SYSTEM FOR INCREASING QUERY-ANSWERING AGENT ACCURACY

§101 §103 §112

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 . Continued Examination Under 37 CFR 1.114 A request for continued examination under 37 CFR 1.114, including the fee set forth in 37 CFR 1.17(e), was filed in this application after final rejection. Since this application is eligible for continued examination under 37 CFR 1.114, and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114. Applicant's submission filed on 08/28/2026 has been entered. Status of Claims Claims 1, 2, 4-12, and 14-20 are pending and currently rejected. Claims 3 and 13 have been canceled. Claims 1, 2, 4, 5, 9, 11, 12, 14, 15, 19, and 20 are currently amended. Claims 1, 11, and 20 are independent. Claims 1, 5, 15, and 20 are objected to. Priority Acknowledgment is made of applicant's claim for the benefit of U.S. Provisional Application No. 63/588,890, filed October 9, 2023 (35 U.S.C. 119(e)). For purposes of this action, the claims are given an effective filing date of October 9, 2023. No foreign priority is claimed. Information Disclosure Statement No Information Disclosure Statement has been filed. Amendment The present Office Action is based upon the original patent application filed on 10/09/2024 as modified by the amendment filed on 08/28/2026. Response to Amendment and Arguments Applicant's remarks (Remarks filed 08/28/2026) state that the amendments to the independent claims overcome the rejections under 35 U.S.C. 101 and 35 U.S.C. 103, without further argument. The amendments have been fully considered. The rejection of the claims under 35 U.S.C. 103 over Deal (US 2014/0101079 A1), Castel (US 2014/0164603 A1), Earle (US 2024/0111960 A1), and Johnson (US 2025/0045311 A1), etc… as set forth in the Final Office Action (Final Rejection dated 06/11/2026), is withdrawn and replaced by the new ground of rejection below. Regarding the rejection under 35 U.S.C. 103, applicant's arguments are moot because the amended claims are rejected below on a new ground of rejection that does not rely on any reference applied in the prior rejection for any teaching or matter specifically challenged in the argument. The new ground was necessitated by amendment and by the updated search conducted upon the filing of the request for continued examination. Regarding the rejection under 35 U.S.C. 101, the amendments have been considered but are not persuasive. The added limitations (storing summaries from one or more summarization agents in a response set, generating an aggregation prompt, identifying elements included in the summaries, and generating an answer based on those elements) are further steps of evaluating, comparing, and composing information. They add to the judicial exception rather than integrating it into a practical application. The rejection is maintained, with the full analysis set out below. Regarding the interview of 18 August 2026, the record reflects a discussion of amending the independent claims to incorporate the subject matter of dependent claims 2, 3, and 9. The filed amendment does not incorporate that subject matter in full. Specifically, the amended independent claims omit: (i) from claim 2, the requirement that the response set comprise "a predetermined minimum number of response summaries" (retained only in amended claims 2 and 12), and the requirement of a "plurality" of summarization, consultation, and evaluation agents (now "one or more"); (ii) from claim 3, the step of "retrieving, using an aggregation agent, the answer based on the response summaries", so that the aggregation prompt recited in the amended claims is never used; and (iii) from claim 9, the requirement that the identified elements be included in "a predetermined proportion of the response summaries" (retained only in amended claims 9 and 19) and that the answer be generated "to include the one or more identified elements" (now "based on" the elements). In addition, the updated search conducted in response to the request for continued examination located the references Du, Jiang, and Zheng, identified below, which were not of record at the time of the interview. As set out in the rejection of claims 2, 9, 12, and 19 below, those references teach the subject matter of claims 2 and 9, and of former claim 3, in the context of a multi-agent large language model system. Accordingly, incorporating claims 2, 3, and 9 in full into the independent claims would no longer, by itself, place the claims in condition for allowance (MPEP 713.04). The Suggestions for Overcoming the Art section below identifies disclosed subject matter that is not taught by the art of record. Claim Objections Claims 1, 5, 15, and 20 are objected to because of the following informalities: • Claim 1 recites "the responses evaluated by one or more evaluation agent". The term should read "one or more evaluation agents." • Claims 5 and 15 each recite "each-consultation agent". The hyphen should be deleted ("each consultation agent"). • Claim 20 recites "responsive to the query, , generate a summarization prompt". The duplicated comma should be deleted. Appropriate correction is required. Claim Interpretation 35 U.S.C. 112(f) The claims recite a "consultation agent", "evaluation agent", and "summarization agent" (and, in claims 4 and 14, an "aggregation agent"). The term "agent" is not modified by functional language that would cause it to stand in for a means-plus-function element. Each agent is recited as being "within at least one large language model", and the claimed operations ("retrieving, using a consultation agent ...") are performed by the recited processor executing instructions stored in memory (claim 1), by the computer-implemented method (claim 11), or by a computer executing stored instructions (claim 20). The specification uses "agent" for a large language model that receives a prompt and returns a response (Spec. ¶¶ [0028], [0030], [0031], [0033]). One of ordinary skill would understand the term as the name for a known class of software structure. The presumption that 35 U.S.C. 112(f) does not apply, which arises because the term "means" is not used, is not overcome (MPEP 2181(I)). The claims are therefore not interpreted under 35 U.S.C. 112(f). Intended-use and descriptive language Claims 1, 11, and 20 recite "the answer to be transmitted to a scheduling system for scheduling repairs to the industrial unit". The claims do not positively recite transmitting the answer or scheduling a repair. The phrase states an intended use of the answer and is given weight only to the extent that the answer must be in a form capable of being transmitted to a scheduling system (MPEP 2103(I)(C), 2111.04). The phrase "comprising a repair plan" is given weight as describing the content of the answer. The phrase "a query related to an industrial unit" is interpreted under its broadest reasonable interpretation in light of the specification. The specification describes, as examples of an industrial unit, a refrigeration unit and industrial equipment monitored by IoT devices, including mining equipment, water treatment facilities, power grid equipment, and construction equipment (Spec. ¶¶ [0013], [0014]). The term is not expressly defined. Under its broadest reasonable interpretation, an industrial unit includes an item of commercial building equipment, such as heating, ventilation, air-conditioning, or refrigeration equipment. Express definitions The specification states that the term "machine-readable medium" refers to a non-transitory medium and excludes signals per se (Spec. ¶ [0126]). Claim 20 is construed accordingly. No other claim term is expressly defined. Conditional language Claim 11 recites "based on determining that the evaluation of the response indicates the response is responsive to the query, generating a summarization prompt based on the response". Because the subsequent steps of retrieving a response summary, storing it, and generating the answer depend on the summarization prompt having been generated, the claim is construed as requiring that the condition be met and the summarization prompt be generated. The same construction is applied to claims 1 and 20 (MPEP 2111.04(II)). Claim Rejections - 35 U.S.C. 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, 2, 4-12, and 14-20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to a judicial exception (an abstract idea) without significantly more. The analysis follows the framework of Alice Corp. v. CLS Bank Int'l, 573 U.S. 208 (2014), and Mayo Collaborative Servs. v. Prometheus Labs., Inc., 566 U.S. 66 (2012), as set out in MPEP 2106. Step 1 - Statutory Category Claim 1 is directed to a system (a machine), claim 11 to a computer-implemented method (a process), and claim 20 to a non-transitory computer-readable storage medium (a manufacture). Each falls within a statutory category. Because claim 20 is limited to a non-transitory medium (Spec. ¶ [0126]), it does not encompass a signal per se. No claim is directed to software per se. (Step 1: Yes.) Step 2A, Prong One - Independent Claims 1, 11, and 20 Claim 1 recites the following limitations, which, apart from the generic computer and model elements addressed at Prong Two, describe an abstract idea: • "responsive to receiving a query related to an industrial unit, generating a consultation prompt based on the query"; • "generating, based on the response, an evaluation prompt"; • "based on determining that the evaluation of the response indicates the response is responsive to the query, generating a summarization prompt based on the response"; • "storing the response summary in a response set"; • "generating, based on the response summaries in the response set, an aggregation prompt"; • "identifying one or more elements included in the response summaries"; and • "generating an answer based on the response set, the response summaries and the one or more identified elements, the answer comprising a repair plan". These limitations describe writing a question about a piece of equipment, writing a request that someone review the answer, deciding whether the answer responds to the question, writing a request for a summary, collecting the summaries, writing a request that the summaries be combined, picking out the points the summaries have in common, and writing a repair plan from those points. Each is an observation, evaluation, judgment, or opinion that can practically be performed in the human mind or with pen and paper. A maintenance manager can draft a question about a malfunctioning refrigeration unit, read an answer and judge whether it addresses the question, condense the answer, compare several condensed answers for common recommendations, and write a repair plan from them. These limitations fall within the mental processes grouping of abstract ideas (MPEP 2106.04(a)(2)(III)). The same limitations, taken together, describe consulting several advisors on a question, having each advisor's answer reviewed by a second person, having a third person summarize the answers that pass review, and reaching a consensus recommendation. This is a method of managing interactions between people, including following rules or instructions, and falls within the certain methods of organizing human activity grouping (MPEP 2106.04(a)(2)(II)(C)). The steps of "retrieving, using a consultation agent within at least one large language model, a response", "retrieving, using an evaluation agent within the at least one large language model, an evaluation", and "retrieving, using a summarization agent within the at least one large language model, a response summary" are not themselves treated as mental steps, since a human cannot practically perform the operation of a large language model. They are treated as additional elements at Prong Two. The abstract idea is identified only from limitations that can practically be performed mentally or that describe the organization of human activity. Claims 11 and 20 recite the same limitations in method and storage-medium form. (Step 2A, Prong One: Yes, for claims 1, 11, and 20.) Step 2A, Prong Two - Independent Claims 1, 11, and 20 The additional elements of claim 1 are: (a) "at least one processor; and at least one memory storing instructions"; (b) the consultation, evaluation, and summarization agents "within at least one large language model", which are used to retrieve a response, an evaluation, and a summary; (c) the requirement that the evaluation agent and the summarization agent each be "distinct from the consultation agent"; and (d) the statement that the answer is "to be transmitted to a scheduling system for scheduling repairs to the industrial unit". Claim 11 recites elements (b)-(d) in a computer-implemented method and claim 20 recites them with a non-transitory computer-readable storage medium and a computer. Elements (a) and the computer and medium of claims 11 and 20 are generic computer components, recited at a high level of generality, that perform the abstract idea. They are mere instructions to apply the exception using a computer (MPEP 2106.05(f)). Elements (b) and (c) have been considered individually and in combination, without evaluating the claim at an overly high level of generality, as required by Ex parte Desjardins, Appeal No. 2024-000567 (ARP Sept. 26, 2025) (precedential), and MPEP 2106.04(d). The claims recite the large language model only by name and recite only its use: a prompt is supplied and an output (a response, an evaluation, or a summary) is retrieved. The claims do not recite how any model is trained, configured, or structured, or any change to the way a model operates. The requirement that the evaluating and summarizing agents be distinct from the answering agent is a choice of which tool performs each step. It does not change how any model operates. Using an existing model to perform the evaluating, summarizing, and answering steps that a person would otherwise perform is the use of a computer as a tool to perform the abstract idea (MPEP 2106.05(f)(2)). It also generally links the abstract idea to the technological environment of language models (MPEP 2106.05(h)). Element (d) is not positively recited (see Claim Interpretation). Even if transmitting the answer to a scheduling system were required, it would be insignificant extra-solution activity in the form of outputting a result (MPEP 2106.05(g)). The reference to an industrial unit limits the field of use only (MPEP 2106.05(h)). Technological-improvement check (MPEP 2106.04(d)(1), 2106.05(a)). The specification makes the following statements of improvement: • "Such strategies increase the overall accuracy of the query-answering system compared to the accuracy of sole query-answering agents, making the query-answering system suitable for inclusion in an autonomous response system." (Spec. ¶ [0014].) • "Thus, the query-answering system has to solve the problems of identifying and aggregating common operations or answer components across a set of responses with varying format, structure, and/or vocabulary. Examples in the disclosure herein refer to a query-answering system that uses summarization query-answering agents to summarize and/or normalize validated output of consultation query-answering agents." (Spec. ¶ [0015].) • "Examples in the disclosure herein refer to a query-answering system that defines a set of configurations, templates and parameters that can be stored, instantiated, and/or tuned as needed." (Spec. ¶ [0016].) • "By aggregating responses from multiple agents, the query-answering system increases overall response accuracy compared to individual agents. By incorporating intermediate automatic response validation and/or a configurable architecture, the query-answering system enables improved accuracy, consistency, and adaptability ..." (Spec. ¶ [0018].) The first and last statements describe more accurate answers obtained by consulting several sources, checking each answer, and taking the consensus. A more accurate answer is an improvement in the information produced by the abstract idea, not an improvement in the functioning of a computer or of the language models. An improvement in the judicial exception itself is not an improvement in technology (MPEP 2106.05(a)(II)). Unlike the claims in Desjardins, the claims do not recite any change to how a model is trained or operates; each model is used for its ordinary purpose of answering a prompt. The second statement (¶ [0015]) identifies a more specific problem: the responses of different agents have varying format, structure, and vocabulary, which makes common answer components hard to identify. The specification's solution is to have the summarization agents normalize each validated response, for example into "a templated and/or normalized output structure" and "using a vocabulary automatically aligned with, or extracted from, one or more domain-specific thesauri, domain-specific corpora, and/or other domain-specific artifacts" (Spec. ¶ [0031]), before an aggregation agent extracts the elements included in a predetermined number or proportion of the responses (Spec. ¶ [0033]). The independent claims do not recite that solution. They recite retrieving "a response summary" without any normalized structure or vocabulary, and they recite "identifying one or more elements included in the response summaries" without any aggregation agent or any criterion for which elements are identified. The third statement (¶ [0016]) concerns stored configurations and templates, which the independent claims also do not recite. Because the claims do not include the components or steps that provide the asserted improvement, the claims do not reflect it (MPEP 2106.04(d)(1); 2106.05(a)). Considered as an ordered combination, the additional elements amount to supplying prompts to off-the-shelf language models, in sequence, and storing their outputs. They do not provide a non-conventional arrangement of computer components. (Step 2A, Prong Two: No. Claims 1, 11, and 20 are directed to the abstract idea.) Step 2B - Independent Claims 1, 11, and 20 For the reasons given at Prong Two, the additional elements, alone and in combination, amount to mere instructions to apply the exception on a generic computer using generic language models, a field-of-use limitation, and, at most, insignificant extra-solution activity. Storing data in memory and transmitting data over a network are well-understood, routine, and conventional computer functions (MPEP 2106.05(d)(II), citing Versata Dev. Group, Inc. v. SAP Am., Inc., 793 F.3d 1306, 1334 (Fed. Cir. 2015), and buySAFE, Inc. v. Google, Inc., 765 F.3d 1350, 1355 (Fed. Cir. 2014)). The use of large language models to generate, assess, and summarize responses was itself conventional before the effective filing date, as shown by Du, Zheng, and Jiang (identified in the Prior Art of Record below). Du describes multiple copies of a large language model answering a question and summarizing one another's responses; Zheng describes using one large language model to grade the answers of others; and Jiang describes ensembling the outputs of multiple large language models (MPEP 2106.05(d)(I)(2)). The claims do not provide an inventive concept. (Step 2B: No.) Case Law Recentive Analytics, Inc. v. Fox Corp., No. 2023-2437 (Fed. Cir. Apr. 18, 2025), held that claims applying established machine learning methods to a new data environment, without disclosing improvements to the machine learning models themselves, are directed to an abstract idea. The present claims likewise use existing large language models, without modification, to answer, check, and summarize queries in the equipment-repair environment. Electric Power Group, LLC v. Alstom S.A., 830 F.3d 1350, 1353-54 (Fed. Cir. 2016), held that collecting information, analyzing it, and presenting the results is an abstract idea. Here, the claims collect responses, analyze them (by evaluation, summarization, and identification of common elements), and output an answer. Enfish, LLC v. Microsoft Corp., 822 F.3d 1327, 1335-36 (Fed. Cir. 2016), found eligible claims directed to a specific improvement in the way a computer stores and retrieves data (a self-referential table). Ex parte Desjardins applied Enfish to a claimed improvement in how a machine learning model is trained. The present claims do not recite an improvement to how the computer or any model functions; the models are used as given. Enfish and Desjardins are therefore not analogous. McRO, Inc. v. Bandai Namco Games Am. Inc., 837 F.3d 1299, 1313-15 (Fed. Cir. 2016), found eligible claims reciting specific rules that enabled a computer to perform a task previously performed by humans in a different way. The present claims do not recite specific rules. They recite results (an evaluation, a summary, identified elements, an answer) without saying how the elements are identified or how the answer is constructed from them. BASCOM Global Internet Servs., Inc. v. AT&T Mobility LLC, 827 F.3d 1341, 1350 (Fed. Cir. 2016), found an inventive concept in a non-conventional and non-generic arrangement of known components. The arrangement claimed here (answer, then review, then summarize, then combine) mirrors the conventional human process of consulting, reviewing, and reaching a consensus, and the art of record shows the same arrangement of language model agents (see the rejection under 35 U.S.C. 103). DDR Holdings, LLC v. Hotels.com, L.P., 773 F.3d 1245, 1257 (Fed. Cir. 2014), concerned a solution to a problem specifically arising in computer networks. The problem addressed here, obtaining a reliable answer from several fallible sources, is not specific to computer networks. Dependent Claims Each dependent claim has been considered individually. Claim 2 adds "the response set comprises a predetermined minimum number of response summaries". Requiring a minimum number of summaries is a further rule for the mental process of gathering enough opinions before reaching a consensus. It is part of the abstract idea. No additional element is added. Prong Two: No. Step 2B: No. Claim 4 adds "an agent of a plurality of agents is at least one of the consultation agent, the evaluation agent, the summarization agent, or an aggregation agent". This labels agents already recited and adds, at most, a generic aggregation agent, which is a further use of a language model as a tool (MPEP 2106.05(f)). Prong Two: No. Step 2B: No. Claim 5 adds that each consultation agent "is bound to a bound evaluation agent" and "to a bound summarization agent", each different from the consultation agent. Assigning each answering agent its own reviewer and summarizer, neither of which is the answering agent itself, is a rule for organizing the review (a person does not review or summarize his own answer). It is part of the abstract idea. To the extent the binding is an additional element, it is a configuration choice about which generic model performs which step and does not change how any model operates (MPEP 2106.05(f)). Prong Two: No. Step 2B: No. This claim is discussed further under Closest to Eligibility. Claim 6 adds that the consultation, evaluation, and summarization prompts are generated using respective "prompt template[s]". Filling in a form question is a mental process and part of the abstract idea. Storing templates is generic data storage (MPEP 2106.05(d)(II)). Prong Two: No. Step 2B: No. Claim 7 adds "storing an entry corresponding to the consultation agent in the response set, the entry comprising one or more of an identifier of the consultation agent, a response status, the response, and the response summary". Recording who answered, the status of the answer, and the answer is record-keeping that can be done with pen and paper and is part of the abstract idea. Storing the entry in memory is a generic computer function (MPEP 2106.05(d)(II)). Prong Two: No. Step 2B: No. Claim 8 adds that "the response status is at least one of at least pending, received, responsive, non-responsive, or timed out". These are labels for the state of a request and a judgment about the answer, which are part of the abstract idea. Prong Two: No. Step 2B: No. Claim 9 adds that "the one or more elements are included in a predetermined proportion of the response summaries". Identifying the points on which a set proportion of advisors agree is a mental process (a vote count) and part of the abstract idea. Prong Two: No. Step 2B: No. Claim 10 adds "waiting for the response for a predetermined maximum time period". Setting a deadline for an advisor's answer is part of organizing the consultation. To the extent it is an additional element, waiting for a response with a time-out is a generic computer function (MPEP 2106.05(d)(II)). Prong Two: No. Step 2B: No. Claim 12 adds the subject matter of claim 2 to the method of claim 11 and is ineligible for the reasons given for claim 2. Claim 14 adds the subject matter of claim 4 to the method of claim 11 and is ineligible for the reasons given for claim 4. Claim 15 adds the subject matter of claim 5 to the method of claim 11 and is ineligible for the reasons given for claim 5. Claim 16 adds the subject matter of claim 6 to the method of claim 11 and is ineligible for the reasons given for claim 6. Claim 17 adds the subject matter of claim 7 to the method of claim 11 and is ineligible for the reasons given for claim 7. Claim 18 adds the subject matter of claim 8 to the method of claim 17 and is ineligible for the reasons given for claim 8. Claim 19 adds the subject matter of claim 9 to the method of claim 11 and is ineligible for the reasons given for claim 9. Closest to Eligibility Claims 5 and 15 come closest to integrating the exception into a practical application, because they specify a configuration of separate agents. They do not cross the line. The binding of each consultation agent to a different evaluation agent and a different summarization agent decides which generic model performs each step, but does not change how any model processes data. Claims 2, 9, 12, and 19, taken together with claim 4 or 14, would recite most of the aggregation workflow of the specification, but they still do not recite the normalization of responses that the specification identifies as the solution to the problem stated in ¶ [0015]. Suggested Amendment (35 U.S.C. 101) Applicant may consider amending the independent claims to recite the specific technical solution described in Spec. ¶¶ [0015] and [0031]-[0033], for example: (i) that the summarization prompt instructs the summarization agent to output each response summary in a predetermined normalized output structure, using a vocabulary drawn from a domain-specific thesaurus or corpus supplied as input (¶ [0031]); (ii) that, after the evaluated responses are summarized, the system determines whether the response set meets a consultation response quorum of at least a predetermined number of responses with associated summaries, and terminates the workflow and returns a predetermined custom response when the quorum is not met (¶¶ [0029], [0032]); and (iii) that the aggregation prompt is submitted to an aggregation agent that extracts the elements included in a predetermined proportion of the normalized summaries, the answer being generated to include those elements (¶¶ [0033], [0071]). Such an amendment would be evaluated under Step 2A, Prong Two, to determine whether the claim, as a whole, reflects an improvement in the operation of the multi-agent system. Any amendment must be supported by the original disclosure. Claim Rejections - 35 U.S.C. 112 The following is a quotation of 35 U.S.C. 112(b): (b) CONCLUSION.-The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention. Claims 1, 2, 4-12, and 14-20 are rejected under 35 U.S.C. 112(b) as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor regards as the invention. (1) Antecedent basis - claims 1, 11, and 20 Claim 1 recites "a response summary" (singular) and then "storing the response summary in a response set generated by one or more summarization agents, the response summaries corresponding to responses ...". There is insufficient antecedent basis for "the response summaries" in this limitation. Claim 1 later recites "the responses evaluated by one or more evaluation agent"; claim 1 previously recites only "a response", and there is insufficient antecedent basis for "the responses". Claims 11 and 20 contain the same language and are indefinite for the same reasons. In addition, claims 11 and 20 recite "a response set generated by the summarization agents", and there is insufficient antecedent basis for "the summarization agents", because claims 11 and 20 previously recite only "a summarization agent". It is also unclear how the recited "consultation agent", "evaluation agent", and "summarization agent" relate to the "one or more consultation agents", "one or more evaluation agent", and "one or more summarization agents" of claim 1. A person of ordinary skill could not determine whether the singular agents are among the "one or more" agents, or whether the claim requires a separate group of agents in addition to them. Claims 2, 4-10, 12, and 14-19 depend from claim 1 or claim 11 and are rejected for the same reasons. Suggested fix: introduce the plural terms expressly and tie the singular agents to them, for example: "storing the response summary in a response set comprising a plurality of response summaries generated by a plurality of summarization agents including the summarization agent, the plurality of response summaries corresponding to a plurality of responses generated by a plurality of consultation agents including the consultation agent, the plurality of responses having been evaluated by a plurality of evaluation agents including the evaluation agent." (2) Gap between steps - claims 1, 11, and 20 Claims 1, 11, and 20 recite "generating, based on the response summaries in the response set, an aggregation prompt", and then "generating an answer based on the response set, the response summaries and the one or more identified elements". The claims do not recite any use of the aggregation prompt: the prompt is not submitted to any agent or model, and the answer is not generated from it. The specification describes the aggregation prompt only as an input that is submitted to an aggregation agent, which returns the answer (Spec. ¶¶ [0033], [0050], [0071]; FIG. 5, operations 504-508). The claims also do not state what identifies the "one or more elements" or how the answer is generated. Because the claims omit the cooperative relationship between the aggregation prompt and the generation of the answer that the specification describes, there is a gap between the steps, and the metes and bounds of the claims are unclear (MPEP 2172.01). A reader cannot tell whether the claim requires the answer to be produced by a language model from the aggregation prompt, or whether the aggregation prompt is a free-standing data item with no role in the claimed method. Claims 2, 4-10, 12, and 14-19 are rejected for the same reason. Suggested fix: recite the step from former claim 3, for example: "submitting the aggregation prompt to an aggregation agent within the at least one large language model; and retrieving, using the aggregation agent, the answer, the aggregation agent identifying one or more elements included in a predetermined proportion of the response summaries and generating the answer to include the one or more identified elements" (Spec. ¶¶ [0033], [0071]). (3) Claim 14 Claim 14 recites "The method of claim 11, generating the answer further comprising: an agent of a plurality of agents is at least one of the consultation agent, the evaluation agent, the summarization agent, or an aggregation agent." The claim purports to add a step to "generating the answer", but the text that follows is a statement of what an agent is, not a step. It is unclear whether claim 14 requires the answer to be generated by an agent, requires an aggregation agent, or merely labels agents already recited. Suggested fix: conform claim 14 to claim 4 ("The method of claim 11, wherein ..."), or recite the intended step, for example "wherein generating the answer comprises retrieving the answer using an aggregation agent." (4) Claim 15 Claim 15 recites "the one or more consultation agents", "the one or more evaluation agents", and "the one or more summarization agents". Claim 11 does not recite "one or more consultation agents", "one or more evaluation agents", or "one or more summarization agents" (it recites "the summarization agents", "the consultation agent", and "the evaluation agent"). There is insufficient antecedent basis for these terms in the claim. Suggested fix: adopt the plural terms in claim 11 as suggested under item (1). Record of close calls (35 U.S.C. 112(b)) The terms "predetermined minimum number" (claims 2 and 12), "predetermined proportion" (claims 9 and 19), and "predetermined maximum time period" (claim 10) are not indefinite. Each refers to a value fixed in advance, and the specification gives examples (Spec. ¶¶ [0029], [0032], [0033]). The phrase "responsive to the query" is not indefinite; the specification explains that the evaluation agent determines whether the response "directly addresses the query" (Spec. ¶ [0031]). Claims 6-8 and 16-18 raise no 35 U.S.C. 112(b) issue apart from their dependency. Written description and enablement (35 U.S.C. 112(a)) No rejection under 35 U.S.C. 112(a) is made. The amended limitations are supported by the specification as originally filed (Spec. ¶¶ [0029]-[0033], [0050], [0063], [0071]; original claims 2, 3, and 9), and the specification describes the claimed agents and prompts in sufficient detail to enable one of ordinary skill to make and use the claimed invention. 35 U.S.C. 112(d) The following is a quotation of 35 U.S.C. 112(d): (d) REFERENCE IN DEPENDENT FORMS.-Subject to subsection (e), a claim in dependent form shall contain a reference to a claim previously set forth and then specify a further limitation of the subject matter claimed. A claim in dependent form shall be construed to incorporate by reference all the limitations of the claim to which it refers. Claims 4 and 14 are rejected under 35 U.S.C. 112(d) as being of improper dependent form for failing to further limit the subject matter of the claim upon which they depend. Claim 4 recites "an agent of a plurality of agents is at least one of the consultation agent, the evaluation agent, the summarization agent, or an aggregation agent." Claim 1 already recites a consultation agent, an evaluation agent distinct from it, and a summarization agent distinct from it, which together are a plurality of agents. Each of those agents is "at least one of the consultation agent, the evaluation agent, the summarization agent, or an aggregation agent." Claim 4 therefore reads on every system covered by claim 1 and does not narrow it. Claim 14 is the same with respect to claim 11. Applicant may cancel the claims, amend them to place them in proper dependent form, rewrite them in independent form, or present a sufficient showing that the dependent claims comply with the statutory requirements. A suggested fix that would add a limitation is "wherein generating the answer comprises submitting the aggregation prompt to an aggregation agent within the at least one large language model, the aggregation agent being distinct from the consultation agent." Claim Rejections - 35 USC § 103 In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status. 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 of this title, 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. The factual inquiries set forth in Graham v. John Deere Co., 383 U.S. 1, 148 USPQ 459 (1966), that are applied for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows: 1. Determining the scope and contents of the prior art. 2. Ascertaining the differences between the prior art and the claims at issue. 3. Resolving the level of ordinary skill in the pertinent art. 4. Considering objective evidence present in the application indicating obviousness or nonobviousness. This application currently names joint inventors. In considering patentability of the claims the examiner presumes that the subject matter of the various claims was commonly owned as of the effective filing date of the claimed invention(s) absent any evidence to the contrary. Applicant is advised of the obligation under 37 CFR 1.56 to point out the inventor and effective filing dates of each claim that was not commonly owned as of the effective filing date of the later invention in order for the examiner to consider the applicability of 35 U.S.C. 102(b)(2)(C) for any potential 35 U.S.C. 102(a)(2) prior art against the later invention. Travalini - US 2023/0259821 A1, published August 17, 2023; Application No. 18/105,424, filed February 3, 2023; first-named inventor Michael Travalini; applicant Mezo Inc. Prior art under 35 U.S.C. 102(a)(1). Travalini describes a computing platform that receives a maintenance request from a user describing a problem with an item, including commercial and industrial equipment and HVAC equipment, uses at least one large language model (e.g., GPT) to parse the request, diagnoses the problem, generates a work order identifying the parts, tools, and skill level for the repair, and transmits the work order to technicians selected according to their schedules and availability. Travalini also obtains technician feedback confirming whether the diagnosis was correct. It is relevant to the claimed query about an industrial unit, the consultation agent within a large language model, the repair plan, and the scheduling of repairs. Du - Yilun Du, Shuang Li, Antonio Torralba, Joshua B. Tenenbaum, and Igor Mordatch, "Improving Factuality and Reasoning in Language Models through Multiagent Debate," arXiv:2305.14325v1, submitted May 23, 2023. Prior art under 35 U.S.C. 102(a)(1). Du describes multiple large language model agents, including different models (chatGPT and Bard), that each answer the same question, summarizes the agents' responses with a language model, and arrives at a final answer by majority vote or by converging on a shared answer. It is relevant to the claimed plurality of consultation agents, the summarization agent, the response set, and the identification of common elements. Zheng - Lianmin Zheng et al., "Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena," arXiv:2306.05685v1, submitted June 9, 2023. Prior art under 35 U.S.C. 102(a)(1). Zheng describes using a strong large language model (GPT-4) as a judge that evaluates the answers of other language model assistants, reports that such judges agree with human evaluators at the level humans agree with one another, and reports that a judge favors its own answers (self-enhancement bias). It is relevant to the claimed evaluation agent distinct from the consultation agent. Jiang - Dongfu Jiang, Xiang Ren, and Bill Yuchen Lin, "LLM-Blender: Ensembling Large Language Models with Pairwise Ranking and Generative Fusion," arXiv:2306.02561v1, submitted June 5, 2023 (also published in Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics, July 2023). Prior art under 35 U.S.C. 102(a)(1). Jiang describes collecting candidate outputs from multiple large language models for one input, ranking the candidates, and passing only a predetermined number (K = 3) of the top-ranked candidates to a generative fusion language model that produces the final output. It is relevant to forwarding only favorably evaluated responses, to a predetermined number of responses, and to generating an answer from the selected responses. Wu - Qingyun Wu et al., "AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation," arXiv:2308.08155v1, submitted August 16, 2023. Prior art under 35 U.S.C. 102(a)(1). Wu describes a multi-agent framework in which large language model agents with different roles generate output, a separate agent checks that output, and the process repeats until the user's question is answered or the process times out. It is relevant to claim 10. Designation of Primary and Secondary References; Level of Ordinary Skill Travalini is the Primary reference. On its own, Travalini teaches receiving a maintenance request describing a problem with an item of equipment, supplying the request to a large language model to parse it, diagnosing the problem, generating a work order that sets out the repair (a repair plan), and transmitting the work order to technicians selected according to their schedules and availability. Travalini uses one language model pipeline for each request and checks its diagnosis only through later human feedback from the technician. The Secondary references fill the following gaps. Zheng supplies the evaluation agent: a separate large language model, prompted with the response, that evaluates another model's response. Du supplies multiple consultation agents answering the same query, a language model that summarizes the responses, a set of responses and summaries, a prompt that combines the summaries, and identification of the answer common to the agents (majority vote or convergence). Jiang supplies forwarding only favorably evaluated responses, a predetermined number of them, to a further language model that generates the final answer. Wu supplies the time-out of claim 10 and the time-out status of claims 8 and 18. A person of ordinary skill in the art at the time of the effective filing date would have had at least a bachelor's degree in computer science, computer engineering, or a related field and about two years of experience building natural language processing or machine learning applications, including applications that prompt pretrained large language models through an application programming interface. The applied references themselves reflect this level of skill (Okajima v. Bourdeau, 261 F.3d 1350, 1355 (Fed. Cir. 2001); MPEP 2141(II)(C)). Du, Zheng, Jiang, and Wu are in the same field of endeavor as the claimed invention, which the specification identifies as "query-answering systems" and "solutions for increasing the accuracy of a query-answering system" (Spec. ¶ [0002]). Each addresses the accuracy or reliability of answers produced by large language models. Travalini is in the same field of endeavor (answering maintenance-repair requests with a language-model-driven chatbot) and is also reasonably pertinent to the problem of producing a repair plan for an industrial unit that is sent to a maintenance scheduling system (Spec. ¶ [0013]) (MPEP 2141.01(a)). Rejection Statements Claims 1, 2, 4-7, 9, 11, 12, 14-17, 19, and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Travalini (US 2023/0259821 A1) in view of Zheng (arXiv:2306.05685), Du (arXiv:2305.14325), and Jiang (arXiv:2306.02561). Claim 1 Claim 1 recites "A system comprising: at least one processor; and at least one memory storing instructions that, when executed by the at least one processor, configure the system to perform operations comprising". Travalini teaches this feature. Travalini's computer server 101 "may have a processor 103 for controlling overall operation of the server 101 and its associated components, including RAM 105, ROM 107, input/output (I/O) module 109, and memory 115" (Travalini, ¶ [0033]), and "Software may be stored within memory 115 to provide instructions to processor 103 for enabling server 101 to perform various functions" (Travalini, ¶ [0034]). Claim 1 recites "responsive to receiving a query related to an industrial unit, generating a consultation prompt based on the query". Travalini teaches this feature. In Travalini, "a computing platform may receive a maintenance request from a user and may configure a chatbot to extract details that describe an item to be repaired" (Travalini, Abstract), and the user describes the problem in response to the chatbot's inquiry (e.g., "My toilet is clogged") (Travalini, ¶ [0049]). Travalini's problem model "may parse the one or more inputs (e.g., the user statement and/or description of the problem, or the like) using at least one NLP algorithm" and "may use at least one large language model (LLM) (e.g., BERT, GPT, Dialogflow, or the like)" for that purpose, and each LLM may be tuned to identify problems including "commercial equipment and/or machinery repairs, industrial equipment and/or machinery repairs" and "HVAC" systems (Travalini, ¶ [0129]). A maintenance request describing a problem with industrial or commercial equipment, such as an HVAC unit, is a query related to an industrial unit (see Claim Interpretation), and the user statement supplied to the large language model is a consultation prompt generated based on the query. Claim 1 recites "retrieving, using a consultation agent within at least one large language model, a response to the query". Travalini teaches this feature. Travalini's large language model receives the user's description of the problem and is used "to classify the data within the one or more inputs, to generate a vectorization structure that may be used to progress through the node graph, to perform entity recognition, or the like" (Travalini, ¶ [0129]), and the result drives the diagnosis of the problem (Travalini, ¶¶ [0061]-[0062]). The large language model that returns the classification of the user's problem is a consultation agent within at least one large language model, and its output is a response to the query. Claim 1 recites "generating, based on the response, an evaluation prompt; retrieving, using an evaluation agent within the at least one large language model, an evaluation of the response, wherein the evaluation agent is distinct from the consultation agent". Primary prior-art reference Travalini does not explicitly teach these features. Travalini evaluates its diagnosis only after the fact, through human feedback: "the technician may confirm whether the diagnosis indicated on the work order (or the enriched work order) corresponds to the technician's diagnosis upon on-site inspection" (Travalini, ¶ [0104]), and its diagnosis engine may "compare the generated prediction to that of the administrator/supervisor" (Travalini, ¶ [0066]). However, secondary prior-art reference Zheng teaches these features as follows. Zheng proposes to "explore the use of state-of-the-art LLMs, such as GPT-4, as a surrogate for humans" in evaluating the answers of chat assistants (Zheng, § 1), and in single-answer grading "an LLM judge is asked to directly assign a score to a single answer," using the prompt shown in Zheng's Figure 5 (Zheng, § 3.1). Zheng's judge receives a prompt containing the question and the assistant's answer and returns an assessment of that answer. That prompt is an evaluation prompt generated based on the response, and the judge model is an evaluation agent within a large language model that returns an evaluation of the response. Zheng's judge (GPT-4) is a different model from the assistants whose answers it grades, and Zheng reports that a judge favors its own answers: "GPT-4 favors itself with a 10% higher win rate; Claude-v1 favors itself with a 25% higher win rate" (Zheng, § 3.3). Before the effective filing date of the claimed invention, it would have been obvious to one of ordinary skill in the art to modify Travalini's checking of the language model's diagnosis against a human reviewer (Travalini, ¶¶ [0066], [0104]) to submit the response, in an evaluation prompt, to a second large language model judge that is distinct from the answering model, as taught by Zheng, so that the system generates an evaluation prompt based on the response and retrieves an evaluation of the response from an evaluation agent distinct from the consultation agent. The reason is that Zheng identifies an LLM judge as "a scalable and explainable way to approximate human preferences" (Zheng, Abstract) that reaches "over 80% agreement, the same level of agreement between humans" (Zheng, Abstract), which is a reason to use it in place of, or before, Travalini's human review, so that an incorrect diagnosis is caught before a work order is dispatched rather than after a technician's site visit. Zheng's finding of self-enhancement bias (Zheng, § 3.3) is a reason to use a judge distinct from the model that produced the answer. One of ordinary skill would have had a reasonable expectation of success because Travalini's language model already produces its output as data that can be placed in a prompt, and Zheng's judge operates on such output through an ordinary prompt. The rationale is MPEP 2143(I)(B) (simple substitution) and (G): Travalini's human review of the model's diagnosis and Zheng's LLM judge are known elements that perform the same function of evaluating a model's output, the substitution was within the skill of the art, and Zheng supplies the reason to make it. The modification does not change Travalini's principle of operation, which remains a language-model-driven chatbot that diagnoses a maintenance request and generates a work order, and the predictable result is an automated check of the diagnosis. KSR Int'l Co. v. Teleflex Inc., 550 U.S. 398, 416 (2007). This reason to combine comes from Zheng itself, not from applicant's disclosure. Claim 1 recites "based on determining that the evaluation of the response indicates the response is responsive to the query, generating a summarization prompt based on the response; retrieving, using a summarization agent within the at least one large language model, a response summary, wherein the summarization agent is distinct from the consultation agent". The combination of Travalini and Zheng does not explicitly teach these features. However, secondary prior-art reference Du teaches retrieving a response summary using a summarization agent within a large language model, as follows. In Du, "Given a question, multiple agents represented as copies of a large language model, generate answers to the question" (Du, § 2.1), and Du states: "We may alternatively first summarize the responses from all other agents into a single response" (Du, § 3.3). Du explains that, "For larger number of agents, we first summarize all agent responses with chatGPT instead of directly concatenating responses" (Du, § 3.3), and reports that "We find this improves the performance of debate" (Du, § 3.3). Du's instruction to chatGPT to summarize the agents' responses is a summarization prompt generated based on the responses, and the summary returned by chatGPT is a response summary retrieved using a summarization agent within a large language model. The summarizing invocation is separate from the agents that generated the responses, so the summarization agent is distinct from the consultation agent. Further, secondary prior-art reference Jiang teaches conditioning the later stage on the evaluation: "The PairRanker module learns to compare all pairs of candidates for each input and subsequently rank the list of candidates. We then select the top K = 3 ranked candidates," which are passed to the GenFuser module (Jiang, § 2.3). Jiang therefore forwards a candidate to the next stage only if the evaluation of that candidate is favorable. Before the effective filing date of the claimed invention, it would have been obvious to one of ordinary skill in the art to modify the combination of Travalini and Zheng to (i) submit each response that the evaluation agent judges favorably to a separate language model with an instruction to summarize it, as taught by Du, and (ii) forward to that summarization step only the responses that the evaluation indicates are responsive, as taught by Jiang, so that, based on determining that the evaluation indicates the response is responsive to the query, the system generates a summarization prompt based on the response and retrieves a response summary from a summarization agent distinct from the consultation agent. The reasons are that Du reports that summarizing agent responses "improves the performance of debate" (Du, § 3.3), and that Jiang teaches that merging only the top-ranked candidates makes it "plausible to generate a superior response by combining their advantages while mitigating their shortcomings" (Jiang, § 4). Discarding responses that a judge has found deficient before further processing is the express purpose of Jiang's ranking stage. One of ordinary skill would have had a reasonable expectation of success because each step uses a language model for its established function (answering, judging, summarizing) and passes text from one step to the next. The rationale is MPEP 2143(I)(A) and (G): the prior art included each element (answering model, judging model, summarizing model, and selection of favorably evaluated outputs), one of ordinary skill could have combined them by known methods, each element performs the same function in combination as it does separately, and Du and Jiang supply the reason to combine them. The modification does not change the principle of operation of Travalini, and the predictable result is a summary produced only for answers that pass evaluation. KSR, 550 U.S. at 416. This reason to combine comes from the references themselves, not from applicant's disclosure. Claim 1 recites "storing the response summary in a response set generated by one or more summarization agents, the response summaries corresponding to responses generated by one or more consultation agents, the responses evaluated by one or more evaluation agent". The combination of Travalini, Zheng, Du, and Jiang teaches this feature, for the reasons given in the preceding paragraphs. Travalini's memory 115 stores "an associated database 121" (Travalini, ¶ [0034]). Du's agents are "multiple agents" (Du, § 2.1) that may be different models, since Du states that "we also investigate the case of mixing different model types, such as chatGPT [21] and Bard [23]" (Du, § 1). Du's summarized responses are "given as context" to the agents in later prompts (Du, § 3.3, Fig. 13), which ask each agent, "Using the opinion of other agents as additional advice, can you give an updated response" (Du, Fig. 3). Supplying the summaries in a later prompt requires that they be stored. In the combination, the stored summaries form a response set generated by one or more summarization agents, corresponding to responses generated by one or more consultation agents, which have been evaluated by one or more evaluation agents (Zheng; Jiang). No further modification is needed. Claim 1 recites "generating, based on the response summaries in the response set, an aggregation prompt". The combination of Travalini, Zheng, Du, and Jiang teaches this feature. Du's prompt that supplies the summarized responses of the other agents and asks for an updated response (Du, § 3.3; Fig. 3) is a prompt generated based on the response summaries. Jiang likewise selects the top-ranked candidates, "concatenate[s] them with the input x, and construct[s] the input sequence for the GenFuser module" (Jiang, § 2.3). The prompt that supplies the summaries for merging is an aggregation prompt. This is the same combination as above, for the reasons given above. Claim 1 recites "identifying one or more elements included in the response summaries". The combination teaches this feature. Du determines the final answer by "performing majority voting" across the agents' answers (Du, § 3.1; Table 1, "Multi-Agent (Majority)"), and Du's agents "are able to converge on a single shared answer after multiple rounds of debate" (Du, § 2.2). Identifying the answer that a majority of the agents' responses share, or the answer shared by all of them, is identifying one or more elements included in the responses. In the combination, those responses are represented by their summaries (Du, § 3.3). This is the same combination, for the reasons given above. Claim 1 recites "generating an answer based on the response set, the response summaries and the one or more identified elements, the answer comprising a repair plan, the answer to be transmitted to a scheduling system for scheduling repairs to the industrial unit". The combination teaches this feature. Du's final answer is the answer that the agents' responses share (Du, §§ 2.2, 3.1), and "The GenFuser module, a seq2seq LM, ultimately generates the final output to serve users" (Jiang, § 2.3). In Travalini, "Based on the diagnosis information, the Recommendation Engine may generate a work order. The work order may identify at least one of the item, the component, the location, the prime symptom, tools/parts required for the repair, a skill level for the repair, and/or a number of trips for the repair" (Travalini, ¶ [0062]), and the enriched work order "may indicate the diagnosis, parts and supplies needed to complete the job, an expected time needed to complete the job" (Travalini, ¶ [0064]). That work order is an answer comprising a repair plan. Travalini sends the work order to a resolution/fulfillment component in which "The technician availability module may be configured to determine a technician's schedule" (Travalini, ¶ [0064]); the chatbot "may offer to schedule a technician visit for the user" (Travalini, ¶ [0045]); and the intelligent routing and assignment engine "may transmit the work order (or enriched work order) to the technicians selected by the regression model" (Travalini, ¶ [0104]). The work order is thus transmitted to a scheduling system for scheduling repairs to the equipment. In the combination, the repair answer of Travalini is the answer generated from the response set, summaries, and identified elements, as taught by Du and Jiang. The reason to apply the combined multi-agent technique to Travalini's repair answer is the one Du states: the multi-agent approach "outperforms single model baselines" (Du, § 1) and improves "the factual validity of generated content, reducing fallacious answers and hallucinations that contemporary models are prone to" (Du, Abstract), which is a reason to use it wherever, as in Travalini, an incorrect model output would lead to an incorrect service action (see Travalini, ¶ [0104], in which a technician discovers on site whether the diagnosis was wrong). This is the same combination, and the same rationale (MPEP 2143(I)(A), (G)), as set out above. Claims 11 and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Travalini in view of Zheng, Du, and Jiang. Claims 11 and 20 Claim 11 recites "A computer-implemented method comprising" the same operations as claim 1, and claim 20 recites "A non-transitory computer-readable storage medium, the computer-readable storage medium including instructions that when executed by a computer, cause the computer to" perform the same operations. Claims 11 and 20 are rejected for the reasons given for claim 1, which are incorporated here. Travalini's methods are performed by a computer server whose processor executes software "stored within memory 115" (Travalini, ¶ [0034]), and Travalini claims "a computing device comprising a processor and a nontransitory memory device storing instructions" (Travalini, claim 1). Travalini also states that "program modules may be located in both local and remote computer storage media including memory storage devices" (Travalini, ¶ [0039]). Claims 11 and 20 recite "a response set generated by the summarization agents, the response summaries corresponding to responses generated by the consultation agent, the responses evaluated by the evaluation agent"; this language is met by the same teachings of Du, Zheng, and Jiang discussed for claim 1. Claims 2 and 12 are rejected under 35 U.S.C. 103 as being unpatentable over Travalini in view of Zheng, Du, and Jiang. Claims 2 and 12 Claim 2 recites "wherein: the response set comprises a predetermined minimum number of response summaries"; claim 12 recites the same for the method of claim 11. The combination teaches this feature. Du runs its multiagent procedure with a set number of agents, for example "fixing the number of agents to three" (Du, § 3.3, "Rounds of Debate"), and studies the effect of that number (Du, § 3.3, "Number of Agents"). Jiang likewise selects a predetermined number of responses: "We then select the top K = 3 ranked candidates" (Jiang, § 2.3). Each run of Du therefore produces responses from a number of agents chosen in advance, and in the combination the response set holds the summaries of those responses. A response set holding the summaries for a number of agents fixed before the query is processed is a response set that comprises a predetermined minimum number of response summaries. This is the same combination as claim 1, for the reasons given for claim 1. Du's study of performance as a function of the number of agents (Du, § 3.3) is a further reason for one of ordinary skill to fix that number in advance. Claims 4 and 14 are rejected under 35 U.S.C. 103 as being unpatentable over Travalini in view of Zheng, Du, and Jiang. Claims 4 and 14 Claim 4 recites "wherein: an agent of a plurality of agents is at least one of the consultation agent, the evaluation agent, the summarization agent, or an aggregation agent"; claim 14 recites the same language for the method of claim 11. The combination teaches this feature. The combination of claim 1 includes a plurality of agents: Travalini's large language model (the consultation agent), Zheng's judge (the evaluation agent), and Du's summarizing model (the summarization agent). Jiang's GenFuser, a language model that merges the selected candidates into the final output (Jiang, Abstract), is an aggregation agent. This is the same combination, for the reasons given for claim 1. Claims 5 and 15 are rejected under 35 U.S.C. 103 as being unpatentable over Travalini in view of Zheng, Du, and Jiang. Claims 5 and 15 Claim 5 recites "each-consultation agent of the one or more consultation agents is bound to a bound evaluation agent of the one or more evaluation agents, the respective consultation agent being different from the bound evaluation agent; and each consultation agent of the one or more consultation agents is bound to a bound summarization agent of the one or more summarization agents, the respective consultation agent being different from the bound summarization agent"; claim 15 recites the same language for the method of claim 11. The combination teaches this feature. In the combination of claim 1, every consultation agent's response is evaluated by Zheng's judge, which is a different model from the assistants it grades (Zheng, § 1), and every response that passes is summarized by a separate summarizing invocation (Du, § 3.3). Assigning each consultation agent a specific evaluation agent and a specific summarization agent is the routing that the combination requires in order to evaluate and summarize each response. Zheng's report that a judge favors its own answers (Zheng, § 3.3) is the reason to ensure that the evaluation agent bound to a consultation agent is a different agent. The same reason applies to the summarization agent, since a summary is a further assessment of what the response says. Before the effective filing date of the claimed invention, it would have been obvious to one of ordinary skill in the art to configure the combination of Travalini, Zheng, Du, and Jiang so that each consultation agent is bound to an evaluation agent and a summarization agent other than itself, so that each response is assessed and summarized by an agent that did not produce it, because Zheng shows that a model assessing its own output favors that output (Zheng, § 3.3). A reasonable expectation of success exists because Du already runs different models (chatGPT and Bard) side by side (Du, § 3.3) and Zheng already uses a judge separate from the graded assistants. The rationale is MPEP 2143(I)(G), with the motivation drawn from Zheng. The predictable result is that no agent evaluates or summarizes its own response. KSR, 550 U.S. at 416. Claims 6 and 16 are rejected under 35 U.S.C. 103 as being unpatentable over Travalini in view of Zheng, Du, and Jiang. Claims 6 and 16 Claim 6 recites "generating the consultation prompt further comprises the use of a consultation prompt template; generating the evaluation prompt further comprises the use of an evaluation prompt template; and generating the summarization prompt further comprises the use of a summarization prompt template"; claim 16 recites the same for the method of claim 11. The combination teaches this feature. Du's method "uses identical procedure and prompts for all tasks we investigate" (Du, Abstract), and in Du's debate prompts the "Responses of other agents to questions are inserted in the middle of the prompt (indicated with [other answers])" (Du, Fig. 3). A fixed prompt reused across queries, with the variable content inserted, is a prompt template. Zheng's judge likewise grades each answer using a fixed judge prompt ("The prompt used for this scenario is in Figure 5 (Appendix)") (Zheng, § 3.1). Before the effective filing date of the claimed invention, it would have been obvious to one of ordinary skill in the art to generate each of the consultation, evaluation, and summarization prompts of the combination of Travalini, Zheng, Du, and Jiang from a respective fixed template into which the query or the response is inserted, as Du and Zheng do for their prompts, because Du shows that one fixed procedure and set of prompts can be reused across all tasks (Du, Abstract). The rationale is MPEP 2143(I)(C): use of a known technique (fixed prompts with inserted content) to improve similar methods in the same way. The predictable result is consistent prompts across queries. KSR, 550 U.S. at 416. Claims 7 and 17 are rejected under 35 U.S.C. 103 as being unpatentable over Travalini in view of Zheng, Du, and Jiang. Claims 7 and 17 Claim 7 recites "storing the response summary in the response set further comprises storing an entry corresponding to the consultation agent in the response set, the entry comprising one or more of an identifier of the consultation agent, a response status, the response, and the response summary"; claim 17 recites the same for the method of claim 11. The combination teaches this feature. The claim requires only one of the listed items. Du supplies each agent with the responses of the other agents (Du, § 3.3; Fig. 3, "[other answers]"), which requires keeping each agent's response associated with the agent that produced it, and in the combination of claim 1 each stored entry includes the response summary. Travalini stores the diagnosis and outcome of each request "in the Home Profile information, for example, for future reference" (Travalini, ¶ [0061]). An entry that stores an agent's response and its summary is an entry corresponding to the consultation agent comprising the response and the response summary. This is the same combination, for the reasons given for claim 1. Claims 9 and 19 are rejected under 35 U.S.C. 103 as being unpatentable over Travalini in view of Zheng, Du, and Jiang. Claims 9 and 19 Claim 9 recites "wherein the one or more elements are included in a predetermined proportion of the response summaries"; claim 19 recites the same for the method of claim 11. The combination teaches this feature. Du selects the final answer by "performing majority voting" across the agents (Du, § 3.1; Table 1), and, where the agents converge, selects the "single shared answer" common to all of them (Du, § 2.2). An answer selected because it appears in a majority of the responses (more than one-half) or in all of the responses is an element included in a predetermined proportion of the responses, and in the combination of claim 1 those responses are represented by their summaries. This is the same combination, for the reasons given for claim 1. Claims 8, 10, and 18 are rejected under 35 U.S.C. 103 as being unpatentable over Travalini in view of Zheng, Du, and Jiang, and further in view of Wu (arXiv:2308.08155). Claims 8, 10, and 18 are rejected under 35 U.S.C. 103 as being unpatentable over Travalini in view of Zheng, Du, and Jiang, and further in view of Wu. Claim 10 Claim 10 recites "wherein retrieving the response to the query using the consultation agent comprises waiting for the response for a predetermined maximum time period". The combination of Travalini, Zheng, Du, and Jiang does not explicitly teach this feature. However, secondary prior-art reference Wu teaches it as follows. In Wu's multi-agent coding application, "The process might be repeated multiple times until the user's question is answered or timed-out" (Wu, § 3, A4). A process that ends when the question is answered or the time limit is reached waits for the answer for at most a predetermined time. Before the effective filing date of the claimed invention, it would have been obvious to one of ordinary skill in the art to modify the retrieval of each consultation agent's response in the combination of Travalini, Zheng, Du, and Jiang to wait for the response only until a time limit, as taught by Wu, so that retrieving the response comprises waiting for it for a predetermined maximum time period, because Wu's framework coordinates several language model agents whose replies arrive at different times and ends the exchange when it is "timed-out" (Wu, § 3, A4) rather than waiting indefinitely for an agent that does not answer. Wu is in the same field of endeavor (multi-agent large language model applications). One of ordinary skill would have had a reasonable expectation of success because a time-out is applied to the same kind of agent call. The rationale is MPEP 2143(I)(D): Wu's time-out is a known technique applied to a known multi-agent system ready for improvement, with predictable results, namely that a slow or unresponsive agent does not stall the system. KSR, 550 U.S. at 416. The modification does not change the principle of operation of the combination. This reason comes from Wu itself. Claims 8 and 18 Claim 8 recites "wherein the response status is at least one of at least pending, received, responsive, non-responsive, or timed out"; claim 18 recites the same for the method of claim 17. The combination of Travalini, Zheng, Du, Jiang, and Wu teaches this feature. In the combination of claims 7 and 17, each entry may record the outcome of the evaluation of claim 1 (whether Zheng's judge found the response acceptable, Zheng, § 3.1; Jiang, § 2.3) and, under the combination of claim 10, whether the agent's response was received or the request "timed-out" (Wu, § 3, A4). Before the effective filing date of the claimed invention, it would have been obvious to one of ordinary skill in the art to record, in the entry for each consultation agent in the combination, a status indicating whether the agent's response was received, timed out, or judged responsive or non-responsive, so that the system can determine which responses to forward to summarization (Jiang, § 2.3) and which agents did not answer in time (Wu, § 3, A4). The rationale is MPEP 2143(I)(A): the outcomes are already produced by the elements of the combination, and recording them in the entry yields the predictable result of a record of each agent's state. KSR, 550 U.S. at 416. Suggestions for Overcoming the Art The applied art does not teach the following disclosed subject matter, which applicant may consider adding to the independent claims (subject to the 35 U.S.C. 101 analysis above): • A consultation response quorum: after the evaluated responses are summarized, determining whether the response set contains at least a predetermined number of responses with associated summaries, continuing to the aggregation step only if the quorum is met, and otherwise terminating the workflow and returning a predetermined custom response retrieved from a configuration file (Spec. ¶¶ [0029], [0032]). Du and Jiang fix the number of agents or candidates in advance, but neither conditions continuation of the workflow on a quorum of evaluated and summarized responses or returns a predetermined custom response when the quorum is not met. Du and Jiang would remain the closest art. • Normalized summaries: the summarization agent outputs each response summary in a templated or normalized output structure, using a vocabulary aligned with or extracted from a domain-specific thesaurus or corpus supplied as input, and the aggregation agent extracts the elements included in a predetermined proportion of those normalized summaries (Spec. ¶¶ [0015], [0031], [0033]). Du summarizes the agents' responses into a single response for the purpose of debate, not into normalized per-response summaries using a domain-specific vocabulary for element extraction. Du would remain the closest art. This feature is most likely to be persuasive in combination with the quorum above. Further search and consideration would be required for either amendment. Conclusion Pertinent Prior Art The prior art made of record and not relied upon is considered pertinent to applicant's disclosure. The following references, including references applied or cited in prior Office actions, are made of record and are discussed for compact prosecution. Each is identified by the limitation of the pending claims to which it is pertinent. Deal - US 2014/0101079 A1, published April 10, 2014; Application No. 13/649,105, filed October 10, 2012; first-named inventor and applicant Steven Vaughn Deal. 35 U.S.C. 102(a)(1). Deal is pertinent to claims 1, 9, 11, 19, and 20, which recite "identifying one or more elements included in the response summaries" and (claims 9 and 19) elements "included in a predetermined proportion of the response summaries". Deal describes a knowledge processing system in which human agents and machine agents solve problems, using "consensus methods, such as negotiation, bargaining, or voting to reach communal decisions" (Deal, ¶ [0017]). Deal uses natural language processing "to identify and cull from a dialogue those that are substantially or statistically identical" agent inputs and to "summarize entire dialogues" (Deal, ¶ [0073]). Culling identical inputs and voting correspond generally to identifying elements common to several responses, and summarizing dialogues corresponds generally to producing summaries. Deal's agents are chiefly human. Deal does not teach large language model agents, an evaluation agent that determines whether a response is responsive, a summary of each individual response, or a repair plan. Castel - US 2014/0164603 A1, published June 12, 2014; Application No. 14/013,817, filed August 29, 2013, claiming the benefit of U.S. Provisional Application No. 61/694,535, filed August 29, 2012; first-named inventor Marc F. Castel. 35 U.S.C. 102(a)(1). Castel is pertinent to claims 1, 11, and 20, which recite "the answer comprising a repair plan, the answer to be transmitted to a scheduling system for scheduling repairs to the industrial unit". Castel describes computerized maintenance management systems that "manage a maintenance departments activities such as work orders, scheduling, reporting, inventory, and personnel" (Castel, ¶ [0006]), and a CMMS that generates and manages work orders (Castel, ¶ [0066]). Castel's system "may recommend service before unscheduled failure" and may "recommend and assist the user in ordering and organizing all of the necessary parts, service providers, contractors and technicians" for scheduled service (Castel, ¶¶ [0178]-[0179]). The CMMS corresponds to a scheduling system that receives repair information. Castel does not teach a query to a language model, or consultation, evaluation, or summarization agents. Earle - US 2024/0111960 A1, published April 4, 2024; Application No. 17/937,383, filed September 30, 2022; first-named inventor Adam Earle; applicant Tenyx, Inc. 35 U.S.C. 102(a)(2). Earle is pertinent to claims 1, 11, and 20, which recite a "summarization agent within the at least one large language model" that is "distinct from the consultation agent" and a "summarization prompt", and to claims 6 and 16, which recite a "consultation prompt template", an "evaluation prompt template", and a "summarization prompt template". In Earle, several LLMs (a customer module, an agent module, a summarizer module, and a manager module) each receive their own prompt (Earle, ¶ [0021]), and the prompt generator "applies the randomly selected state information into a template prompt such that the template prompt becomes unique to a particular LLM" (Earle, ¶ [0022]). "The summarizer module 208 is an LLM configured to extract information from synthetic chats" (Earle, ¶ [0051]), and the prompt generator generates a separate prompt for the summarizer module (Earle, ¶ [0060]). A scorer module then compares the summary to the prompts given to the other LLMs (Earle, ¶ [0070]). The summarizer LLM with its own prompt corresponds to a summarization agent, distinct from the conversing LLMs, that receives a summarization prompt, and the template prompt corresponds to the claimed prompt templates. Earle's scorer evaluates the conversation after summarization, using similarity algorithms rather than a language model (Earle, ¶¶ [0064], [0072]). Earle does not teach evaluating whether a response is responsive before summarizing it, multiple agents answering the same query, identification of elements common to several responses, or a repair plan. Johnson - US 2025/0045311 A1, published February 6, 2025; Application No. 18/362,339, filed July 31, 2023; first-named inventor Steven Todd Johnson; applicant American Express Travel Related Services Company, Inc. 35 U.S.C. 102(a)(2). Johnson is pertinent to claims 1, 11, and 20, which recite "responsive to receiving a query ... generating a consultation prompt based on the query" and "retrieving, using a consultation agent within at least one large language model, a response to the query". Johnson's claims agent executes "a content parser 114 comprising a large language model (LLM) 117 that is configured to parse an unstructured claim request in order to return a structured response to the requester" (Johnson, ¶ [0022]), and states that "A large language model 117 can be configured to return a response to a prompt" (Johnson, ¶ [0050]). Parsing a natural-language query with an LLM and returning a response corresponds generally to retrieving a response to a query from a consultation agent. Johnson concerns verification of credentials, not industrial equipment. It does not teach evaluation, summarization, or aggregation agents, or a repair plan. Krishna - US 2016/0086114 A1, published March 24, 2016; Application No. 14/845,008, filed September 3, 2015, claiming priority to Indian Application No. 4594/CHE/2014, filed September 22, 2014; first-named inventor Venkata Murali Krishna; applicant Infosys Limited. 35 U.S.C. 102(a)(1). Krishna is pertinent to claims 1, 6, 11, 16, and 20, which recite a "consultation prompt" and a "consultation prompt template". Krishna's consulting system selects a methodology that "includes consulting plan comprising a template with a list of questions" under different categories, sends the questionnaire to stakeholders and consultants, and analyzes their responses (Krishna, ¶ [0008]). Responses are used to prepare a question-and-answer document, and a further document is generated "after a process of Voting and prioritizing by the stakeholder" (Krishna, ¶ [0038]). The question template sent to several consultants corresponds generally to a consultation prompt template sent to several consultation agents, and the voting corresponds generally to identifying elements common to several responses. Krishna's consultants are human, and Krishna does not teach language model agents, evaluation or summarization agents, or a repair plan. Ferenczi - US 2025/0111141 A1, published April 3, 2025; Application No. 18/375,794, filed October 2, 2023; first-named inventor [ ] Ferenczi; applicant American Express Travel Related Services Company, Inc. 35 U.S.C. 102(a)(2). Ferenczi is pertinent to claims 1, 11, and 20 (multiple consultation agents, evaluation of responses, and "identifying one or more elements included in the response summaries"), to claims 2 and 12 ("a predetermined minimum number of response summaries"), to claims 9 and 19 ("a predetermined proportion"), and to claim 10 ("waiting for the response for a predetermined maximum time period"). Ferenczi describes "multiple LLM service providers to receive a prompt," each of which "can generate a response to the prompt by applying the prompt to one or more LLM models," after which "a threshold number of stored responses can be compared to one another to determine a level of similarity between responses and to identify any response that may correspond to a hallucination" (Ferenczi, ¶ [0007]). A similarity detection service excludes "any provider responses that are potential hallucinations," so that similar responses (e.g., two answers of "Sacramento") are returned as the accurate response (Ferenczi, ¶ [0016]). The distributed agent determines whether a threshold number of responses have been stored; if not, it "determines if a threshold time period has elapsed to receive provider responses" (Ferenczi, ¶¶ [0085]-[0086]), and where no accurate response is determined it "notifies the client application 115 that an accurate response 109 could not be determined" (Ferenczi, ¶ [0088]). For example, "if the distributed agent 118 requires at least four responses to be evaluated for a given prompt 112, the threshold number can be four" (Ferenczi, ¶ [0091]). The multiple LLM service providers correspond to multiple consultation agents; the threshold number of responses corresponds to a predetermined minimum number of responses; the threshold time period corresponds to a predetermined maximum waiting time; and selecting the responses that are similar to one another corresponds to identifying elements common to the responses. Ferenczi does not teach a separate language-model evaluation agent that determines whether a response is responsive, a summarization agent that summarizes each response, an aggregation prompt, or a repair plan for an industrial unit. Full - US 2022/0405664 A1, published December 22, 2022; Application No. 17/354,803, filed June 22, 2021; first-named inventor Cameron Full; applicant Physicians Software Solutions, LLC. 35 U.S.C. 102(a)(1). Full is pertinent to claims 1, 11, and 20 to the extent they recite "a scheduling system for scheduling repairs". Full describes "a schedule stored in the memory system, the schedule including slots for performing services," an incoming request module that receives "an inquiry from a client to reserve one of the slots in the schedule for performing a requested service," and a scheduling module that "reserve[s] a scheduled slot among the slots for performing the requested service" (Full, ¶ [0004]). The scheduling module corresponds generally to a scheduling system that schedules a requested service. Full concerns scheduling of professional (medical) services. It does not teach repairs to an industrial unit, language model agents, or the generation of a repair plan. Wilson - US 2011/0298629 A1, published December 8, 2011; Application No. 12/792,405, filed June 2, 2010; first-named inventor David C. Wilson; assignee Trane International Inc. 35 U.S.C. 102(a)(1). Wilson is pertinent to claims 1, 11, and 20, which recite "a query related to an industrial unit" and "the answer to be transmitted to a scheduling system for scheduling repairs to the industrial unit". In Wilson, the monitored mechanical equipment "may be configured as one of the components of a building heating ventilation and air-conditioning ('HVAC') system (e.g., a chiller, a compressor, fans, a cooling" tower) (Wilson, ¶ [0010]), and a remote monitoring system analyzes each alarm value and the maintenance schedule "to determine whether or not a work order 62 should be generated to prompt a technician to troubleshoot the mechanical equipment 14" (Wilson, ¶ [0020]). The chiller corresponds to an industrial unit, and the work order generated against the maintenance schedule corresponds to repair information provided for scheduling repairs. Wilson does not teach a natural-language query or language model agents. Uyeki - US 2009/0089134 A1, published April 2, 2009; Application No. 11/866,372, filed October 2, 2007; inventor Robert Uyeki. 35 U.S.C. 102(a)(1). Uyeki is pertinent to claims 1, 11, and 20, which recite "the answer comprising a repair plan, the answer to be transmitted to a scheduling system for scheduling repairs". Uyeki's system analyzes received diagnostic trouble codes "to determine an underlying problem: determine parts needed to repair the underlying problem; estimate time needed to repair the underlying problem; identify a servicing dealer, and select a proposed date/time slot for the service appointment" (Uyeki, ¶ [0009]). The parts and repair time determined for the problem correspond to a repair plan, and selecting a service appointment slot corresponds to scheduling the repair. Uyeki concerns vehicles rather than an industrial unit and does not teach language model agents. Hua - US 2016/0005242 A1, published January 7, 2016; Application No. 12/398,736, filed March 5, 2009; first-named inventor Tri T. Hua; assignee United States Government as represented by the Secretary of the Navy. 35 U.S.C. 102(a)(1). Hua is pertinent to claims 1, 11, and 20, which recite "a query related to an industrial unit" and "an answer comprising a repair plan". Hua's predictive automated maintenance system lets remote users "troubleshoot and repair the UUT in the event that a test results in a UUT fault" (Hua, ¶ [0013]), and users access a central knowledge database "to obtain information about previous system failures and repairs, suggested troubleshooting and recommended actions based on past experience of similar problems" (Hua, ¶ [0039]). Each system is "configured for a pre-scheduled maintenance activity based on the type of equipment, its operational status, and its level of usage" (Hua, ¶ [0047]). Retrieving recommended repair actions for a faulted unit corresponds generally to answering a query about an industrial unit with a repair plan. Hua does not teach language model agents or the claimed evaluation, summarization, and aggregation. Mansfield - US 2017/0004508 A1, published January 5, 2017; Application No. 15/197,122, filed June 29, 2016, claiming the benefit of U.S. Provisional Application No. 62/187,006, filed June 30, 2015; first-named inventor Stephen K. Mansfield; applicant Cross Country Home Services, Inc. 35 U.S.C. 102(a)(1). Mansfield is pertinent to claims 1, 11, and 20, which recite "the answer comprising a repair plan, the answer to be transmitted to a scheduling system for scheduling repairs". Mansfield's service response platform receives failure messages from smart appliances, determines "an appropriate responsive action" by "performing a triage process to determine a root failure" (Mansfield, ¶ [0014]), and "schedules a service request/call 149, which in turn ... causes the ERP 150 to dispatch 152 the optimal service provider 132" (Mansfield, ¶ [0064]). Mansfield's event normalizer converts manufacturer-specific appliance statuses "into a common format, a process referred to herein as 'Normalization'" (Mansfield, ¶ [0012]). The responsive action and dispatch correspond to a repair plan sent to a scheduling system, and the normalization of heterogeneous messages is relevant to the normalization of responses described in applicant's specification (Spec. ¶¶ [0015], [0031]). Mansfield does not teach language model agents or the normalization of language model responses by a summarization agent. Mazur - US 2024/0163295 A1, published May 16, 2024; Application No. 18/107,002, filed February 7, 2023, claiming the benefit of U.S. Provisional Application No. 63/425,117, filed November 14, 2022; first-named inventor David C. Mazur; applicant Rockwell Automation Technologies, Inc. 35 U.S.C. 102(a)(2). Mazur is pertinent to claims 1, 11, and 20, which recite "an industrial unit" and "a scheduling system for scheduling repairs to the industrial unit". Mazur describes computerized maintenance management systems that monitor "state data (e.g., run time, operating conditions, fault data) of the industrial automation devices 20" (Mazur, ¶ [0080]), and a CMMS process that manages "reactive maintenance, preventative maintenance, and predictive maintenance operations" and may "consistently monitor and automate scheduling of maintenance operations" for industrial automation devices (Mazur, ¶ [0145]). The CMMS corresponds to a scheduling system for scheduling repairs to an industrial unit. Mazur's subject is secure deployment of configuration data to industrial devices; it does not teach a query, language model agents, or a repair plan generated from agent responses. Allouch - Merav Allouch, Amos Azaria, and Rina Azoulay, "Conversational Agents: Goals, Technologies, Vision and Challenges," Sensors, vol. 21, no. 24, article 8448, published December 17, 2021. 35 U.S.C. 102(a)(1). Allouch is pertinent to claims 1, 11, and 20, which recite consultation agents answering a query, "an evaluation of the response", and "generating an answer" from multiple responses. Allouch surveys conversational agents across domains including "industrial" domains (Allouch, Abstract). Allouch reports that "the ensemble-based CA developed by Serban et al. [74] for the Amazon Alexa Prize competition utilizes an ensemble of NLG and retrieval models" and that "Their dialogue manager is trained to select an appropriate response" (Allouch, § 4.2, p. 12). Allouch also describes evaluation of agent responses, noting that "humans can evaluate whether the CA responses seem appropriate," that "since human evaluation procedures are expensive, several automatic metrics have been proposed for the evaluation process" (Allouch, § 7, p. 26), and that "machine-based evaluation or hybrid human-machine-based evaluation are widely used to examine the quality of CAs" (Allouch, § 7.2). Selecting a response from an ensemble of models corresponds generally to generating an answer from multiple responses, and automatic evaluation of agent responses corresponds generally to evaluating a response. Allouch is a survey; it does not teach a large language model evaluation agent that gates summarization, summarization agents, or a repair plan. Chan - Chi-Min Chan, Weize Chen, Yusheng Su, Jianxuan Yu, Zhiyuan Liu, Jie Fu, Wei Xue, and Shanghang Zhang, "ChatEval: Towards Better LLM-based Evaluators through Multi-Agent Debate," arXiv:2308.07201v1, submitted August 14, 2023 (35 U.S.C. 102(a)(1)). Chan is pertinent to the limitations of claims 1, 11, and 20 reciting an "evaluation agent", a "summarization agent" distinct from the other agents, and "identifying one or more elements included in the response summaries". In Chan, several LLM agents act as a referee team, and "each agent autonomously receives responses from the others and, in turn, delivers its own responses to them" (Chan, § 2). In Chan's simultaneous-talk-with-summarizer strategy, "we additionally employ another LLM as a summarizer," and "we prompt this extra LLM to summarize the messages conveyed so far" (Chan, § 2). The final result is derived "from the majority vote among various annotators" (Chan, § 2). The separate summarizer LLM corresponds to a summarization agent distinct from the agents whose output it summarizes, and the majority vote corresponds to identifying elements included in a proportion of the agents' outputs. Chan applies these techniques to evaluating text produced by other models, not to answering a query about an industrial unit, and it does not summarize each response separately or generate a repair plan. Madisetti - US 2024/0394289 A1, published November 28, 2024; Application No. 18/795,345, filed August 6, 2024, a continuation of Application No. 18/470,487, filed September 20, 2023, which is a continuation of an application filed July 7, 2023, and claiming the benefit of U.S. Provisional Application Nos. 63/463,913 (filed May 4, 2023) and 63/469,571 (filed May 30, 2023); first-named inventor Vijay Madisetti; assignee Vijay Madisetti. Madisetti is pertinent to the limitations of claims 1, 11, and 20 reciting multiple consultation agents and "generating an answer based on the response set". Madisetti describes an input broker that generates "multiple derived prompts for different categories," where "The derived prompts 822 are sent multiple h-LLMs 824 which produce the results," and an output broker that "computes various scores and assigns weights for ranking the results" (Madisetti, ¶ [0063]); "the best result is sent to the user 900" (Madisetti, ¶ [0064]). The scoring of each LLM's result corresponds to evaluating the responses, and returning the best-ranked result corresponds to generating an answer from the set of responses. Madisetti does not teach that a separate language model evaluates or summarizes each response, or that the answer is generated from elements common to the summaries. Suggested Interview The examiner invites applicant's representative to a telephone or video interview before filing a response, to discuss amendments that may advance prosecution. Applicant may request an interview using the Automated Interview Request (AIR) form. The following topics are suggested: • Prior art: whether the independent claims can be amended to recite the consultation response quorum and custom-response termination (Spec. ¶¶ [0029], [0032]), together with the normalized response summaries and the extraction by an aggregation agent of elements included in a predetermined proportion of those summaries (Spec. ¶¶ [0015], [0031], [0033]). Claims 1, 11, and 20 are affected. Du and Jiang would remain the closest art. • 35 U.S.C. 101: whether the same amendments, recited with the specificity set out in the Suggested Amendment under 35 U.S.C. 101, would reflect the improvement described at Spec. ¶ [0015] and integrate the claims into a practical application. Claims 1, 11, and 20, and claims 5 and 15 as the dependent claims closest to eligibility, are affected. • 35 U.S.C. 112(b): correcting the antecedent basis of "the response summaries," "the responses," and "the summarization agents" in claims 1, 11, and 20, and the "one or more" agent terms of claim 15, by introducing a plurality of each agent that includes the recited agent. • 35 U.S.C. 112(b) and 112(d): reciting the use of the aggregation prompt by an aggregation agent to produce the answer (former claim 3; Spec. ¶¶ [0033], [0071]), which would also resolve the improper dependency of claims 4 and 14 and the form of claim 14. Applicant is requested to send any proposed claim amendments to the examiner before the interview, so that the examiner can state, for each proposed amendment, whether it would overcome each outstanding rejection and why (MPEP 713.01, 713.04). Any comments considered necessary by applicant must be submitted no later than the payment of the issue fee and, to avoid processing delays, should preferably accompany the issue fee. Such submissions should be clearly labeled “Comments on Statement of Reasons for Allowance.” THIS ACTION IS MADE FINAL Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). 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 extension fee 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 date of this final action. Contact Information Any inquiry concerning this communication or earlier communications from the examiner should be directed to MATTHEW T. SITTNER whose telephone number is (571) 270-7137 and email: matthew.sittner@uspto.gov. The examiner can normally be reached on Monday-Friday, 8:00am - 5:00pm (Mountain Time Zone). Please schedule interview requests via email: matthew.sittner@uspto.gov If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Sarah M. Monfeldt can be reached on (571) 270-1833. Information regarding the status of an application may be obtained from the Patent Application Information Retrieval (PAIR) system. Status information for published applications may be obtained from either Private PAIR or Public PAIR. Status information for unpublished applications is available through Private PAIR only. For more information about the PAIR system, see http://pair-direct.uspto.gov. Should you have questions on access to the Private PAIR system, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative or access to the automated information system, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000. /MATTHEW T SITTNER/ Primary Examiner, Art Unit 3629b