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and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114.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 03/04/2026 has been entered.
Response to Arguments
This communication is in response to the Amendments and Arguments filed on 03/04/2026. Claims 1-4, 6-11, 13-18, 20, and 21-23 are pending and have been examined. Claims 1, 2, 4, 6-11, 13-18, and 20 are amended. New claims are 21-23. Cancelled claims are 5, 12 and 19.
Applicant's arguments filed 03/04/2026 are fully considered , but they are not persuasive.
With regard to the 35 U.S.C. 101 rejection, Examiner respectfully disagree with assertion. Please see reasoning below.
Applicant asserts: “prompt with augmentation information to improve the accuracy and/or relevancy of the response generated by the LLM without incurring the huge time and resource costs required to retrain the LLM. See, e.g., Spec. ,-i,i [0016]-[0018]. Furthermore, by combining information in parametric memory and information in non-parametric memory, the LLM can improve the accuracy of the response based on the information in the non-parametric memory. Id The reduced reliance on information in the parametric memory allows for a smaller model..” (Remark page 11-12, last -1st para.)
Examiner Notes: Independent claims does not recite this limitation. Examiner maintained and updated 101 rejection, based on amendment.
Regarding the rejection under 103, applicant’s arguments with respect to claims 1-20 have been considered but are moot because the new ground of rejection does not rely on any reference applied in the prior rejection of record for any teaching or matter specifically challenged in the argument.
Applicant’s assertion: “Applicant respectfully asserts that Neelakantan, Pasupalak, and Tang, taken individually or in combination, fail to teach or suggest a "pre-trained machine learning model configured to prioritize the retrieved piece of augmentation information over information used to train the pretrained machine learning model in the generation of the response," as recited in independent claim 1. Neelakantan relates to performing a query of an embedding space based on semantic similarity between a reference vector representation and a generated vector representation. See, e.g., FIG. 5. Pasupalak relates to a conversation agent that interfaces with one or more domain models to respond to a user. See, e.g., Col. 33, 11. 1-24. Tang relates to retrieval augmented generation. Applicant respectfully asserts that Neelakantan, Pasupalak, and Tang, taken individually or in combination, fail to teach or suggest a model that prioritizes a piece of augmentation information over information used to train the model in the generation of the response.” (Remark page 13, middle paragraph).
Examiner Notes: Examiner has brought new reference for this new limitation, therefore applicant’s arguments are moot. Please see mapping below.
Hence, Applicant’s arguments are not persuasive.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1-4, 6-11, 13-18, 20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more.
Regarding claim(s) 1, 8 and 15, the limitation(s) of “receiving”, “generating”, “comparing”, “retrieving”, “providing”, and “receiving” as drafted, are processes that, under broadest reasonable interpretation, covers performance of the limitation in the mind but for the recitation of generic computer components. More specifically, the mental process of a human. Human can collect question or prompt/text, convert words/phrase into vector representation or assigning number to the text, compare multiple vector representation (near-neighbor) on the vector space and they can be related subpart of other vectors and see comparison result meets threshold requirement, recall previous conversation or some additional information related to conversation/query and responding answer including additional information. In the amended claim, applicant talks about “determine a subset of second feature vectors of the plurality of second feature vectors …” human knows subset and sub-list of feature vectors and can compare them. Human can change/ edit and update data and can even retrieve latest edited data. Human can prioritize information when providing additional information for training.
The claim recites additional element pre-trained machine learning model (i.e. LLM), this model is nothing but computer components. LLM is well-understood, routine, and conventional computer function in view of NEELAKANTAN et al. US 20240249186 A1. (“[0044] … In some embodiments, the converter/encoder 208 may include or may be a Transformer encoder.”) (“[0065] … For example, other pre-trained models 214 may include pre-trained generative language models such as, e.g., Generative Pre-trained Transformer (GPT) models or Codex models (e.g., OpenAI Codex). …”). If a claim limitation, under its broadest reasonable interpretation, covers performance of the limitation in the mind but for the recitation of generic computer components(processor), then it falls within the --Mental Processes-- grouping of abstract ideas. Also, Claim recites training pre-train model with additional data, but missing information about how models are training. Accordingly, the claim(s) recite(s) an abstract idea. The claims are not patent eligible.
This judicial exception is not integrated into a practical application because the recitation of “large language model” and in claim 1, 2, 7, 8, 14 and 15, and “a processor” in claim 8, and “memory” in claim 8, and “storage medium” in claim 15 and 16, reads to generalized computer components, based upon the claim interpretation wherein the structure is interpreted using [0062] and [0069] in the specification. Accordingly, these additional elements do not integrate the abstract idea into a practical application because it does not impose any meaningful limits on practicing the abstract idea. The claim(s) is/are directed to an abstract idea.
The claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception. As discussed above with respect to the integration of the abstract idea into a practical application, the additional element of using generalized computer components to “receiving”, “generating”, “comparing”, “retrieving”, “providing”, and “receiving”, amounts to no more than mere instructions to apply the exception using a generic computer component. Mere instructions to apply an exception using a generic computer component cannot provide an inventive concept. generic computer components (“a large language model”) performing task, MPEP 2106.05(f) and MPEP 2106.05(d) and (f).
Claims 2-7, 9-14, and 16-20 depend from claims 1, 8 and 15 respectively, do not remedy any of the deficiencies of claim 1, 8 and 15 respectively, and therefore are rejected on the same grounds as claim 1 above.
With respect to claim(s) 2, 9 and 16, the claim(s) recite(s) “providing a user interface …”, and “providing the response …” pre and post solution activity. Collecting querry via user interface and displaying output via interface is nothing but pre and post solution activity. The claim recites additional element, “user interface” recites in the claim. Using user interface to present interaction is common use in this area. For example, Pasupalak teaches (“(39) In some contexts, the user input may be a speech input query 302, but responses (output) from the services for presentation to the user by user interface manager 103 on smartphone may be any one or combination of speech (e.g. synthesized automated voice),. …” Col. 30, Lines 44-47) (“(237) In one embodiment, speech service 112 may include general language models 904a,b,c and user-specific language models 902a,b. General language models 904a,b,c may be applicable to all users of the Conversational Agent 150 and ...” Col. 44, lines 41-56 ) by Pasupalak et al. US 10853582 B2
With respect to claim(s) 3, 10, and 17, the claim(s) recite(s) “determining cosine similarities …” between multiple vectors, which reads math algorithm/ math calculation. Therefore, Claims recite abstract idea. No additional elements are present.
With respect to claim(s) 4, 11, and 18, the claim(s) recite(s) “determining the second feature vectors having a cosine similarity to the first feature vector that satisfies…” how close two vectors are and naming them first and second vectors, and each vectors satisfy predetermined threshold value, also detecting maximum/ highest match found in the sub-vector with parent vector. This claim recites math algorithm to decide maximum cosine similarity. Therefore, Claims recite abstract idea. No additional elements are present.
With respect to claim(s) 6, 13, and 20, the claim(s) recite(s) “encode the query into a low-dimensional dense vector …” human can represent text into numbers or vectors. This claim recites additional element GPT based encoder from BERT. This additional element is common to this technical area. GPT based encoder and BERT is well-understood, routine, and conventional computer function in view of NEELAKANTAN et al. US 20240249186 A1. (“[0044] … In some embodiments, the converter/encoder 208 may include or may be a Transformer encoder.”) (“[0065] … For example, other pre-trained models 214 may include pre-trained generative language models such as, e.g., Generative Pre-trained Transformer (GPT) models or Codex models (e.g., OpenAI Codex). …”) Therefore, Claims recite abstract idea. The claims are not patent eligible.
With respect to claim(s) 7, and 14 the claim(s) recite(s) “augmentation information comprises …”, reads on a human various type/kind of additional information can be in specific domain, entity-specific, product specific, or updated information specific/old data or new data- human can detect all of these category’s information by reading text. Human can update/edit/change information/data periodically. Also, human can retrieve latest updated data. Also, the claim recites additional element large language model (LLM), this model is nothing but computer components. This additional element is common to this technical area. GPT based encoder and BERT is well-understood, routine, and conventional computer function in view of NEELAKANTAN et al. US 20240249186 A1. (“[0044] … In some embodiments, the converter/encoder 208 may include or may be a Transformer encoder.”) (“[0065] … For example, other pre-trained models 214 may include pre-trained generative language models such as, e.g., Generative Pre-trained Transformer (GPT) models or Codex models (e.g., OpenAI Codex). …”). If a claim limitation, under its broadest reasonable interpretation, covers performance of the limitation in the mind but for the recitation of generic computer components(processor), then it falls within the --Mental Processes-- grouping of abstract ideas. Accordingly, the claim(s) recite(s) an abstract idea. The claims are not patent eligible.
These claims further do not remedy the judicial exception being integrated into a practical application and further fail to include additional elements that are sufficient to amount to significantly more than the judicial exception. Therefore, the claims are directed to an abstract idea.
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 1, 3, 6, 7, 8, 10, 13, 14, and 15, 17, 20 is/are rejected under 35 U.S.C. 103 as being unpatentable over NEELAKANTAN et al. US 20240249186 A1, in view Araki, US 20220269863 A1 and in view of Syed et al. US 12080428 B1 And further in view of Tang (“Integrating ChatGPT with internal knowledge base and question-answer platform”) {IDS provided}
Regarding Claim 1, NEELAKANTAN teaches:
Claim 1, 8 AND 15 is/are rejected under 35 U.S.C. 103 as being unpatentable over NEELAKANTAN et al. US 20240249186 A1, in view of Tang (“Integrating ChatGPT with internal knowledge base and question-answer platform”) {IDS provided}
Regarding Claim 1, NEELAKANTAN teaches:
1. A computer-implemented method for, comprising:
receiving a query; NEELAKANTAN teaches (“[0016] Embodiments of the present disclosure may also include a system for generating a semantic similarity result, the system including at least one processor configured to receive, via a user interface of a user device, a query for semantic similarity, the query including a natural language input. …”) by NEELAKANTAN et al. US 20240249186 A1
generating a first feature vector based on the query; NEELAKANTAN teaches (“[0077] … . As used herein, a reference vector may include one or more converted reference data values. For example, a received data value including natural language text may be converted to one or more reference vector data values representing a reference vector in a reference vector representation. In some embodiments, the reference data value may be generated based on the natural language input of the query 402 …”) (“[0067] … generating a vector representation may be provided. … … The steps may further include converting the training data set into at least one first vector of a vector representation …”) by NEELAKANTAN et al. US 20240249186 A1
comparing the first feature vector to a plurality of second feature vectors, each of which corresponding to a piece of augmentation information, to determine a subset of second feature vectors of the plurality of second feature vectors that satisfies a predetermined condition with respect to the first feature vector; NEELAKANTAN teaches positive/negative example pair (i.e.subset). at least one reference vector representation(s) 410 may include a plurality of reference vectors (e.g., a matrix), which may be usable for comparisons between vectors, such as vectors included in an embedding space. And at least one processor may be configured to access an embedding space storing a plurality of vector representations. and portions of text data within a predetermined distance threshold of each other. And encode the at least one additional positive/negative example pair to at least one additional vector (i.e. piece of augmentation information )of the vector representation, and include the at least one additional vector in subsequent training data for training the artificial machine learning model. NEELAKANTAN teaches (“[0077] The process 500 may also include a step 530 of transforming, e.g., via at least one hardware processor, the natural language input to a reference vector representation (e.g., reference vector representation(s) 410). A reference vector representation may include any digital representation of multiple vectors, which may be usable to compare with one or more vectors stored in an embedding space. For example, at least one reference vector representation(s) 410 may include a plurality of reference vectors (e.g., a matrix), which may be usable for comparisons between vectors, such as vectors included in an embedding space… … …”) (“[0076] The retrieved data may be usable for a comparison, such as those discussed with respect to step 540. An embedding space may include any data component storing at least one vector representation generated by a machine learning model, as discussed above with respect to embedding space 414.”) by NEELAKANTAN et al. US 20240249186 A1
retrieving the piece of augmentation information corresponding to the determined subset of second feature vectors (“[0076] … Accessing an embedding space (e.g., embedding space 414) may include one or more of querying, requesting, locating, searching, retrieving, or receiving (e.g., with respect to a local or remote storage source). For example, accessing an embedding space may include retrieving data (e.g., values) relating to one or more vectors within at least one vector representation embedded in the embedding space 414. The retrieved data may be usable for a comparison, such as those discussed with respect to step 540. An embedding space may include any data component storing at least one vector representation generated by a machine learning model, as discussed above with respect to embedding space 414.”) by NEELAKANTAN et al. US 20240249186 A1
NEELAKANTAN further teaches accessing embedding space to retrieve data. (“[0076] The process 500 may further include a step 520 of accessing, e.g., via at least one hardware processor, an embedding space storing a plurality of vector representations generated by a machine learning model trained using positive example pairs and negative example pairs, as discussed above with respect to FIG. 4. Accessing an embedding space (e.g., embedding space 414) may include one or more of querying, requesting, locating, searching, retrieving, or receiving (e.g., with respect to a local or remote storage source). For example, accessing an embedding space may include retrieving data (e.g., values) relating to one or more vectors within at least one vector representation embedded in the embedding space 414. The retrieved data may be usable for a comparison, such as those discussed with respect to step 540. An embedding space may include any data component storing at least one vector representation generated by a machine learning model, as discussed above with respect to embedding space 414.”) by NEELAKANTAN et al. US 20240249186 A1
NEELAKANTAN does not explicitly teach the pieces of augmentation information comprising at least one of: recent information unavailable at generation of the large language model, or information changed after generation of the large language model.
Araki teaches:
the pieces of augmentation information comprising at least one of: pre-trained machine learning model, recent information unavailable at ; Araki teaches (“[0021] The new sentences 30 with the proposed labels are displayed to or otherwise provided for review by a domain expert 20 who interacts with the system to verify the new sentences 30 and correct any errors. Particularly, the domain expert 20 may correct the proposed labels for the new sentences 30 or correct grammatical or similar issues in the text of the new sentences 30. Using the verified new sentences 30, one or both of the sentence generator 40 and the sentence classifier 50 are retrained. In this way, the performance of the data augmentation continuously improves, and less and less time is required for the domain expert 20 to verify each newly generated sentence 30.”) (“[0026] … The newly labeled sentences are added to the initial training data and the model is retrained on the newly expanded training data. …”) by Araki, US 20220269863 A1
Araki is considered to be analogous to the claimed invention because it relates that enable rapid and cost-effective human-in-the-loop synthesis of domain-specific textual training data for a deep learning model.
Therefore, it would have been obvious for someone of ordinary skill in the art before the effective filing date of the claimed invention to modify NEELAKANTAN, to incorporate the teachings of Araki in order to include retrieved information as its updated or changed
One could have been motivated to do so because system will may improve the performance of the data augmentation continuously improves, and less and less time is required for the domain expert 20 to verify each newly generated sentence. (“[0021] … , the performance of the data augmentation continuously improves, and less and less time is required for the domain expert 20 to verify each newly generated sentence 30.”)) by Araki, US 20220269863 A1
The combination does not explicitly teach the pre-trained machine learning model configured to prioritize the retrieved piece of augmentation information over information used to train the pre-trained machine learning model in the generation of the response.
Syed teaches:
the pre-trained machine learning model configured to prioritize the retrieved piece of augmentation information over information used to train the pre-trained machine learning model in the generation of the response; Syed teaches (“(29) At Block 114, the system augments data for providers and previously serviced patients and cases. FIG. 3 illustrates data augmentation at Block 114 according to embodiments of the present disclosure. In some embodiments, the system augments the retrieved historical data 312 (e.g., data 112 retrieved in Block 110) by either: at Block 304, replacing or imputing characteristics that were otherwise attempted to be retrieved for patients and cases, but were missing or found to be noisy; or at Block 308, augmenting the characteristics retrieved for providers, patients and cases with additional attributes. For example, the retrieved historical data 312, e.g., corresponding to patient data 102, can be augmented via Block 304 or Block 308 to obtain augmented historical data 316. In some examples, data can be augmented via both Blocks 304 and 308, e.g., by both replacing or imputing characteristics that were otherwise attempted to be retrieved for patients and cases, but were missing or found to be noisy and augmenting the characteristics retrieved for patients and cases with additional attributes. As used herein the process of replacing, imputing, and/or augmenting characteristics can refer generally to augmenting characteristics of the retrieved patient data, e.g., retrieved historical patient data 112, 312.” Col 8, lines 1-25 ) (“(59) At Block 136, the system can augment data for new patient(s) requiring prioritization for non-emergent care and/or services. FIG. 6 illustrates patient data augmentation 636, e.g., corresponding to Block 136. For example, the retrieved patient data 634 corresponding to retrieved patient characteristics and retrieved past care service records can be augmented to obtain augmented patient data 638 that include augmented patient characteristics and augmented past care service records. In one or more examples, the system can augment the data for the new patient in a manner similar to the augmentation of the data for patients as described in Block 114. For example, the system can augment the patient data 634 (e.g., data retrieved in Block 130) by either replacing or imputing characteristics that were otherwise attempted to be retrieved for providers, patients and cases, but were missing or found to be noisy or augmenting the characteristics retrieved for providers, patients and cases with additional attributes.”) (“(72) FIG. 9B illustrates a flow chart of a process 900B for determining a prioritization of care for a patient, according to embodiments of the present disclosure. At Block 912, the system can retrieve patient data. As shown in the figure, the system can retrieve patient data for one or more patients, e.g., Patient 1-N. In some embodiments, Block 912 can correspond to Block 130. At Block 914, the system can input patient data into a first machine-learning model. As shown in the figure, the machine-learning model 922 can receive inputs corresponding to datasets for a plurality of patients, e.g., Patients 1-N. The model can be configured to output a predicted risk value for a corresponding patient. In some examples, model 922 can correspond to model 122, model 422, and/or model 722 described above. In some examples, Block 914 can correspond to Block 140 described above.” Col. 14, lines 31-49) (“(78) FIG. 10 illustrates a flow chart of a process 1000 for training a machine-learning model according to embodiments of this disclosure. In some embodiments process 1000 can correspond to one or more blocks of process 100. At Block 1002, the system can receive patient-specific datasets corresponding to a plurality of patients. For example, the patient data can include patient characteristics and case history, as discussed above. In some examples, Block 1002 can correspond to Block 110 described above. At Block 1004, the system can modify the received patient data to obtain respective augmented patient data. In some examples, Block 1004 can correspond to Block 114 described above. At Block 1006, the system can determine training data based on the augmented patient data. In some examples, the training data can be determined as described above with respect to FIGS. 4B-4D. At Block 1008, the system can train the machine-learning model based on the training data. In some examples, Block 1008 can correspond to Block 120 described above.” Col. 18, lines 21-40) by Syed et al. US 12080428 B1
Syed is considered to be analogous to the claimed invention because it relates generally to machine-learning techniques, and more specifically to machine-learning techniques for prioritizing medical procedures and visits.
Therefore, it would have been obvious for someone of ordinary skill in the art before the effective filing date of the claimed invention to modify NEELAKANTAN, and Araki to incorporate the teachings of Syed in order to include prioritize retrieve information.
One could have been motivated to do so because system provides accurate prediction on prioritization. (“(15) … Embodiments of the present disclosure therefore provides a machine intelligence-based data-driven approach for prioritization that is predictive, accurate, quantitative and deeply personalized and can model the time-varying risk trajectories of patients if they do not receive non-emergent care while dealing with the issues above. , …” col. 5, lines 20-25) by Syed et al. US 12080428 B1
The combination does not explicitly teach augmenting a prompt.
Tang teaches:
augmenting a prompt to a pre-trained machine learningpre-trained machine learning model, an augmented prompt comprising a request for a response to the query based on the retrieved piece of augmentation information, and the retrieved piece of augmentation information, and Tang teaches (“Prompt engineering applied to the case of an internal knowledge base means feeding relevant data from the knowledge base to ChatGPT every time we interact with it. …”) (“This is where Retrieval Augmentation Generation workflow comes in.” page 5, ) (“… the process first uses the user question to perform a search to retrieve relevant documents from the internal dataset and then provides these documents together with the question to ChatGPT, With the additional context, ChatGPT can answer as though it has been trained with the internal dataset.” Page 5, last para. ) by Tang (“Integrating ChatGPT with internal knowledge base and question-answer platform”) {IDS provided}
receiving the response from the pre-trained machine learning model. Tang teaches With the additional context, ChatGPT can answer as though it has been trained with the internal dataset.” Page 5, last para. ) by Tang (“Integrating ChatGPT with internal knowledge base and question-answer platform”) {IDS provided}
Tang is considered to be analogous to the claimed invention because it relates Integrating ChatGPT with internal knowledge base and question-answer platform.
Therefore, it would have been obvious for someone of ordinary skill in the art before the effective filing date of the claimed invention to modify NEELAKANTAN, Araki and Syed, to incorporate the teachings of Tang in order to include augmented input to the LLM.
One could have been motivated to do so because system may improve accuracy and adding multimodal capabilities to support images, videos, speech. (“ … using human-assisted answers to improve accuracy and adding multimodal capabilities to support images, videos, speech.” page 14, second para) by Tang (“Integrating ChatGPT with internal knowledge base and question-answer platform”) {IDS provided}
Claim 8 is a system claim with limitations similar to the limitations of method Claim 1 and is rejected under similar rationale. Additionally,
Regarding Claim 8, NEELAKANTAN teaches:
8. A system for augmenting a large language model, comprising: a processor; a memory device that stores program code structured to cause the processor to: NEELAKANTAN teaches (“[0025] An exemplary operating environment for implementing various aspects of this disclosure is illustrated in FIG. 1. As illustrated in FIG. 1, an exemplary operating environment 100 may include a computing device 102 (e.g., a general-purpose computing device) in the form of a computer. Components of the computing device 102 may include, but are not limited to, various hardware components, such as one or more processors 106, data storage 108, a system memory 104, other hardware 110, …”) by NEELAKANTAN et al. US 20240249186 A1
Claim 15 is a computer-readable storage medium claim with limitations similar to the limitations of method Claim 1 and is rejected under similar rationale. Additionally,
Regarding Claim 15, NEELAKANTAN teaches:
15. A computer-readable storage medium comprising computer-executable instructions, that when executed by a processor, cause the processor to: NEELAKANTAN teaches (“[0030] Computing device 102 includes at least one logical processor 106. The computing device 102, like other suitable devices, also includes one or more computer-readable storage media, which may include, but are not limited to, memory 104 and data storage 108. In some embodiments, memory 104 and data storage 108 may be part of a single memory component. The one or more computer-readable storage media may be of different physical types. …”) by NEELAKANTAN et al. US 20240249186 A1
Regarding Claim 3, the combination teaches the method claim 1 as identified above.
NEELAKANTAN further teaches:
3. The method of claim 1, wherein said comparing the first feature vector to the plurality of second feature vectors comprises: determining cosine similarities between at least a portion of the first feature vector and corresponding portions of the plurality of second feature vectors. NEELAKANTAN teaches (“[0054] … For example, a similarity score may be calculated based on a cosine similarity between a first vector within the vector representation and a second vector within the vector representation, where the first vector may represent the converted or encoded first data unit and the second vector may represent the converted or encoded second data unit. As used herein, a node and a vector may be used interchangeably. …”) by NEELAKANTAN et al. US 20240249186 A1
Claim 10 is a system claim with limitations similar to the limitations of method Claim 3 and is rejected under similar rationale.
Claim 17 is a computer-readable storage medium claim with limitations similar to the limitations of method Claim 3 and is rejected under similar rationale.
Regarding Claim 6, the combination teaches the method claim 1 as identified above
NEELAKANTAN further teaches:
6. The method of claim 1, wherein said generating the first feature vector comprises: encoding the query into a low-dimensional dense vector using a Generative Pre-Trained Transformer (GPT)-based or a Bidirectional Encoder Representations from Transformers (BERT)-based encoder, and NEELAKANTAN teaches [0052] … Converting may include one or more of generating, transforming, embedding, mapping, and/or encoding. For example, converting the training data set into at least one first vector of a vector representation may include mapping one or more portions of text data (e.g., words) from text to numerical (e.g., embedding space) values, such as according to a mapping function. A vector may include a numerical representation of a data unit from a paired data sample, and the numerical representation may be mapped to a high-dimensional space. A vector representation may include or may be part of a matrix, a data structure, an embedding (e.g., a word embedding, a document embedding, an image embedding, an audio embedding, or a code embedding), a representation of one-hot encoded vectors, a latent representation (e.g., low dimensional vectors representing higher dimensional inputs), a high-dimensional space digital representation, or any digital representation of multiple vectors. … … converting the training data set into at least one first vector of a vector representation may include encoding each positive example pair within the training data set (e.g., using a converter/encoder 208). For example, given one or more input sequences comprising data values of an identified positive example pair, (x, y), a converter/encoder 208 may process “x” and “y,” either together or independently, thereby mapping (e.g., digitally mapping) the values “x” and “y” to an embedding space (e.g., a vector representation) and generating a vector for value “x” and a vector for value “y.” In some embodiments, the vectors may share an embedding space of the vector representation (e.g., a vector representation 218) and thereby have a relationship represented by or based on a distance from one another.by NEELAKANTAN et al. US 20240249186 A1
(“[0065] In some embodiments, artificial machine learning model 210 may be initialized with one or more other pre-trained models 214 as input prior to training the artificial machine learning model 210 to generate additional vectors. For example, other pre-trained models 214 may include pre-trained generative language models such as, e.g., Generative Pre-trained Transformer (GPT) models …”) by NEELAKANTAN et al. US 20240249186 A1
wherein said second feature vectors are generated by encoding the pieces of augmentation information into low-dimensional dense vectors using the GPT-based or the BERT-based encoder. NEELAKANTAN teaches (“[0047] Consistent with disclosed embodiments, operating environment 200 including an embedding platform 204 may be further configured to convert, by one or more hardware processors, one or more negative example pairs into one or more second vectors of a vector representation (e.g., one or more second vectors of an embedding). For example, converting one or more negative example pairs into one or more second vectors of a vector representation may include using a converter/encoder 208 to, for example, parse and/or encode an negative example pair, discussed further with respect to step 340. Converting identified negative example pairs into one or more second vectors of a vector representation may include any aspect described herein with respect to converting positive example pairs into at least one first vector.”) (“[0052] … A vector representation may include or may be part of a matrix, a data structure, an embedding (e.g., a word embedding, a document embedding, an image embedding, an audio embedding, or a code embedding), a representation of one-hot encoded vectors, a latent representation (e.g., low dimensional vectors representing higher dimensional inputs), a high-dimensional space digital representation, or any digital representation of multiple vectors. …”) [0063] … and the one or more second vectors of the vector representation. For example, at least one processor 106 may include the at least one first vector and the one or more second vectors in training data (e.g., training data set 202) for training the artificial machine learning model. … ... one additional positive example pair and/or at least one negative example pair from the training data, encode the at least one additional positive/negative example pair to at least one additional vector of the vector representation, and include the at least one additional vector in subsequent training data for training the artificial machine learning model.”) (“[0065] In some embodiments, artificial machine learning model 210 may be initialized with one or more other pre-trained models 214 as input prior to training the artificial machine learning model 210 to generate additional vectors. For example, other pre-trained models 214 may include pre-trained generative language models such as, e.g., Generative Pre-trained Transformer (GPT) models …”) by NEELAKANTAN et al. US 20240249186 A1
Claim 13 is a system claim with limitations similar to the limitations of method Claim 6 and is rejected under similar rationale.
Claim 20 is a computer-readable storage medium claim with limitations similar to the limitations of method Claim 6 and is rejected under similar rationale.
Regarding Claim 7, the combination teaches the method claim 1 as identified above
The combination does not teach domain-specific information.
Syed further teaches:
7. The method of claim 1, wherein the pieces of augmentation information comprise at least one of: domain-specific information; (“(29) At Block 114, the system augments data for providers and previously serviced patients and cases. FIG. 3 illustrates data augmentation at Block 114 according to embodiments of the present disclosure. In some embodiments, the system augments the retrieved historical data 312 (e.g., data 112 retrieved in Block 110) by either: at Block 304, replacing or imputing characteristics that were otherwise attempted to be retrieved for patients and cases, but were missing or found to be noisy; or at Block 308, augmenting the characteristics retrieved for providers, patients and cases with additional attributes. For example, the retrieved historical data 312, e.g., corresponding to patient data 102, can be augmented via Block 304 or Block 308 to obtain augmented historical data 316. In some examples, data can be augmented via both Blocks 304 and 308, e.g., by both replacing or imputing characteristics that were otherwise attempted to be retrieved for patients and cases, but were missing or found to be noisy and augmenting the characteristics retrieved for patients and cases with additional attributes. As used herein the process of replacing, imputing, and/or augmenting characteristics can refer generally to augmenting characteristics of the retrieved patient data, e.g., retrieved historical patient data 112, 312.”) by Syed et al. US 12080428 B1
Syed is considered to be analogous to the claimed invention because it relates generally to machine-learning techniques, and more specifically to machine-learning techniques for prioritizing medical procedures and visits.
Therefore, it would have been obvious for someone of ordinary skill in the art before the effective filing date of the claimed invention to modify NEELAKANTAN, Araki to incorporate the teachings of Syed in order to include prioritize retrieve information.
One could have been motivated to do so because system provides accurate prediction on prioritization. (“(15) … Embodiments of the present disclosure therefore provides a machine intelligence-based data-driven approach for prioritization that is predictive, accurate, quantitative and deeply personalized and can model the time-varying risk trajectories of patients if they do not receive non-emergent care while dealing with the issues above. , …” col. 5, lines 20-25) by Syed et al. US 12080428 B1
Claim 14 is a system claim with limitations similar to the limitations of method Claim 7 and is rejected under similar rationale
Claim 2, 9, and16, is/are rejected under 35 U.S.C. 103 as being unpatentable over NEELAKANTAN, Araki, Syed and Tang and further in view of Mostafazadeh et al. US 20220343903 A1
Regarding Claim 2, the combination teaches the method claim 1 as identified above.
The combination does not explicitly teach pre-trained machine learning model is configured to generate the response based on domain-specific information contained in the retrieved piece of augmentation information.
Mostafazadeh teaches:
2. (Currently Amended) The method of claim 1, further comprising: providing a user interface [[for]]enabling querying of domain-specific information, wherein the query is received from a user through the user interface; and providing the response to the user through the user interface, FIG. 2B, Mostafazadeh teaches (“[0046] Referring now to FIG. 2B, one example implementation of the domain general AI platform or system150 is described. In particular, FIG. 2B illustrates example couplings, signals and data that are passed among the multi-modal immersive content recommender 200, the AI-enabled visualization & reporting engine 202, the natural language AI engine 204, the human-in-the-loop ML module 206, the distributed runtime engine 208, the AI-enabled block-world graphical user interface 210, the function module 212 and the AI-enabled distributed deep semantic data compositor 214. The domain general AI platform 150 addresses the problem of scalability by leveraging transfer learning as the learning paradigm for training of the AI and AI-enabled modules 202, 204, 210 and 214. The domain general AI platform 150 advantageously learns to generalize and scale its various modules, from a particular domain to another, by having each module pre-trained on the most general domain. The pretrained versions of the modules come imbued with fundamental linguistic, visual, world, and commonsense knowledge, which are domain-agnostic. These components are then orchestrated to allow the domain general AI platform or system 150 to collect high-quality targeted data sets in new domains over time and automatically train fine-tuned AI modules for various underlying tasks.”)
wherein the . FIG. 2B, Mostafazadeh teaches (“[0046] … The domain general AI platform 150 advantageously learns to generalize and scale its various modules, from a particular domain to another, by having each module pre-trained on the most general domain. The pretrained versions of the modules come imbued with fundamental linguistic, visual, world, and commonsense knowledge, which are domain-agnostic. These components are then orchestrated to allow the domain general AI platform or system 150 to collect high-quality targeted data sets in new domains over time and automatically train fine-tuned AI modules for various underlying tasks.”) (“[0051] … The query rewriter 403 combines the text from the AI-enabled visualization & reporting engine 202 and the text from the speech recognition module 402 and uses them to generate a new more accurate query that is void of any vocabulary mismatch in the underlying domain. This new query is then provided to the neural QA system 404, the neural semantic parser 406, and the deep information retrieval module 408. In addition to the text query, the neural QA system 404 also receives corpora of curated facts from the AI-enabled Distributed Deep Semantic Data Compositor 214, and teaching actions and instances from the human-in-the loop ML 206. The neural QA system 404 is a retrieval augmented generation model that is trained to dynamically retrieve or generate factual responses to novel queries based on the corpora of curated facts, providing provenance for its decisions. The neural QA system 404 first performs a top-K neural retrieval to find the closest relevant documents to the query based on the indexing of the corpora of curated facts, which is then augmented with the deep representation of the query using an encoder transformer architecture, that is then used for retrieving a part of the document or generating a novel response to the user query, using a decoder transformer architecture. In some implementations, the encoder-decoder transformer architecture may be backpropagating to the neural retriever, and in some implementations, it may be taking the retriever as non-parametric memory. The neural semantic parser 406 is the next module that answers the query based on the existing structured datasets. The neural semantic parser 406 is coupled to receive teaching actions and instances from the human-in-the loop ML 206 and the external data and/or schema from the AI enable distributed deep semantic data compositor 214, in addition to the text query from the query rewriter 403. The neural semantic parser 406 performs deep natural language understanding grounded in the underlying datasets by learning to link the incoming text query tokens to the matches against the schema and particular values present in the datasets. One implementation of the neural semantic parser 406 transforms the linked tagged query into a vector representation using an encoder transformer architecture that is then used for generating the corresponding executable program using a decoder transformer architecture. This module also provides provenance for its decisions by its grounding in the underlying data. The neural semantic parser 406 is capable of generating any domain-specific programming language such as SQL or any general-purpose programming language such as Python, depending on the needs. The deep information retrieval module 408 is coupled to receive teaching actions and instances from the human-in-the loop ML 206, in addition to the text query from the query rewriter 403. The deep information retrieval module 408 performs multimodal similarity search for the query against all the data records available to the system, tuned to have the highest recall and hence the lowest precision. In some implementations, the deep information retrieval module 408 is a Siamese network that learns to represent a textual query close to its most relevant multimodal documents in the high-dimensional space. Hence, given an input query from the user, this module 406 can efficiently retrieve the most relevant data records, which gets sent to the dialog manager 410. The dialogue manager 410 is responsible for managing the state and flow of the interaction with the user, as well as combining the results from the specialized modules 404, 406, and 408. The dialogue manager 410 then provides an output to the natural language generation module 412. The natural language generation module 412 also receives from the distributed runtime engine 208 the execution results from executing the executable program sent by the dialog manager 410. The natural language generation module 412 uses an encoder transformer architecture to encode both the query and a summary of the execution results, and uses a transformer decoder to generate a coherent textual verbalization of the system response. …”) by Mostafazadeh et al. US 20220343903 A1
Mostafazadeh is considered to be analogous to the claimed invention because it relates to domain-general artificial intelligence (AI) platform that enables data-informed decision making in any domain, for anyone, without needing to code.
Therefore, it would have been obvious for someone of ordinary skill in the art before the effective filing date of the claimed invention to modify NEELAKANTAN, Araki, Syed, and Tang to incorporate the teachings of Mostafazadeh in order to include user interface that conversationally interacts with a user.
One could have been motivated to do so because system will may improve over time efficiently interacting with its teacher. [0042] The human-in-the-loop ML module 206 may include software and/or logic to provide the functionality for combining the power of an intelligent human in the loop, as the teacher, with the domain general AI platform or system150 that learns to improve over time through efficiently interacting with its teacher. …”) by Mostafazadeh et al. US 20220343903 A1
Claim 9 is a system claim with limitations similar to the limitations of method Claim 2 and is rejected under similar rationale.
Claim 16 is a computer-readable storage medium claim with limitations similar to the limitations of method Claim 2 and is rejected under similar rationale.
Claim 4, 11, and 18 is/are rejected under 35 U.S.C. 103 as being unpatentable over NEELAKANTAN, Araki, Syed and Tang in view of DOTAN-COHEN et al. US 20230014775 A1 and further in view of Chien et al. US 20230135293 A1
Regarding Claim 4, the combination teaches the method claim 3 as identified above.
The combination does not explicitly teach comparing feature vector detailed example with additional detail.
DOTAN-COHEN teaches:
4. The method of claim 3, wherein said comparing the first feature vector to a plurality of second feature vectors, each of which corresponding to a piece of augmentation information, to determine second feature vectors that satisfy a predetermined condition with respect to the first feature vector comprises: DOTAN-COHEN teaches (“[0074] … For example, the task completion predictor 268 can convert the natural language sentence (and other contextual input) into a first feature vector in feature space and compare against trained word embeddings or second feature vectors that represent individual task candidates and then compute a distance between the feature vectors. If the distance (e.g., Euclidian or Cosine) meets or is within a threshold, then the candidate natural language sentence can be marked as an indication that the corresponding candidate task has been completed. …”) by DOTAN-COHEN et al. US 20230014775 A1
determining the second feature vectors having a cosine similarity to the first feature vector that satisfies a first predetermined relationship with a first predetermined threshold; DOTAN-COHEN teaches (“[0074] … For example, the task completion predictor 268 can convert the natural language sentence (and other contextual input) into a first feature vector in feature space and compare against trained word embeddings or second feature vectors that represent individual task candidates and then compute a distance between the feature vectors. If the distance (e.g., Euclidian or Cosine) meets or is within a threshold, then the candidate natural language sentence can be marked as an indication that the corresponding candidate task has been completed.”) (“[0094] … For example, John may simply attach a file (with no message) named “sports department sales figures May 2021.PDF” Accordingly, the candidate task completion model(s) 314 can then determine the file name (“sports department sales figures May 2021”) and the file type (“PDF”) (for example, the attachment name/type 312 input) and compare this information to what is already known. … …For example, the user metadata input 310 may be “John,” as John is the one responsible for completing the task. Because the contextual data input 306 and the candidate task 308 indicate that John is to send sports department sales numbers to Jane by Friday [May 22.sup.nd, embodiments, such as NLP models (e.g., WORD2VEC) can responsively determine the semantic similarity of words in the file attachment (e.g., “sports department”) to words in the candidate task detection models 302 (e.g., “sports department”), by for example, converting such words into feature vectors and then responsively determining a distance (e.g., a Euclidian or Cosine distance) between the words and/or aggregation of the words (e.g., via a dot product function). And based on the semantic similarity score exceeding or meeting a threshold, the file attachment can be marked as an indication that the candidate task has been completed.”) (“[0095] … and determine a distance to a second feature vector of “[John]: can you send [sports department sales numbers] to [Jane] by Friday [May 22.sup.nd],” which is the candidate task. Based on the distance being within or meeting a distance threshold, the phrase “I just sent you the numbers” can be marked (e.g., highlighted) as an indication that the candidate task has been completed, which is the output 316.”) (“[0098] … For example, in Question Answering systems, the multi-head attention layer 406-1 determines how relevant the ith word (or particular word in a block) is for answering the question or relevant to other words in the same or other blocks, the output of which is an attention vector. For every word, some embodiments generate an attention vector, which captures contextual relationships between other words in the same sentence, block, and or line. For a given word, some embodiments compute a weighted average or otherwise aggregate attention vectors of other words that contain the given word (e.g., other words in the same line or block) to compute a final attention vector. …”) by DOTAN-COHEN et al. US 20230014775 A1
DOTAN-COHEN is considered to be analogous to the claimed invention because it relates user-data collection component 210 receives or accesses user-related data continuously, periodically,
Therefore, it would have been obvious for someone of ordinary skill in the art before the effective filing date of the claimed invention to modify NEELAKANTAN, Araki, Syed, and Tang to incorporate the teachings of DOTAN-COHEN in order to include cosine similarity algorithm.
One could have been motivated to do so because system will may improve existing technologies by automatically detecting indications that tasks have been completed via new logic or rules and improving the functionality. (“Abstract … These systems and methods improve existing technologies by automatically detecting indications that tasks have been completed via new logic or rules and improving the functionality and computing resource consumption relative to existing machine learning models. These systems also improve the way computers operate by reducing computing resource consumption, such as memory, network latency, I/O, and the like.) by DOTAN-COHEN et al. US 20230014775 A1
The combination does not explicitly teach determining a first predetermined number of second feature vectors having highest cosine similarities to the first feature vector;
Chien teaches:
determining a first predetermined number of second feature vectors having highest cosine similarities to the first feature vector; Chien teaches threshold td-idf metric (i.e. first predetermined number). Also, those secondary content items having the highest cosine similarity measures (e.g., the top three, the top five, the top ten, etc.) and/or cosine similarity measures above a threshold (e.g., above a similarity measure above 0.70, above 0.75, etc.) may be selected as being “relevant.” And determining that the at least one second content relates to the at least one assertion in response to the td-idf metric exceeding a threshold td-idf metric. (“[0031] Among the trusted content source(s), server(s) 116 may identify content from content source(s) 119 (e.g., “secondary content”) that may be relevant to the topic of the primary content from content source 114. For example, server(s) 116 may generate a term frequency-inverse document frequency (td-idf) vector representing the primary content (or may access such a vector that may already be generated and stored by content source 114). Server(s) 116 may similarly generate or access td-idf vectors representing respective contents from content source(s) 119, and may compare these td-idf vectors to the td-idf vector representing the primary content, such as by applying a cosine similarity calculation. For example, the td-idf vectors may be generated for the primary content and/or the secondary content in text form, or from text transcripts that may be generated from the primary content and/or the secondary content in non-text form (e.g., video or audio). Those secondary content items having the highest cosine similarity measures (e.g., the top three, the top five, the top ten, etc.) and/or cosine similarity measures above a threshold (e.g., above a similarity measure above 0.70, above 0.75, etc.) may be selected as being “relevant.” It should be noted that other processes of determining relevant content (e.g., similar content) may be applied, such as using different types of feature/vector representations of the respective content, using a different distance/similarity metric, such as Euclidean distance, Jaccard similarity, etc. In one example, server(s) 116 may also apply a semantic enrichment/enhancement process (or context enhancement/enrichment process) to the primary content and/or to the secondary content. For instance, a language database with known synonyms may be used to add additional words or phrases and metadata terms to be included in a td-idf vector generation or similar process.”) (“[0055] In one example, step 340 may include obtaining the at least one second content, calculating a term frequency-inverse document frequency (td-idf) metric between the at least one second content and the first content, and determining that the at least one second content relates to the at least one assertion in response to the tf-idf metric exceeding a threshold td-idf metric. In one example, step 340 may further include generating a td-idf vector for each of the first content and the at least one second content. In one example, the metric may be based upon a cosine similarity or other distance/similarity metrics. In one example, step 340 may include applying a semantic enhancement process to at least one of the first content or the at least one second content to add additional relevant terms to the at least one of the first content or the at least one second content. In such case, the td-idf metric or similar similarity metric may be based at least in part upon the additional relevant.”) by Chien et al. US 20230135293 A1
Chien is considered to be analogous to the claimed invention because it relates generally to machine learning and natural language processing,
Therefore, it would have been obvious for someone of ordinary skill in the art before the effective filing date of the claimed invention to modify NEELAKANTAN Araki, Syed, Tang and DOTAN-COHEN to incorporate the teachings of Chien in order to include content items having the highest cosine similarity measures.
One could have been motivated to do so because computer algorithms whose outputs improve with experience. (“[0002] Machine learning is a subset of artificial intelligence encompassing computer algorithms whose outputs improve with experience. …) by Chien et al. US 20230135293 A1
Claim 11 is a system claim with limitations similar to the limitations of method Claim 4 and is rejected under similar rationale
Claim 18 is a computer-readable storage medium claim with limitations similar to the limitations of method Claim 4 and is rejected under similar rationale.
Claim 21, 22, and 23 is/are rejected under 35 U.S.C. 103 as being unpatentable over NEELAKANTAN, Araki, Syed, Tang and further in view of Lewis et al. (“Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks”)
The combination does not explicitly teach the pre-trained machine learning model is configured to generate the response by combining information in parametric memory and information in non-parametric memory.
Lewis teaches:
21. (New) The method of claim 1, wherein the pre-trained machine learning model is configured to generate the response by combining information in parametric memory and information in non-parametric memory. Lewis teaches (“We endow pre-trained, parametric-memory generation models with a non-parametric memory through a general-purpose fine-tuning approach which we refer to as retrieval-augmented generation (RAG). We build RAG models where the parametric memory is a pre-trained seq2seq transformer, and the non-parametric memory is a dense vector index of Wikipedia, accessed with a pre-trained neural retriever. …”page 2, first paragraph) by Lewis et al. (“Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks”)
Lewis is considered to be analogous to the claimed invention because it relates we bring hybrid parametric and non-parametric memory to the “workhorse of NLP,” i.e. sequence-to-sequence (seq2seq) models.
Therefore, it would have been obvious for someone of ordinary skill in the art before the effective filing date of the claimed invention to modify NEELAKANTAN, Araki, Syed, and Tang to incorporate the teachings of Lewis in order to include combining information in parametric memory and information in non-parametric memory.
One could have been motivated to do so because computer algorithms whose outputs improve with experience. Our work opens up new research directions on how parametric and non-parametric memories interact and how to most effectively combine them, showing promise in being applied to a wide variety of NLP tasks.. …” page 9, Discussion section, last lines) by Lewis et al. (“Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks”)
Claim 12 is a system claim with limitations similar to the limitations of method Claim 21 and is rejected under similar rationale
Claim 13 is a computer-readable storage medium claim with limitations similar to the limitations of method Claim 21 and is rejected under similar rationale.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to FOUZIA HYE SOLAIMAN whose telephone number is (571)270-5656. The examiner can normally be reached M-F (8-5)AM.
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, Paras D. Shah can be reached at (571) 270-1650. 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.
/F.H.S./
Examiner, Art Unit 2653
/Paras D Shah/Supervisory Patent Examiner, Art Unit 2653
04/12/2026
Prosecution Insights