DETAILED ACTION
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
All previous objections and rejections directed to the Applicant’s disclosure and claims not discussed in this Office Action have been withdrawn by the Examiner.
Response to Amendments and Arguments
The 101 and 112 rejections have been remedied and thus removed.
The applicant’s amendments with respect to claim 1 have been carefully considered, but are not persuasive. The applicant argues that the present invention discloses a processor-implemented method for recommending suitable assets in enterprise problem-solving contexts by employing descriptor-based prompt-tuning, embedding generation, similarity scoring, and optimal prompt construction, performing a specialized set of operations that go beyond the generic RAG techniques of Jain. These
The applicant also states that although Jain describes the use of embeddings and LLMs in combination, it does not teach or suggest descriptor-based prompt-tuning prior to generating embeddings, nor the augmentation of descriptors to enrich the embedding process.
The applicant next argues that Jain does not disclose or imply the construction of an optimal prompt that restricts the LLM's processing to a specific, limited context of identified assets and a user query.
Lastly, the applicant states that Also, notably absent from Jain is the use of few-shot prompting as expressly required by the amended claim. The specification teaches that few-shot prompting enables the LLM to provide effective recommendations with minimal training examples, thereby reducing computational load and improving efficiency. The examiner notes that original claim 7 describing use of few-shot prompting has been incorporated into claim 1. Few-shot prompting is clearly taught by Jain et al., para [0038].
Claim Rejections - 35 USC § 102
The following is a quotation of the appropriate paragraphs of 35 U.S.C. 102 that form the basis for the rejections under this section made in this Office action:
The following is a quotation of the appropriate paragraphs of 35 U.S.C. 102 that form the basis for the rejections under this section made in this Office action:
A person shall be entitled to a patent unless –
(a)(2) the claimed invention was described in a patent issued under section 151, or in an application for patent published or deemed published under section 122(b), in which the patent or application, as the case may be, names another inventor and was effectively filed before the effective filing date of the claimed invention.
Claim(s) 1, 3, 5-8, 10, and 12-14 is/are rejected under 35 U.S.C. 102(a)(2) as being anticipated by US 20250181619, hereinafter referred to as Jain et al.
Regarding claim 1 (Currently Amended), Jain et al. discloses a method for recommendation of suitable assets using a Large Language Model (LLM) Personalized recommendations are made across multiple domains by employing a domain-specific biasing approach. similarity distance scores are normalized, and a weighted confidence score is calculated for each knowledge domain, considering the current domain context, domain popularity, and conversational history. A tunable global confidence score serves as a threshold for comparing and suggesting knowledge domains, enhancing the grounding of LLM results,” Jain et al., para [0017].), the method comprising:
receiving descriptors, by a processor (Jain et al., para [0021]), for each asset from amongst a plurality of assets (“At step 310, a query is received. The query may be received from an interface, such as user interface (102) of FIG. 1,” Jain et al., para [0105]. And, Jain et al., para [0107], explains that queries are classified into domains (i.e., a description relating to the query). The domains are interpreted as assets.), wherein each asset along with a corresponding descriptor is used to prompt-tune the LLM prior to generating embeddings (“If the response is deemed suboptimal, the feedback is used to adjust the model's parameters or to re-prompt the LLM with adjusted embeddings or additional context. The evaluation may be performed by additional machine learning models trained for quality assessment or through human-in-the-loop interventions. This iterative process may continue until the generated response meets the desired standards,” Jain et al., para [0115]. Here, LLM’s parameters are adjusted (i.e., tuned) with a re-prompt. The tuning of the LLM allows for acquiring more accurate embeddings (post-tuning).), and wherein the descriptors are augmented to generate additional descriptors (“The use case example demonstrates how the knowledge service (500) may provide context-aware responses to Jane's queries. Using the reasoning capabilities of the LLM augmented with specific information retrieved from the particular domains, the knowledge service significantly improves Jane's efficiency and productivity by delivering precise, grounded responses that are focused within the domain of her particular query,” Jain et al., para [0135].);
generating embeddings, by the processor, for each asset from amongst the plurality of assets, by employing the LLM, based on the provided descriptors of each asset from amongst the plurality of assets (“At step 320, the query is classified into a first domain within a plurality of domains. Query classification may be performed using one or more natural language processing (NLP) techniques, including preprocessing, tokenization, and normalization. The preprocessed text is then transformed into numerical vectors using word embedding techniques, allowing machine learning models to process and understand the linguistic patterns,” Jain et al., para [0107].);
creating a database , by the processor, of assets by storing the generated embeddings for each asset from amongst the plurality of assets in the database in an indexed form to enable quick identification and retrieval of specific assets (“At step 330, an index is retrieved of domain-specific vector embeddings that corresponds to the recommended domains. When a query is classified into a domain, the system queries the database or search engine according to the domain specific index corresponding to the identified domain. In other words, once a query is classified into a domain, the system utilizes this classification to retrieve the corresponding vector embeddings, which encapsulate the domain-specific semantic context. Retrieval of the vector embeddings may be performed using databases or search engines such as Elasticsearch, which are capable of managing the complex data structures that vector embeddings represent. For example, using the dense vector data type, Elasticsearch is able to index and search through vector fields, allowing for efficient retrieval based on similarity scoring, such as cosine similarity or Euclidean distance measures,” Jain et al., para [0109]-[0110].);
receiving, by the processor, a user query pertaining to a request for recommendation to solve an enterprise problem (“At step 320, the query is classified into a first domain within a plurality of domains. Query classification may be performed using one or more natural language processing (NLP) techniques, including preprocessing, tokenization, and normalization. The preprocessed text is then transformed into numerical vectors using word embedding techniques, allowing machine learning models to process and understand the linguistic patterns. NLP frameworks such as Hugging Face's Transformer models (i.e., BERT or GPT) may be used to classify queries by outputting probabilities for each domain, with the highest probability indicating the query's domain, thus informing the next steps in query processing or response generation. Probabilities from these models may then be augmented, as shown in figures one and two, to determine the recommended domain. At step 330, an index is retrieved of domain specific vector embeddings that corresponds to the recommended domains,” Jain et al., para [0107]-[0109].);
generating cosine similarity scores, by the processor, wherein each cosine similarity score is generated between the user query and a corresponding generated embedding for a given asset stored in the database of assets (“Retrieval of the vector embeddings may be performed using databases or search engines such as Elasticsearch, which are capable of managing the complex data structures that vector embeddings represent. For example, using the dense vector data type, Elasticsearch is able to index and search through vector fields, allowing for efficient retrieval based on similarity scoring, such as cosine similarity or Euclidean distance measures,” Jain et al., para [0110].);
identifying, by the processor, a predefined number of similar assets from amongst the database of assets, based on highest cosine similarity scores (Jain et al., para [0110].);
constructing an optimal prompt, by the processor, based on the identified predefined number of similar assets and the user query (“At step 340, the LLM is prompted with the query and the domain-specific vector embeddings. The system then prompts the LLM with both the original query and these domain-specific embeddings,” Jain et al., para [0111].), wherein the constructed optimal prompt instructs the LLM to use only provided context comprising the user query and the identified predefined similar assets (“The orchestrator includes functionality to prompt the large language model based on the original user query, and domain specific embeddings retrieved by the information retrieval system,” Jain et al., para [0024]. Here, the context comprises the user query and the domain specific embeddings retrieved by the information retrieval system (i.e., identified predefined similar assets).); and
prompting the LLM, by the processor, using the constructed optimal prompt for generating a response as a recommendation of the identified predefined number of similar assets as the suitable assets for solving the user query (“At step 350, a query response is received from the LLM as grounded with the most relevant index results. For example, the process may deliver the query response to the end-user or another downstream system through a combination of web technologies and communication protocols,” Jain et al., para [0114].), wherein the prompting is implemented by using a technique of few-shot prompting (“This volume of data allows the expansion of GPT-3 to 175 billion parameters using 96 attention layers, each with a 96×128 dimension head, enabling few or zero-shot training paradigms. By prompting the model with a few response paradigms, the GPT-3 model understands the context, produces results, and may structure responses automatically, without retraining parameters,” Jain et al., para [0038].).
As to claim 8, system claim 8 and method claim 1 are related as method and system of using same, with each claimed element’s function corresponding to the method step. Accordingly claim 8 is similarly rejected under the same rationale as applied above with respect to method claim. Also, Jain et al., para [0127] and [0139], teaches processor, memory, CRM, and instructions.
Regarding claim 3 (Original), Jain et al. discloses the method according to claim 1, wherein the descriptors of each asset from amongst the plurality of assets comprises: an asset title, an asset description, asset metadata (“At step 310, a query is received. The query may be received from an interface, such as user interface (102) of FIG. 1,” Jain et al., para [0105]. And, Jain et al., para [0107], explains that queries are classified into domains (i.e., a description relating to the query).).
As to claim 10, system claim 10 and method claim 3 are related as method and system of using same, with each claimed element’s function corresponding to the method step. Accordingly claim10 is similarly rejected under the same rationale as applied above with respect to method claim. Also, Jain et al., para [0127] and [0139], teaches processor, memory, CRM, and instructions.
Regarding claim 6 (Original), Jain et al. discloses the method according to claim 1, wherein the user query is received from a user device associated with a user (Jain et al., fig. 1(100) – user device. “Questions or prompts from a user start here,” Jain et al., para [0023].).
As to claim 13, system claim 13 and method claim 6 are related as method and system of using same, with each claimed element’s function corresponding to the method step. Accordingly claim13 is similarly rejected under the same rationale as applied above with respect to method claim. Also, Jain et al., para [0127] and [0139], teaches processor, memory, CRM, and instructions.
Regarding claim 5 (Currently Amended), Jain et al. discloses the method according to claim 1, wherein the generated embeddings for each asset from amongst the plurality of assets are stored in the database The information retrieval system (110) provides the searchable indexes (116), query logic, and the payload (query response). The various search indexes, including index (116A, 116B . . . 116N) may contain vectors or non-vector content. The indexes (116) are created in advance based on a user defined schema and loaded with content (118) that is sourced from files, databases, or storage,” Jain et al., para [0026].).
As to claim 12, system claim 12 and method claim 5 are related as method and system of using same, with each claimed element’s function corresponding to the method step. Accordingly claim12 is similarly rejected under the same rationale as applied above with respect to method claim. Also, Jain et al., para [0127] and [0139], teaches processor, memory, CRM, and instructions.
Claim 7 canceled.
Claim 14 canceled.
Claim Rejections – 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claim(s) 2 and 9 is/are rejected under 35 U.S.C. 103 as being unpatentable over US 20250181619, hereinafter referred to as Jain et al., in view of EP 4513327, hereinafter referred to as Chen et al.
Regarding claim 2 (Currently Amended), Jain et al. discloses the method according to claim 1, but not wherein the plurality of assets comprises at least one of: a machine learning model, a login module, a software, an object detection model, and the like. Chen et al. is cited to disclose wherein the plurality of assets comprises at least one of: a machine learning model, a login module, a software, an object detection modelOne of the notable capabilities of modern LLMs is that of code generation models to produce responses that include suggestions for software code. By integrating programming code snippets within their training datasets,
these models offer insightful and contextually relevant code suggestions to user queries,” Chen et al., para [0014].). Chen et al. benefits Jain et al. by providing datasets encompassing software code snippets, thereby enhancing the capability of LLMs to understand and generate programming-related content (Chen et al., para [0003]). Therefore, it would be obvious for one skilled in the art to combine the teachings of Jain et al. with those of Chen et al. to improve the LLM versatility of Jain et al.
As to claim 9, system claim 9 and method claim 2 are related as method and system of using same, with each claimed element’s function corresponding to the method step. Accordingly, claim 9 is similarly rejected under the same rationale as applied above with respect to method claim. Also, Jain et al., para [0127] and [0139], teaches processor, memory, CRM, and instructions.
Claim(s) 4 and 11 is/are rejected under 35 U.S.C. 103 as being unpatentable over US 20250181619, hereinafter referred to as Jain et al., in view of US 20250131247, hereinafter referred to as Mondlock et al.
Regarding claim 4 (Currently Amended), Jain et al. discloses the method according to claim 1, wherein the LLM (202) used for implementing the step of generating embeddings for each asset from amongst the plurality of assets, is a text_embedding_ada model. Mondlock et al. is cited to disclose wherein the LLM (202) used for implementing the step of generating embeddings for each asset from amongst the plurality of assets, is a text_embedding_ada model (“Alternatively, the document/asset/expert module 122 may transmit the text chunks and/or data chunks, via the LLM interface module 132, to the LLM service 170 (e.g., using the text-embedding-ada-002 model) and receive embeddings from the LLM service 170,” Mondlock et al., para [0038].). Mondlock et al. benefits Jain et al. by incorporating retrieval-augmented generation (RAG), thereby supplementing user queries with relevant supplied data to enable LLMs to provide improved responses (Mondlock et al., para [0002]). Therefore, it would be obvious for one skilled in the art to combine the teachings of Jain et al. with those of Mondlock et al. to improve the LLM versatility of Jain et al.
As to claim 11, system claim 11 and method claim 4 are related as method and system of using same, with each claimed element’s function corresponding to the method step. Accordingly, claim 11 is similarly rejected under the same rationale as applied above with respect to method claim. Also, Jain et al., para [0127] and [0139], teaches processor, memory, CRM, and instructions.
Conclusion
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 nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to ANNE L THOMAS-HOMESCU whose telephone number is (571)272-0899. The examiner can normally be reached Mon-Fri 8-6.
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, Bhavesh M Mehta can be reached on 5712727453. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/ANNE L THOMAS-HOMESCU/Primary Examiner, Art Unit 2656