Notice of Pre-AIA or AIA Status
1. The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
DETAILED ACTION
2. This Office Action is sent in response to Applicant’s Communication received on 08/12/2026 for application number 18/393,019.
Response to Amendments
3. The Amendment filed 08/12/2026 has been entered. Claims 1, 3-5, 7, 8, 10-12, 14, 15, 17, and 18 have been amended. Claims 1-20 remain pending in the application.
4. Applicant’s amendments to Claims 3, 4, 7, 10, 11, 14, 17, and 18 have been fully considered and are persuasive. The objections to these Claims are respectfully withdrawn.
5. Applicant’s amendments to Claims 1, 8, and 15 have been fully considered and are persuasive. The amendments provided to overcome the §101 rejection set forth in the previous office action are sufficient. Accordingly, the 35 U.S.C § 101 rejection of Claims 1-20 is respectfully withdrawn.
Response to Arguments
Applicant argues that Friedman discusses a completely different method of generating responses using a dialogue manager that simply generates "sensible, useful, grounded responses at each turn" based on context information. For example, Friedman describes "Dialogue management is the central module of a CRS, acting as the interface between the user and the rest of the system. It is responsible for guiding the user through a multi-turn exploration of the recommendation corpus and generating sensible, useful, grounded responses at each turn." (Friedman, Section 3.1, pages 3-4). Therefore, Friedman at most describes using a neural network model to generically generate responses based on general context information. Friedman does not discuss "the second neural network model makes a decision based on the conversational data, and generates a response based on the decision, wherein the decision includes one of (i) a knowledge search including generating a training query based on the conversation data and retrieving at least one knowledge article in the one or more knowledge documents using the training query, and (ii) a product search including generating the training query and retrieving at least one top-ranking product from the list of products using the training query, and wherein the response includes a concatenation of the at least one knowledge article or a concatenation of the at least one top-ranking product" recited in amended claim 1.
Examiner respectfully disagrees and notes that Applicant relies on Friedman’s general statement that the dialogue manager generates “sensible, useful, and grounded responses at each turn,” but Friedman’s more specific disclosure that the unified LLM processes the prior conversation context and generates output that trigger different system actions. In particular, Friedman teaches that "Response: <message>" will cause message to be shown as a user facing response, and "Request: <query>" will cause query to be sent to the recommendation engine backend to retrieve a slate of recommendations. Friedman further teaches, in its Search API Lookup section, that the LLM processes the session history, output a search query, and supplies that query to a search algorithm to retrieve items, after which a ranker determines which retrieved items are included in the recommendation slate and in what order. Thus, Friedman does not merely generate a generic contextual response, it teaches an LLM making an action determination from conversational context and, for a retrieval action, generating a query, retrieving candidate items, ranking those items, and incorporating the selected recommendations into the system output. Accordingly, Friedman in view of Zeng teaches the amended claim 1.
Claim Rejections – 35 USC § 103
6. 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.
7. Claims 1-6, 8-13, and 5-20 are rejected under 35 U.S.C. 103 as being unpatentable over Friedman et al. (Leveraging Large Language Models in Conversational Recommender Systems, arXiv, published 16 May 2023, pages 1-24) in view of Zeng et al. (U.S. Patent Application Pub. No. US 20180053119 A1).
Claim 1: Friedman teaches a method for a training framework of an artificial intelligence (AI) conversation agent to provide assisted shopping experience for shoppers (i.e. A Conversational Recommender System (CRS) offers increased transparency and control to users by enabling them to engage with the system through a real-time multi-turn dialogue. In this section we discuss the use of LLM-powered user simulators to generate realistic data at scale and techniques for tuning system components using larger amounts of data. Recommender systems are one of the most prominent success stories of machine learning in industry, delivering personalized content to billions of users over a wide range of domains such as Search, Videos, News, and Shopping; abs, Section 1, 4, pages 1, 7), comprising:
retrieving, from a unit, a list of products for recommendation to a user and one or more knowledge documents regarding the list of products (i.e. Recommender systems are one of the most prominent success stories of machine learning in industry, delivering personalized content to billions of users over a wide range of domains such as Search, Videos, News, and Shopping. The purpose of the retrieval phase is to take the full corpus, which for some domains such as videos or urls may contain hundreds of millions of items, and based on the context select a small number of candidate items (e.g. 100) that will be fed to a downstream ranker. A recommendation engine then uses a tractable search algorithm to retrieve candidates from the item corpus. In the case of a YouTube video recommender, this metadata consists of information such as the title, knowledge graph entities associated with the video, developer description of the video, transcript of the video, and user comments; Section 3.2.1, 3.2.2, pages 1, 5-6), retrieving candidate items for recommendation and using item-related knowledge/metadata documents or textual artifacts regarding those items;
causing an interactive simulation between a first neural network model and a second neural network model to generate conversational data (i.e. we propose new implementations for user preference understanding, flexible dialogue management and explainable recommendations as part of an integrated architecture powered by LLMs. We propose techniques for building a controllable LLM-based user simulator to generate synthetic conversations. Given a fixed CRS and such a user simulator 𝑓, we can generate a new sample session by having the CRS and 𝑓 interact for a given number of turns (i.e. the CRS generates each 𝑠𝑖 and 𝑓 generates each 𝑢𝑖); abs, Section 4.1, pages 1, 8), an interactive simulation between a first LLM model, the user simulator, and a second LLM model, the CRS/dialogue manager, wherein:
the first neural network model generates a query relating to a product of the list of products (i.e. The LLM takes as input the full history of the ongoing conversation and outputs the next user utterance, analogous to how a CRS dialogue manager can use an LLM to generate system utterances. The LLM takes as input the full history of the ongoing conversation and outputs the next user utterance, analogous to how a CRS dialogue manager can use an LLM to generate system utterances; Section 3.2.1, Section 4.1, pages 5, 6, 8), an LLM user simulator generates user utterances/queries in a recommendation session and those queries relate to candidate items from the corpus, and
the second neural network model makes a decision based on the conversational data (i.e. An LLM takes as input the full session context and outputs a sequence of messages ending in a terminal output that triggers a system action, such as a response to the user … From an end-to-end point of view, given context information (dialogue history, a user profile, item summaries, etc.), the goal of the dialogue manager is to generate system actions to take, as well as an appropriate system utterance; Section 3.1, page 3), and generates a response based on the decision (i.e. Dialogue management is the central module of a CRS, acting as the interface between the user and the rest of the system. It is responsible for guiding the user through a multi-turn exploration of the recommendation corpus and generating sensible, useful, grounded responses at each turn … an output "Response: <message>" will cause message to be shown as a user facing response, and "Request: <query>" will cause query to be sent to the recommendation engine backend to retrieve a slate of recommendations; Section 3.1, pages 3, 4), CRS/dialogue manager generates a response based on context information, including item summaries; wherein the decision includes one of (i) a knowledge search including generating a training query based on the conversation data and retrieving at least one knowledge articles in the one or more knowledge documents using the training query, and (ii) a product search including generating the training query (i.e. Search API Lookup. In this method, the LLM directly outputs a search query, which gets fed into a black-box search API to retrieve items, Analogous to Concept Based Search, in Search API Lookup the LLM can be taught to output relevant search queries using a small number of examples (see Section 3.1), but the quality of retrieval is limited by the extent to which a search query can properly represent the full context of a conversation; Section 3.2.1, 4.2, pages 6, 9) and retrieving at least one top-ranking product from the list of products using the training query (i.e. the LLM directly outputs a search query, which gets fed into a black-box search API to retrieve items … A recommendation engine then uses a tractable search algorithm to retrieve candidates from the item corpus … After candidate items have been retrieved, a ranker decides which of them will be included in the recommendation slate and in what order; Section 3.2.1, 3.2.2, pages 5-6), and wherein the response includes a concatenation of the at least one knowledge article or a concatenation of the at least one top-ranking product (i.e. Once triggered by the dialogue management module, it is the responsibility of the recommendation module to return a slate of high quality, relevant, and diverse recommendations that will be shown to the user … After candidate items have been retrieved, a ranker decides which of them will be included in the recommendation slate and in what order … According to the conversational recommender setup considered in this paper (see Section 2), a session consists of a sequence 𝑆 = {𝑠1,𝑢1, 𝑠2,𝑢2, ..., 𝑠𝑛,𝑢𝑛}, where each 𝑢𝑖 is a natural language utterance by the user and each 𝑠𝑖 is a combination of a natural language utterance and possibly a slate of recommendations by the CRS; Section 3.2, 4.1, pages 4-9);
the conversational data including the training query and the response from the interactive simulation (i.e. Given a fixed CRS and such a user simulator 𝑓, we can generate a new sample session by having the CRS and 𝑓 interact for a given number of turns (i.e. the CRS generates each 𝑠𝑖 and 𝑓 generates each 𝑢𝑖). Let R be a set of sessions generated by having real users interact with a particular CRS, and Q be a set of simulated sessions sampled from the CRS and a user simulator 𝑓 according to the procedure outlined above; Section 4.1, 4.2, page 8), a set of sessions, simulated sessions, or training examples used for later tuning/training;
training the second neural network model based on the conversational data (i.e. In this section we discuss the use of LLM-powered user simulators to generate realistic data at scale and techniques for tuning system components using larger amounts of data. Using this data, we can tune the ranking LLM to predict the ground truth labels as a regression problem; Section 4, pages 7-9); and
generating a recommendation response using trained second neural network model to a user query from a shopper (i.e. Once triggered by the dialogue management module, it is the responsibility of the recommendation module to return a slate of high quality, relevant, and diverse recommendations that will be shown to the user … This can either be an initial recommendation slate or a refinement of an earlier slate from the session based on feedback from the user; Section 1, 3.2. pages 4-5).
Friedman does not explicitly teach a memory unit and storing.
However, Zeng teaches a method for a training framework of an artificial intelligence (Al) conversation agent to provide assisted shopping experience for shoppers (i.e. the present teaching discloses methods for designing and developing intelligent virtual agents, which can automatically generate and recommend response/reply messages for assisting human representatives or acting as virtual representatives/agents to communicate with customers in a more efficient and effective way … a virtual sales agent 182 … The virtual agent may also perform product/service recommendation to the user based on a user database 132; para. [0046, 0053, 0055]), comprising: retrieving, from a memory unit (i.e. The exemplary computer platform includes an internal communication bus 1710, program storage and data storage of different forms, e.g., disk 1770, read only memory (ROM) 1730, or random access memory (RAM) 1740, for various data files to be processed and/or communicated by the computer, as well as possibly program instructions to be executed by the CPU; para. [0173]), a list of products for recommendation to a user and one or more knowledge documents (i.e. The service virtual agent 1 142 in this example comprises a dynamic dialog state analyzer 210, a dialog log database 212, one or more deep learning models 225, a customized FAQ generator 220, a customized FAQ database 222, various databased (e.g., a knowledge database 134, a publisher database 136, . . . , and a customized task database 139); para. [0061]) regarding the list of products (i.e. There are additional product sources 1460 where a plurality of products sources 1460-1 . . . 1460-2 that the user may be connected to and be able to search for products via conversations with the service virtual agents 140; para. [168]); storing the conversational data including the training query and the response from the interactive simulation (i.e. The dynamic dialog state analyzer 210 record dialog logs including both the dialog states and other metadata related to the dialog, into the dialog log database 212, which can be used by the customized FAQ generator 220 for further generating customized FAQs. The dynamic dialog state analyzer 210 may also estimate user intent based on the dialog state determined by analyzing the received user input. The estimated user intent is then sent to the real-time task manager 230 for real-time task management; para. [0063, 0078, 0100]); training the second neural network model based on the conversational data (i.e. The semi-supervised learning mechanism 300 is provided to obtain the deep learning models 225 via semi-supervised learning and comprises a parser 342, a structured information identifier 346, an entity identifier 348, an unstructured information identifier 350, a semi-supervised training seeds generator 354, and a learning engine 352. In operation, the parser 342 takes conversation data from actual dialogs with users and training seeds 359 (generated by the semi-supervised training seeds generator 354 as input. Based on natural language models 340 and dictionaries 344, the parser 342 parses the input conversations and sends such processed results to various identifiers to extract relevant information; para. [0081-084], 0094); and generating a recommendation response using trained second neural network model to a user query from a shopper (i.e. If the next task type is determined at 308 to recommend a product or service to the user, the recommendation engine 250 is invoked to analyze, at 330, the user information from the user database 132 and recommends, at 332, one or more products or services that match the dynamically estimated user intent (interest) and/or the user information. Then the process goes to 334 for storing dialog logs in a database; para. [0079]).
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to modify the invention of Friedman to include the feature of Zeng. One would have been motivated to make this modification because it provides methods for designing and developing intelligent virtual agents, which can automatically generate and recommend response/reply messages for assisting human representatives or acting as virtual representatives/agents to communicate with customers in a more efficient and effective way, to achieve similar or even better customer satisfaction with minimum human involvement.
Claim 2: Friedman and Zeng teach the method of claim 1. Friedman does not explicitly teach wherein the model is located on an external server accessible via a communication network.
However, Zeng teaches wherein the model is located on an external server accessible via a communication network (i.e. the implementation of various components described above may be embodied in a hardware device, it may also be implemented as a software only solution—e.g., an installation on an existing server; para. [0168, 0169, 0177]).
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to modify the invention of Friedman to include the feature of Zeng. One would have been motivated to make this modification because it provides centralized, scalable, and remotely accessible AI training services for generating simulated user utterances and synthetic conversations.
Claim 3: Friedman and Zeng teach the method of claim 1. Friedman further teaches wherein the interactive simulation is caused by: feeding the list of products (i.e. Once triggered by the dialogue management module, it is the responsibility of the recommendation module to return a slate of high quality, relevant, and diverse recommendations that will be shown to the user; Section 3.2. pages 4-5) and a first prompt instructing the first neural network model to generate questions relating to one or more characteristics of the product as input to the first neural network model (i.e. The LLM takes as input the full history of the ongoing conversation and outputs the next user utterance, analogous to how a CRS dialogue manager can use an LLM to generate system utterances. The LLM takes as input the full history of the ongoing conversation and outputs the next user utterance, analogous to how a CRS dialogue manager can use an LLM to generate system utterances; Section 3.2.1, Section 4.1, pages 5, 6, 8).
Claim 4: Friedman and Zeng teach the method of claim 3. Friedman further teaches wherein the interactive simulation is caused by: feeding the query (i.e. The user outputs a natural language message on their turn, and the agent responds with a natural language message, optionally updating the slate of recommendations based on the conversation; Section 2, page 3) and the one or more knowledge documents together (i.e. given context information (dialogue history, a user profile, item summaries, etc.), the goal of the dialogue manager is to generate system actions to take, as well as an appropriate system utterance; Section 3.1, pages 3-4) with a second prompt instructing the second neural network model to generate answers using at least one knowledge document as context information as input to the second neural network model (i.e. given context information (dialogue history, a user profile, item summaries, etc.), the goal of the dialogue manager is to generate system actions to take, as well as an appropriate system utterance; Section 3.1, pages 3-4).
Claim 5: Friedman and Zeng teach the method of claim 4. Friedman further teaches wherein the second neural network model comprises a retriever model that retrieves the at least one knowledge document as relevant to the query or the at least one top-ranking product as relevant to the query (i.e. The purpose of the retrieval phase is to take the full corpus, which for some domains such as videos or urls may contain hundreds of millions of items, and based on the context select a small number of candidate items (e.g. 100) that will be fed to a downstream ranker. A recommendation engine then uses a tractable search algorithm to retrieve candidates from the item corpus. In the case of a YouTube video recommender, this metadata consists of information such as the title, knowledge graph entities associated with the video, developer description of the video, transcript of the video, and user comments; Section 3.2.1, 3.2.2, pages 5-6).
Claim 6: Friedman and Zeng teach the method of claim 1. Friedman further teaches wherein the interactive simulation is caused by iteratively feeding the response as an input to the first neural network model to generate a next query (i.e. According to the conversational recommender setup consid ered in this paper (see Section 2), a session consists of a sequence 𝑆 = {𝑠1,𝑢1,𝑠2,𝑢2,...,𝑠𝑛,𝑢𝑛}, where each 𝑢𝑖 is a natural language utterance by the user and each 𝑠𝑖 is a combination of a natural language utterance and possibly a slate of recommendations by the CRS. Therefore, a user simulator is defined by a function 𝑓 (𝑆′) = 𝑈𝑖, where 𝑆′ = {𝑠1,𝑢1,𝑠2,𝑢2,...,𝑠𝑖} is a partial session and 𝑈𝑖 is a dis tribution over possible user utterances 𝑢𝑖 continuing the session. Given a fixed CRS and such a user simulator 𝑓, we can generate a new sample session by having the CRS and 𝑓 interact for a given number of turns (i.e. the CRS generates each 𝑠𝑖 and 𝑓 generates each 𝑢𝑖); Section 4.1; pages 7-8).
Claim 8 is similar in scope to Claim 1 and is rejected under a similar rationale.
Zeng further teaches a memory storing (i.e. The exemplary computer platform includes an internal communication bus 1710, program storage and data storage of different forms, e.g., disk 1770, read only memory (ROM) 1730, or random access memory (RAM) 1740, for various data files to be processed and/or communicated by the computer, as well as possibly program instructions to be executed by the CPU; para. [0173]), and a plurality of processor-executable instructions; and one or more processors executing the instructions to perform operations (i.e. The exemplary computer platform includes an internal communication bus 1710, program storage and data storage of different forms, e.g., disk 1770, read only memory (ROM) 1730, or random access memory (RAM) 1740, for various data files to be processed and/or communicated by the computer, as well as possibly program instructions to be executed by the CPU; para. [0173]).
Claims 9-13 are similar in scope to Claims 2-6 and are rejected under a similar rationale.
Claim 15 is similar in scope to Claim 1 and is rejected under a similar rationale.
Zeng further teaches a non-transitory processor-readable storage medium storing a plurality of processor-executable instructions (i.e. machine readable non-transitory medium is disclosed, wherein the medium has information for generating knowledge for a chat bot; para. [0010]) for a training framework of an artificial intelligence (AI) conversation agent, the processor-executable instructions executed by one or more processors to perform operations (i.e. The exemplary computer platform includes an internal communication bus 1710, program storage and data storage of different forms, e.g., disk 1770, read only memory (ROM) 1730, or random access memory (RAM) 1740, for various data files to be processed and/or communicated by the computer, as well as possibly program instructions to be executed by the CPU; para. [0173]).
Claims 16-20 are similar in scope to Claims 2-6 and are rejected under a similar rationale.
8. Claims 7 and 14 are rejected under 35 U.S.C. 103 as being unpatentable over Friedman in view of Zeng, and further in view of Rajagopal et al. (U.S. Patent Application Pub. No. US 20210064826 A1).
Claim 7: Friedman and Zeng teach the method of claim 1. Friedman further teaches wherein the training the second neural network model based on the conversational data comprises:
feeding a first query from the conversational data to the second neural network model (i.e. The user outputs a natural language message on their turn, and the agent responds with a natural language message, optionally updating the slate of recommendations based on the conversation; Section 2, page 3); generating, by the second neural network model, a predicted response conditioned on the one or more knowledge documents and the conversational data (i.e. The conversation is used as context for the user’s preferences and the video metadata is used as context for the item. The LLM takes in summaries of the item side and context side to produce a score for the item and an explanation for the score; Fig. 6, page 6); computing a loss (i.e. Given this data, we can tune a Generalized Dual Encoder Model (see Section 3.2.1), in which the initial context representation and item representations are each encoded by an LLM. Regardless of whether we choose to tune only the adapter layers of the two tower model or the LLM params as well, the loss is fully differentiable and normal supervised learning with gradient descent suffices; Section 4.2, page 9); and updating the second neural network model based on the loss (i.e. Further tune the unified LLM on simulated sessions through reinforcement learning to optimize for proxy rewards generated by these reward models; Section 4.2, page 9).
Friedman does not explicitly teach computing a loss by computing a first response from the conversational data and the predicted response.
However, Rajagopal teaches computing a loss by computing a first response from the conversational data and the predicted response (i.e. The conversation simulator 800 may further compare the observed Bot response with expected Bot response to assess Bot's performance in response to the training … verifying the expected Bot response against the observed bot response, which may include an intent mismatch, not an identified utterance, and business logic failure; para. [0098, 0101]); and updating the second neural network model based on the loss (i.e. The recommendation may include how each of the Bots improves its accuracy and authenticity, which may be used for a next round of training; para. [0039, 0061, 0116]).
Therefore, it would have been obvious to one of ordinary skill in the art before the effective filling date of the claimed invention to modify the combination of Friedman and Zeng to include the feature of Rajagopal. One would have been motivated to make this modification because it improves training efficiency, test coverage, and the accuracy of the conversational agent.
Claim 14 is similar in scope to Claim 7 and is rejected under a similar rationale.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant’s disclosure.
Zhang et al. (Pub. No. US 20210358007 A1), Implementations of the disclosed subject matter provide systems and methods of recommending products of an electronic product catalog of online commercial retailers. Interactions by a user with an electronic product catalog may be tracked. The tracked activities may tokenized, so that there may be sequences of tokens for each user that are translated from the user's activity with an online electronic product catalog.
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.
It is noted that any citation to specific pages, columns, lines, or figures in the prior art references and any interpretation of the references should not be considered to be limiting in any way. A reference is relevant for all it contains and may be relied upon for all that it would have reasonably suggested to one having ordinary skill in the art. In re Heck, 699 F.2d 1331, 1332-33, 216 U.S.P.Q. 1038, 1039 (Fed. Cir. 1983) (quoting In re Lemelson, 397 F.2d 1006, 1009, 158 U.S.P.Q. 275, 277 (C.C.P.A. 1968)).
Any inquiry concerning this communication or earlier communications from the examiner should be directed to TAN TRAN whose telephone number is (303)297-4266. The examiner can normally be reached on Monday - Thursday - 8:00 am - 5:00 pm MT.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Matt Ell can be reached on 571-270-3264. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
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.
/TAN H TRAN/Primary Examiner, Art Unit 2141
Prosecution Insights