in MPEP § 2106.05(f)) and they operate in a well‐understood, routine, and conventional manner.
MPEP § 2106.0S(d)(II) sets forth the following:
The courts have recognized the following computer functions as well-understood, routine, and conventional functions when they are claimed in a merely generic manner (e.g., at a high level of generality) or as insignificant extra-solution activity.
• Receiving or transmitting data over a network, e.g., using the Internet to gather data, Symantec ... ; TLI Communications LLC v. AV Auto. LLC ... ; OIP Techs., Inc., v. Amazon.com, Inc ... ; buySAFE, Inc. v. Google, Inc ... ;
• Performing repetitive calculations, Flook ... ; Bancorp Services v. Sun Life ... ;
• Electronic recordkeeping, Alice Corp ... ; Ultramercial ... ;
• Storing and retrieving information in memory, Versata Dev. Group, Inc. v. SAP Am., Inc ... ;
• Electronically scanning or extracting data from a physical document, Content Extraction and Transmission, LLC v. Wells Fargo Bank ... ; and
• A web browser's back and forward button functionality, Internet Patent
• Corp. v. Active Network, Inc. ...
. . . Courts have held computer-implemented processes not to be significantly more than an abstract idea (and thus ineligible) where the claim as a whole amounts to nothing more than generic computer functions merely used to implement an abstract idea, such as an idea that could be done by a human analog (i.e., by hand or by merely thinking).
In addition, when taken as an ordered combination, the ordered combination adds nothing that is not already present as when the elements are taken individually. There is no indication that the combination of elements integrate the abstract idea into a practical application. Their collective functions merely provide conventional computer implementation. Therefore, when viewed as a whole, these additional claim elements do not provide meaningful limitations to transform the abstract idea into a practical application of the abstract idea or that the ordered combination amounts to significantly more than the abstract idea itself.
The dependent claims have been fully considered as well, however, similar to the findings for claims above, these claims are similarly directed to the “Mental Processes” grouping of abstract ideas set forth in the 2019 PEG, without integrating it into a practical application and with, at most, a general purpose computer that serves to tie the idea to a particular technological environment, which does not add significantly more to the claims. The ordered combination of elements in the dependent claims (including the limitations inherited from the parent claim(s)) add nothing that is not already present as when the elements are taken individually. There is no indication that the combination of elements improves the functioning of a computer or improves any other technology. Their collective functions merely provide conventional computer implementation. Accordingly, the subject matter encompassed by the dependent claims fails to amount to significantly more than the abstract idea.
Looking at the claim as a whole does not change this conclusion and the claim is ineligible.
Regarding claims 2, 3, 4, and 18 (Flagging, Classification, and Content Moderation),
The claim recites:
Identifying flagged data (claim 2 and 18)
Classifying chat data into content cluster (claims 3 and 18)
Remove flagged data (claims 4 and 13)
This merely applies rules to: evaluate content, categorize content, take action (remove flag). There are no changes to: how classification is performed, how clusters are generated, or any underlaying technical mechanism. These fall under: Mental Process (evaluation/categorization), and Data Processing.
This does not change the nature of the abstract idea. It does not add a technical improvement to an abstract idea, such as improving computer functionality, data structure, or processing architecture.
There is no practical application, and no inventive step, the claims are still considered abstract.
Regarding claims 5, 6, 10, and 11 (Verification and Comparison Logic),
The claim recites:
Verification process to verify interaction data (claim 5 and 10)
Generating a vector, comparing to clusters, and verifying (claims 6 and 11)
This merely performs: data analysis, comparison, decision making. There are no changes to: how vectors are generated, how similarity is computed, or any technical improvement in comparison algorithms. These fall under: Mental Process (comparison and evaluation), and Mathematical Algorithm (vector representation and similarity).
This does not change the nature of the abstract idea. It does not add a technical improvement to an abstract idea, such as improving computer functionality, data structure, or processing architecture.
There is no practical application, and no inventive step, the claims are still considered abstract.
Regarding claims 7, and 12 (Notification/Output),
The claim recites:
Generating a notification base on guardrail updates
This merely reports results of the previous analysis. There are no changes to: how notifications are generated technically, any communication protocol, or system performance. These fall under: Insignificant Extra-Solution Activity.
This does not change the nature of the abstract idea. It does not add a technical improvement to an abstract idea, such as improving computer functionality, data structure, or processing architecture.
There is no practical application, and no inventive step, the claims are still considered abstract.
Regarding claims 8, 16, and 20 (Training Data Generation and Model Training),
The claim recites:
Generating training data from interaction data
Training the LLM using that data
This merely performs: data preparation, model training. There are no changes to: the LLM architecture, how training is technically performed, or any learning algorithm. These fall under: Mental Process (data preparation), and Mathematical Algorithm (model training).
This does not change the nature of the abstract idea. It does not add a technical improvement to an abstract idea, such as improving computer functionality, data structure, or processing architecture.
There is no practical application, and no inventive step, the claims are still considered abstract.
Regarding claims 13, and 14 (Canceled).
Regarding claim 15 (Model Refinement/Fine-Tuning),
The claim recites:
Fine-tuning cluster membership using the LLM
Based on expanded query data
This merely refine prior results using additional data. There are no changes to: how the model in trained technically, any optimization algorithm, or computational efficiency. These fall under: Mental Process (refinement and adjustment), and Mathematical Algorithm (model updating).
This does not change the nature of the abstract idea. It does not add a technical improvement to an abstract idea, such as improving computer functionality, data structure, or processing architecture.
There is no practical application, and no inventive step, the claims are still considered abstract.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claim(s) 1-12, and 15-20 are rejected under 35 U.S.C. 103 as being unpatentable over Taylor; Bret Steven et al. (US 20250258856 A1) [Taylor] in view of D'Agostino; Dino Paul et al. (US 20250307562 A1) [D'Agostino] in view of Hu; Chenhui et al. (US 20250373627 A1) [Hu] in view of Crabtree; Jason et al. (US 20250259041 A1) [Crabtree].
Regarding claim 1, Taylor discloses, a system for managing a [guardrail] data structure, comprising: one or more processors; and one or more transitory or non-transitory computer-readable media storing instructions that are executable to cause the one or more processors to perform operations (see Fig. 5), the operations comprising: receiving, from a first user of a plurality of users of a guardrail data structure, interaction data associated with chat data (receive user input conversation ¶ [0037]. Including inputs (e.g., user queries or session variables) and outputs (e.g., responses or actions) ¶ [0026], [0039], multi-user chat interaction/chatbots ¶ [0022], [0023], [0062]. the classification module may deploy guardrails when using the LLM to classify intent of an end-user ¶ [0033]-[0035], [0051]);
fine-tuning, using the LLM, at least one cluster membership of a plurality of content clusters [based on the expanded query datasets] (FIG. 3A illustrates a process of using LLM to classify conversations. For instance, an agent receives a new message (e.g., a new classification target) from a user, “Hey, I want to set up my XYZ.” The agent receives the input and uses an LLM (e.g., Generative AI 140) to determine whether the message can be classified into an existing category. … . The data store 214 may include both existing categories and the proposed categories, and the proposed categories may include both the previously proposed categories and the newly proposed categories ¶ [0037]-[0040]; In some embodiments, the classification module 204 may periodically update the categories based on the most recent data and/or refine categories through light human configuration and customer feedback. The classification module 204 may include a recomputing cluster history with a rolling window which continuously updates and recalculates clusters/categories of data over a specified time frame. A rolling window may be used to analyze and update the categories based on most recent data while removing old data outside the window. The categories may be recomputed regularly, using only data within the most recent time window. The categories may be updated dynamically, and as new data comes in, older data falls out of the window. In some examples, the categories may be updated in real time. By continuously recalculating categories, the classification module 204 can adapt to emerging trends, shifts in customer behavior, or changes in conversation topics, maintaining the relevance of the categories. The classification module 204 may limit the number of additional categories to around a certain number, e.g., 10, 20 or so. As the rolling window moves forward and new data comes in, some categories that are no longer relevant or less frequently occurring will “roll off,” e.g., removed from the data store 214 ¶ [0043]);
[instructing the remote computing device] to update the guardrail data structure based on the update to the at least one cluster membership (deploy guardrails…rules, constraints, safety mechanism ¶ [0033], rules define when guardrails override classification ¶ [0034], guardrails modify system behavior ¶ [0051]. Also see ¶ [0037]);
[instructing a remote computing device] to identify, using the guardrail data structure (In some embodiments, the classification module 204 may deploy guardrails when using the LLM to classify intent of an end-user. Guardrails may refer to predefined categories, rules, constraints, or safety mechanisms that override or guide the LLM's classification. In one implementation, the classification module 204 may define guardrails based on situations where the set of categories may not align with pre-defined criteria. These scenarios typically include cases where sentiment, compliance, or priority considerations take precedence over a purely topic-based classification ¶ [0033]), flagged data ( In some embodiments, the classification module 204 may establish specific rules and conditions that dictate when a guardrail should override an LLM classification. For example, the classification module 204 may determine the rules based on keyword matching, sentiment analysis, contextual cues, compliance triggers, etc. For instance, if an input contains strong dissatisfaction indicators such as “unhappy,” “frustrated,” or “not acceptable,” the classification module 204 may automatically classify it under “Dissatisfaction” rather than its originally assigned category. By defining these conditions, the classification module gains a structured mechanism for refining LLM output and ensuring proper categorization ¶ [0034]) based on the chat data (For example, while the LLM analyzes the term “return” and will output an initial classification in category A, the external software system 115 may deploy a sentimental trigger, e.g., a “dissatisfaction” guardrail, where when a user expresses dissatisfaction, the user is directed to a certain outcome (e.g., a human representative or some other intervention). The classification module 204 determines a “Dissatisfaction” guardrail based on the sentimental trigger. When such a guardrail is deployed, “I am dissatisfied with the return process” may be categorized by the “Dissatisfaction” guardrail despite it matching the Return category “A” ¶ [0035]);
generating, using a large language model (LLM), a plurality of embeddings (FIG. 3A illustrates a process of using LLM to classify conversations. For instance, an agent receives a new message (e.g., a new classification target) from a user, “Hey, I want to set up my XYZ.” The agent receives the input and uses an LLM (e.g., Generative AI 140) to determine whether the message can be classified into an existing category. If the message can be classified into an existing category, such as “Account Setup” or “Installation Instructions,” the LLM outputs a tag and the agent tags the conversation with the identified existing category. If the LLM cannot identify an existing category to classify the conversation, the LLM may determine whether the message is aligned with a previously proposed category. If an aligned previously proposed category is identified, the LLM outputs the tag of the previously proposed category and the agent tags the conversation with the previously proposed category. If an aligned previously proposed category is not identified, the LLM may output a proposed new category for tagging the conversation. In some implementations, a human agent may receive the new category and evaluate whether the new category is valid, actuate, etc. The tagged conversation and the corresponding tag are stored in a data store 214 which is accessible by the LLM in subsequent clustering steps. The data store 214 may include both existing categories and the proposed categories, and the proposed categories may include both the previously proposed categories and the newly proposed categories ¶ [0037]-[0040]);
[generating, using a query expansion model, an expanded query dataset] based on the plurality of embeddings (The classification module 204 may generate one or more embeddings to represent the user input and determines a category of the user input by clustering the respective embeddings. Embeddings are dense vector representations of words, sentences, documents, etc. that capture semantic relationships in a continuous vector space. To generate embeddings, the classification module 204 may use text vectorization that transforms textual data into a numerical format that the LLM can understand and process. In some implementations, the classification module 204 may pre-process the text data, such as removing punctuation, special characters, etc., and normalize the text data to generate the embeddings ¶ [0038]-[0040]);
However, Taylor does not explicitly facilitate transmitting the update of the at least one cluster membership to a remote computing device; instructing the remote computing device.
D’Agostino discloses, transmitting the update of the at least one cluster membership to a remote computing device (tracking conversations across systems ¶ [0063], generating outputs/reports ¶ [0099], system processing and responses ¶ [0120]. Adaptation process may involve fine-tuning existing parameters, updating weights, or training new layers within the model to better capture the nuances of the conversation and generate more contextually appropriate responses ¶ [0162]. Also see ¶ [0098], [0135]).
It would have been obvious to one ordinary skilled in the art before the effective filing date of the claimed invention to combine the teachings of the cited references because D’Agostino’s system would have allowed Taylor to facilitate transmitting the update of the at least one cluster membership to a remote computing device; instructing the remote computing device. The motivation to combine is apparent in the Taylor’s reference, because there is a need to improve identifying subsets/contextual information in chatbot interactions.
However, neither Tylor nor D’Agostino explicitly facilitates instructing a remote computing device; based on the flagged data.
Hu discloses, instructing a remote computing device (The backdoor attack 431 can be an attack against the LLM. The backdoor attack 431 can be a hidden trigger which can be embedded during training and can be configured to exhibit malicious behaviors, for example on specific inputs while functioning normally otherwise. The backdoor attack 431 can define an input-trigger, which can create poisoned data with specific triggers. In other aspects the backdoor attack 431 can define a prompt-trigger which can modify prompts to generate harmful outputs. In yet other aspects, the backdoor attack 431 can define demonstration triggers which can alter demonstration data and can lead to incorrect outputs. In yet further aspects, the backdoor attack 431 can define instruction triggers, which can introduce harmful instructions during tuning via for example crowdsourcing. An example of the backdoor attack 431 can include ProAttack which can be a clean-label backdoor attack 431 and can leverage prompts as triggers without external markers and can ensure correct labeling of poisoned samples. The backdoor attack 431 can be configured to make LLMs perform normally on most inputs but maliciously on specific triggers ¶ [0048]-[0046]. Some aspects of the present disclosure pertain to the defense against jailbreaking. More specifically, the disclosure provides a benchmark dataset of LLM defense on multi-turn AI-Human conversation. As used herein, the term “jailbreaking” generally refers to a process of manipulating a LLM to bypass or partially bypass its built-in safety protocols, restrictions, or ethical guidelines. Moreover, jailbreaking an LLM can incorporate some strategy to elicit harmful content from the LLM. Jailbreaking can involve creating carefully designed prompts that deceive the LLM into ignoring the safety constraints and encourage the generation of malicious responses that are otherwise restricted ¶ [0060]-[0061], [0064]-[0066], [0070]-[0072]. Also see Fig. 1);
based on the flagged data (Such an approach can include employing a one-or-more step process fine-tuning procedure, for example by first combining backdoor weights optimized on poisoned data with pretrained weights and then refining these combined weights on a small set of clean data. The defense 404 can include embedding purification. Embedding purification can include targeting potential backdoors in word embedding and refining the embeddings to remove malicious triggers. The defense 404 can include clustering-based approaches ¶ [0049], [0058]-[0059], also see Figs. 4 and 5);
It would have been obvious to one ordinary skilled in the art before the effective filing date of the claimed invention to combine the teachings of the cited references because Hu’s system would have allowed Taylor and D’Agostino to facilitate transmitting the update of the at least one cluster membership to a remote computing device; instructing the remote computing device. The motivation to combine is apparent in Taylor and D’Agostino’s reference, because there is a need for advanced methods for LLM security.
However, neither on of Taylor, D’Agostino, or Hu explicitly facilitates generating, using a query expansion model, an expanded query dataset; based on the expanded query datasets.
Crabtree discloses, generating, using a query expansion model, an expanded query dataset; based on the expanded query datasets (This architecture contrasts with traditional retrieval methods that pre-fetch or summarize the entire domain corpus upfront. Instead, the enhanced LazyGraphRAG performs iterative best-first retrieval and dynamic expansion of partial queries, minimizing overhead while preserving maximum context relevance. When an agent or LLM subtask encounters a question or partial request—for example, a specialized medical agent needing to confirm dosing guidelines—the system sends a targeted subquery to the knowledge graph ¶ [0071], [0072], [0076], [0308]-[0309]);
It would have been obvious to one ordinary skilled in the art before the effective filing date of the claimed invention to combine the teachings of the cited references because Crabtree’s system would have allowed Taylor, D’Agostino, and Hu to facilitate generating, using a query expansion model, an expanded query dataset; based on the expanded query datasets. The motivation to combine is apparent in Taylor, D’Agostino, and Hu’s reference, because there is a need for managing distributed agent interactions efficiently while maintaining rigorous compliance with ethical, legal, and operational constraints, all while scaling effectively across heterogeneous computing environments and dynamically adapting to evolving regulatory and operational requirements.
Regarding claim 2, the combination of Taylor, D’Agostino, Hu, and Crabtree discloses, wherein the operations further comprise instructing the remote computing device to identify, [using the updated guardrail data structure], flagged data based on the chat data (D’Agostino: detects violations…flags the non-compliant conversations ¶ [0098]. Applying compliance rules to identify violations ¶ [0120]).
using the updated guardrail data structure (Taylor: deploy guardrails…rules, constraints, safety mechanism ¶ [0033], rules define when guardrails override classification ¶ [0034], guardrails modify system behavior ¶ [0051]. Also see ¶ [0037]).
Regarding claim 3, the combination of Taylor, D’Agostino, Hu, and Crabtree discloses, wherein identifying the flagged data comprises classifying (D’Agostino: detects violations…flags the non-compliant conversations ¶ [0098]. Applying compliance rules to identify violations ¶ [0120], [0122]. Grouping chat messages into subsets ¶ [0106], labeling based on contextual attributes ¶ [0175]), using the guardrail data structure, the chat data into one or more content clusters of the plurality of content clusters based on contextual data (Taylor: clustering/classification ¶ [0040], updating clusters ¶ [0043]).
Regarding claims 4, the combination of Taylor, D’Agostino, Hu, and Crabtree discloses, wherein the operations further comprise instructing the remote computing device to remove the flagged data from the chat data (D’Agostino: detects violations…flags the non-compliant conversations ¶ [0098]. Applying compliance rules to identify violations ¶ [0120]).
Regarding claims 5 and 10, the combination of Taylor, D’Agostino, Hu, and Crabtree discloses, wherein the operations further comprise verifying, by the LLM (Taylor: input…verify whether the user input meets predefined validation…LLM used to extract data and perform tasks related to validation ¶ [0045]. Rules and compliance triggers ¶ [0033]-[0034]), the interaction data using a verification process (D’Agostino: ensuring compliance with regulatory standards ¶ [0098], NLP and ML used to detect risks and identify compliance violations ¶ [0100]-[0101], verify user identity and validate transaction legitimacy ¶ [0123]).
Regarding claims 6 and 11, the combination of Taylor, D’Agostino, Hu, and Crabtree discloses, processing the chat data associated with the interaction data to generate a vector (D’Agostino: contextual attributes…labels to the vector…stored within vector database ¶ [0079]. Transforming these communications into vector representation using NLP techniques ¶ [0100]);
comparing the plurality of embeddings to the plurality of content clusters (D’Agostino: vectors are compared to regulatory requirements and industry standards ¶ [0120], analyze…to identify patterns….grouping/categorization via ML models); and
verifying the interaction data based on the comparing (D’Agostino: ensure compliance ¶ [0098], identifying compliance related to issues, assessing compliance status ¶ [0100]. Also see ¶ [0123]).
Regarding claims 7 and 12, the combination of Taylor, D’Agostino, Hu, and Crabtree discloses, wherein the operations further comprise generating a notification based on the update to the guardrail data structure (D’Agostino: generating reports, insights, alerts ¶ [0098]-[0102], outputs based on compliance analysis ¶ [0120]-[0124]).
Regarding claims 8, 16, and 20, the combination of Taylor, D’Agostino, Hu, and Crabtree discloses, processing the interaction data to generate training data, wherein the training data comprises examples of chat data correlated to examples of cluster memberships (D’Agostino: collecting chat conversation…analyze to extract actionable insights ¶ [0100], communications transformed into vector representation…used to identify deviations/compliance issues ¶ [0120]); and
training the LLM using the plurality of training data (D’Agostino: feedback loop to refine models ¶ [0098], generating training material from detected issues ¶ [0135]).
Regarding claim 13, (Canceled).
Regarding claim 14, (Canceled).
Regarding claim 15, the combination of Taylor, D’Agostino, Hu, and Crabtree discloses, wherein the method further comprises fine-tuning, using the LLM, one or more cluster memberships of one or more content clusters of the plurality of content clusters based on the expanded query dataset (D’Agostino: adjust parameters, weights, or architecture…fine-tuning existing parameters…updating weights ¶ [0162]. Assigning categories, compliance classes and risk levels ¶ [0100], [0120]).
Regarding claim 9, the combination of Taylor, D’Agostino, Hu, and Crabtree clearly show a method for performing the process for the method in claims 1, 2 and 3. Therefore, the rejections of claims 1, 2 and 3 applies to claim 9.
Regarding claim 17, the combination of Taylor, D’Agostino, Hu, and Crabtree clearly show a method for performing the process for the method in claims 1 and 8. Therefore, the rejections of claims 1 and 8 applies to claim 17.
Regarding claim 18, the combination of Taylor, D’Agostino, Hu, and Crabtree clearly show a method for performing the process for the method in claims 2 and 3. Therefore, the rejections of claims 2 and 3 applies to claim 18.
Conclusion
THIS ACTION IS MADE FINAL. Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to MOHAMMAD S ROSTAMI whose telephone number is (571)270-1980. The examiner can normally be reached Mon-Fri From 9 a.m. to 5 p.m..
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Boris Gorney can be reached at (571)270-5626. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
10/26/2026
/MOHAMMAD S ROSTAMI/Primary Examiner, Art Unit 2154