Prosecution Insights
Last updated: November 20, 2026
Application No. 18/668,685

LARGE LANGUAGE MODEL (LLM)-BASED KNOWLEDGE RESOURCE RETRIEVER AND RANKER

Non-Final OA §103§112
Filed
May 20, 2024
Examiner
WITHEY, THEODORE JOHN
Art Unit
2655
Tech Center
2600 — Communications
Assignee
SAP SE
OA Round
3 (Non-Final)
44%
Grant Probability
Moderate
3-4
OA Rounds
6m
Est. Remaining
80%
With Interview

Examiner Intelligence

Grants 44% of resolved cases
44%
Career Allowance Rate
15 granted / 34 resolved
-17.9% vs TC avg
Strong +36% interview lift
Without
With
+36.1%
Interview Lift
resolved cases with interview
Typical timeline
3y 0m
Avg Prosecution
18 currently pending
Career history
73
Total Applications
across all art units

Statute-Specific Performance

§101
18.9%
-21.1% vs TC avg
§103
54.3%
+14.3% vs TC avg
§102
15.0%
-25.0% vs TC avg
§112
10.3%
-29.7% vs TC avg
Black line = Tech Center average estimate • Based on career data from 34 resolved cases

Office Action

§103 §112
DETAILED ACTION This office action is in response to Applicant’s Request for Continued Examination (RCE), received on 06/25/2026. Claims 1, 9, and 17 have been amended. Claims 1-2, 5-10, 13-18, 21-26 are pending and have been considered. 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and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114.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 06/25/2026 has been entered. Response to Arguments Applicant’s arguments, see pgs. 10-12, filed 06/25/2026, with respect to the rejection(s) of claim(s) 1 under 35 U.S.C. 103 (Matson in view of He) have been fully considered and are persuasive. Therefore, the rejection has been withdrawn. However, upon further consideration, a new ground(s) of rejection is made in view of Fei et al. (US-20220318255-A1), hereinafter Fei. Fei discloses “embodiments of two novel retrieval-oriented pretraining tasks are presented herein to further pretrain cross-lingual language models for downstream retrieval tasks, such as cross-lingual ad-hoc retrieval (CUR) and cross-lingual question answering (CLQA)” (abstract). See updated rejections below. Applicant’s arguments with respect to claim(s) 2, 5-8, 10, 13-16, 18, 21-26 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. Claim Rejections - 35 USC § 112 The following is a quotation of 35 U.S.C. 112(b): (b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention. The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph: The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention. Claims 9, 10, 13-16, 23-24 are rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention. Claim 9 recites the limitation "each of the plurality of sentences in the knowledge resources that was randomly masked", “the first predetermined percentage”, “the first masked sentences” in the first element of the “wherein the first transformer-based language model is trained by:…” portion of the claim. There is insufficient antecedent basis for this limitation in the claim. The examiner believes that Applicant may have incorrectly placed the amended portion of the claim to be associated with the first transformer-based language model as opposed to the second transformer-based language model (as seen with independent claims 1 and 17). For further analysis of the claim, the previously recited limitations will be amended to “wherein the first transformer-based language model is trained by: for each sentence of the plurality of sentences in the knowledge resources that is randomly masked based on a first predetermined percentage to generate first masked sentences, randomly masking a second predetermined percentage of second words included in the same sentence to generate second masked sentences.” Claims 10, 13-16, 23-24 are rejected as being dependent upon a rejected base claim. Claim Rejections - 35 USC § 103 In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status. The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action: A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made. Claim(s) 1, 5, 9, 13, 17, 21, 23, and 25 is/are rejected under 35 U.S.C. 103 as being unpatentable over Matson et al. (US-20240419710-A1), hereinafter Matson, in view of He et al. (US-20240362286-A1), hereinafter He, further in view of Fei et al. (US-20220318255-A1), hereinafter Fei. Regarding claim 1, Matson discloses: a computer-implemented method (Abstract, Methods, computer systems, computer-storage media), comprising: receiving a query for knowledge resources from one or more knowledge bases ([0022] each time a search query is issued, the information must be searched for and located at a particular computer storage address of a storage device, [The information being searched, i.e. knowledge resources, of a storage device, i.e. knowledge base]), wherein the query comprises one or more search terms ([0088] generate a text embedding, for example in the form of a vector, to represent the search query (e.g., search query 252), [The ability to generate a text embedding based on a received search query indicates that query to be comprised of one or more search terms, i.e. words]), and wherein the knowledge resources comprise a plurality of sentences ([0059] Using logical breaks can ensure a more coherent separation of content (e.g., not in the middle of a sentence or word). In this regard, content may be separated into portions based on word structure, sentence structure, [Summarizing text (which is also semantically and lexically analyzed) based on separation into sentences indicates the knowledge resources to be comprised of sentences]); obtaining, based on the query, a first set of candidate resources from the one or more knowledge bases having a lexical similarity to the one or more search terms ([Fig. 2, Lexical Search Manager 236], [0094] The lexical search manager 236 is generally configured to manage a lexical search to identify search results related to or relevant to search queries. A lexical search generally refers to a search in which a character string is recognized by search for a sequence of segmented patterns that fits a string in a lexicon, [0099] Based on similarities or matches between a search query (e.g., lexical search query) and lexical search data representing various content items, a set of content items or search results can be identified as relevant to a search query, [0101] search results manager 238 provides search results relevant to a search query, [Identifying content items, i.e. candidate resources, by the lexical search manager to be provided to the search results manager indicates a required obtaining before being provided to the search results manager]); obtaining, based on first embeddings representative of the knowledge resources and a second embedding representative of the one or more search terms of the query ([0019] a text embedding model may be used to produce text embeddings for content. When a search query is obtained, the text embedding model produces a text embedding for the search query and, thereafter, the search query text embedding and the content text embeddings are compared), a second set of candidate resources from the one or more knowledge bases having a semantic similarity to the one or more search terms ([Fig. 2, Semantic Search Manager 234], [0086] The semantic search manager 234 is generally configured to perform semantic searches. As described, using a semantic search enables a search with semantic meaning, thereby enhancing search accuracy by understanding a searcher's intent and contextual meaning of terms as they appear in a searchable dataspace, [0102] the search results manager 238 aggregates or integrates search results identified via the semantic search manager 234, [Aggregating semantic search results using the search results manager indicates a required obtaining of search results, i.e. candidate resources, to be provided to the search results manager for aggregation/integration]). Matson does not disclose: wherein the first embeddings and the second embedding are generated by a first transformer-based language model and a second transformer-based language model, wherein the first transformer-based language model is trained by: for each sentence of the plurality of sentences in the knowledge resources, randomly masking a first predetermined percentage of first words included in the sentence to generate first masked sentences; and providing the first masked sentences to the first transformer-based language model, wherein the first transformer-based language model is configured to predict the masked first words based on unmasked words included in the sentences and generate a sentence embedding for each of the first masked sentences. He discloses: wherein the first embeddings and the second embeddings are generated by a first transformer-based language model ([0121] the ML model 312 may comprise a search model 704 trained to transform document content contained within an electronic document 706 into semantically searchable document content. For example, the search model 704 may implement an artificial neural network 400, such as a recurrent neural network (RNN) for an Embeddings from Language Models (ELMo), Bidirectional Encoder Representations from Transformers (BERT), a BERT variant, and so forth) and a second transformer-based language model ([0121] the ML model 312 may comprise a generative AI model 728 to implement generative AI techniques to assist in summarizing some or all of the search results in a natural language such as a human language for better readability and understanding by a human reader. For example, the generative AI model 728 may implement a language model such as a generative pre-trained transformer (GPT) language model, among others, [Wherein the search model 704 comprised of BERT and generative AI model 728 comprised of GPT of Fig. 7 correspond to respective first and second transformer-based language models]), wherein the first transformer-based language model is trained by: for each sentence of the plurality of sentences in the knowledge resources, randomly masking a first predetermined percentage of first words included in the sentence to generate first masked sentences ([0126] BERT is pre-trained on a large corpus of unlabeled text data using two objectives: (1) masked language modeling; and (2) next sentence prediction. During the masked language modeling objective, a certain percentage of the input tokens are randomly masked); and, providing the first masked sentences to the first transformer-based language model ([0126] encourages the model to learn meaningful representations for each token that take into account the context in which it appears, [The ability of the model to learn meaningful representations for tokens indicates the necessary reception of the tokens to be learned by the model]), wherein the first transformer-based language model is configured to predict the masked first words based on unmasked words included in the sentences and generate a sentence embedding for each of the first masked sentences ([0126] the model is trained to predict the original values of the masked tokens based on their surrounding context, [0138] the search manager 124 may use the search model 704 to encode a search query 144 using BERT to generate a contextualized embedding for the search query 144, [Predicting masked tokens based on context indicates the context to be unmasked tokens]). Matson and He are considered analogous art within semantic and lexical parsing of input queries. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the teachings of Matson to incorporate the teachings of He, because of the novel way to introduce advances to intra-document search capabilities (consisting of at least lexical and semantic searching, see [0036]) that return relevant results based on semantic and lexical similarity while creating abstractive summaries of relevant snippets, improving the search experience over current keyword based lexicon search (He, [0032]-[0036]). Matson in view of He does not disclose: wherein the second transformer-based language model is trained by: for each sentence of the plurality of sentences in the knowledge resources that was randomly masked based on the first predetermined percentage to generate the first masked sentences, randomly masking a second predetermined percentage of second words included in the same sentence to generate second masked sentences. Fei discloses: wherein the second transformer-based language model ([0091] pretraining a cross-lingual language model with global+sliding window attention using a query language modeling task, according to embodiments of the present disclosure. Given a pair of cross-lingual query 510 and document 520, a percentage of query tokens 515 are masked (605), which may be masked at random [In view of the first and second transformer-based language models of He, a cross-lingual LM containing transformers (see Fig. 2) indicates the cross-lingual LM to be a second transformer-based model]) is trained by: for each sentence of the plurality of sentences in the knowledge resources that was randomly masked based on the first predetermined percentage to generate the first masked sentences ([Fig. 8, S805 “Sample a batch of cross-lingual query and document pairs…”], [0011] sampling a batch of cross-lingual query and document pairs, in which the relevance ranking batch comprises (1) at least one positive sample cross-lingual query and document pair, in which the query and the document are in different languages and the document is relevant to the query and (2) at least one negative sample cross-lingual query and document pair, in which the query and the document are in different languages and the document is not relevant to the query, [A batch of queries/documents indicates a plurality of knowledge resources in view of the previously disclosed knowledge resources of Matson which were masked using the first method of He]), randomly masking a second predetermined percentage of second words included in the same sentence to generate second masked sentences ([0150] Masking query with 30% probability was regarded as a baseline. Two model embodiments were trained with different query masking probabilities (15% and 45%). The other two model embodiments were trained with a masking probability of 30%, [Different model embodiments (without disclosing different input data) indicates the same set of data applied to different masking probabilities/embodiments]). Matson, He, and Fei are considered analogous art within masked text prediction. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the teachings of Matson in view of He to incorporate the teachings of Fei, because of the novel way to develop transformers which are capable of accepting longer input text sequences, significantly improving upon general multilingual language model in both cross-lingual retrieval and cross-lingual transfer (Fei, [0009]). He further discloses: providing the sentence embedding for each of the first masked sentences to the second transformer-based language model ([0147] generative AI model 728 uses a large language module (LLM) to assist in summarizing the search results 146 to produce an Abstractive summary 148. The generative AI may provide an Abstractive summary 148 of the search results 146 relevant to a given search query 144, [Generating an Abstractive summary based on search results, i.e. embeddings of first previously masked sentences gathered using BERT search model 704, indicates the AI model, i.e. second transformer-based model, must necessarily be provided the sentence embeddings for the search results to know what to generate/vary summaries for]), wherein the second transformer-based language model is configured to predict the masked second words of the second masked sentences based on the sentence embedding for each of the first masked sentences ([0122] search manager 124 may prepare a prompt with both the search query 144 and some or all of the search results 146 (e.g., the top k sections) from the electronic document 706, and send it to the generative AI model 728 to create an Abstractive summary 148, [Generating an abstractive summary from search results, i.e. those gathered through the masking operation performed by search model 704, see previously cited [0121], wherein the AI model generating the summary is a GPT model (necessarily performing masking operations), indicting the generated summary is a prediction of masked words based on the sentence embeddings, i.e. search results, from first masked sentences, i.e. the search results]); and, for each of the first set of candidate resources and the second set of candidate resources, determining a level of confidence indicating a relevance of the candidate resource to the query ([0070] The search manager 124 may implement a set of improved search tools and algorithms to perform lexical searching, semantic searching, or a combination of both, [0138] search model 704 can compute the attention scores between the query embedding and each document token embedding using the dot product or the cosine similarity, [0139] The search model 704 may normalize the scores using a softmax function. Softmax is a mathematical function that is commonly used in machine learning and neural networks. It is used to convert a vector of real numbers into a probability distribution, where each element in the vector represents the probability of a particular outcome, [0143] the search manager 124 may use the lexical search generator 732 to perform a lexical full-text search to produce and rank a first set of search results 146, [0214] Another example is the Latent Dirichlet Allocation (LDA) algorithm, which is a probabilistic model that identifies topics within a corpus and assigns probabilities to documents based on their relevance to these topics, [Performing ranking based on a normalized probability distribution, wherein the probability relates to a similarity, indicates the probability to be a confidence as related to being able to solve the input query for lexical search results in view of the LDA algorithm applied to semantic search results. Further, generation of a ranked list necessarily has a related “confidence” to the ordering of the list in terms of correctness]). Matson (in view of the confidence score related to relevance of He) further discloses: ranking the first set of candidate resources and the second set of candidate resources based on at least the level of confidence determined for each of the first set of candidate resources and the second set of candidate resources to generate a ranked list of candidate resources ([0093] In embodiments, the search results may correspond with a weight or a rank indicating an extent of relevance or relatedness to the search query. For instance, a semantic similarity or distance may be used to identify or determine a weight or rank associated with a search result, [0099] the search results may correspond with a weight or a rank indicating an extent of relevance or relatedness to the search query. For instance, a lexical similarity or distance may be used to identify or determine a weight or rank associated with a search result, [Wherein a similarity/relevance/relatedness in the context of Matson will clearly correspond to a confidence of relevance as disclosed in the instant application in view of the probabilities, i.e. confidences, of He]); and, providing a response to the query comprising at least a subset of the ranked list candidate resources to a graphical user interface for display thereby ([0113] a comprehensive set of search results is provided in response to the search query. The comprehensive search results can include a combination of the first set of search results and the second set of search results. In this way, the comprehensive set of search results includes search results generated from performing a semantic search and performing a lexical search. In some cases, the search results from the semantic search and lexical search may be interleaved with one another based on ranking of relevance to the search query). Regarding claim 5, Matson in view of He, further in view of Fei discloses: the computer-implemented method of claim 1. He further discloses: wherein the second predetermined percentage is greater than the first predetermined percentage ([0138] the search manager 124 may use the search model 704 to encode a search query 144 using BERT to generate a contextualized embedding for the search query 144, [0147] The search manager 124 may implement or access a generative AI model 728 for the model inferencer 208. The generative AI model 728 uses a large language module (LLM) to assist in summarizing the search results 146 to produce an Abstractive summary 148, [Using masking in the context of BERT and generative AI necessarily indicates the percentage of masks of a generative task, i.e. to generate a summary, to be greater than that used in BERT, i.e. single/span token masking. Consider the differences in size between generating a sentence embedding of a search query ([Fig. 14, 1406[Wingdings font/0xE0]1408]) against prompting generation of an entire summary of multiple sentences based on multiple search results (see [Fig. 10]) to be generated ([Fig. 14, 1414[Wingdings font/0xE0]1416])]). Regarding claim 9, Matson discloses: a system (Abstract, Methods, computer systems, computer-storage media), comprising: a memory ([Fig. 9, Memory 912]); and at least one processor coupled to the memory ([Fig. 9, Processor(s) 914 coupled to memory 912 through bus 910]) and configured to: receive a query for knowledge resources from one or more knowledge bases ([0022] each time a search query is issued, the information must be searched for and located at a particular computer storage address of a storage device, [The information being searched, i.e. knowledge resources, of a storage device, i.e. knowledge base]), wherein the query comprises one or more search terms ([0088] generate a text embedding, for example in the form of a vector, to represent the search query (e.g., search query 252), [The ability to generate a text embedding based on a received search query indicates that query to be comprised of one or more search terms, i.e. words]), and wherein the knowledge resources comprise a plurality of sentences ([0059] Using logical breaks can ensure a more coherent separation of content (e.g., not in the middle of a sentence or word). In this regard, content may be separated into portions based on word structure, sentence structure, [Summarizing text (which is also semantically and lexically analyzed) based on separation into sentences indicates the knowledge resources to be comprised of sentences]); obtain, based on the query, a first set of candidate resources from the one or more knowledge bases having a lexical similarity to the one or more search terms ([Fig. 2, Lexical Search Manager 236], [0094] The lexical search manager 236 is generally configured to manage a lexical search to identify search results related to or relevant to search queries. A lexical search generally refers to a search in which a character string is recognized by search for a sequence of segmented patterns that fits a string in a lexicon, [0099] Based on similarities or matches between a search query (e.g., lexical search query) and lexical search data representing various content items, a set of content items or search results can be identified as relevant to a search query, [0101] search results manager 238 provides search results relevant to a search query, [Identifying content items, i.e. candidate resources, by the lexical search manager to be provided to the search results manager indicates a required obtaining before being provided to the search results manager]); obtain, based on first embeddings representative of the knowledge resources and a second embedding representative of the one or more search terms of the query ([0019] a text embedding model may be used to produce text embeddings for content. When a search query is obtained, the text embedding model produces a text embedding for the search query and, thereafter, the search query text embedding and the content text embeddings are compared), a second set of candidate resources from the one or more knowledge bases having a semantic similarity to the one or more search terms ([Fig. 2, Semantic Search Manager 234], [0086] The semantic search manager 234 is generally configured to perform semantic searches. As described, using a semantic search enables a search with semantic meaning, thereby enhancing search accuracy by understanding a searcher's intent and contextual meaning of terms as they appear in a searchable dataspace, [0102] the search results manager 238 aggregates or integrates search results identified via the semantic search manager 234, [Aggregating semantic search results using the search results manager indicates a required obtaining of search results, i.e. candidate resources, to be provided to the search results manager for aggregation/integration]). Matson does not disclose: wherein the first embeddings and the second embedding are generated by a first transformer-based language model and a second transformer-based language model. He discloses: wherein the first embeddings and the second embeddings are generated by a first transformer-based language model ([0121] the ML model 312 may comprise a search model 704 trained to transform document content contained within an electronic document 706 into semantically searchable document content. For example, the search model 704 may implement an artificial neural network 400, such as a recurrent neural network (RNN) for an Embeddings from Language Models (ELMo), Bidirectional Encoder Representations from Transformers (BERT), a BERT variant, and so forth) and a second transformer-based language model ([0121] the ML model 312 may comprise a generative AI model 728 to implement generative AI techniques to assist in summarizing some or all of the search results in a natural language such as a human language for better readability and understanding by a human reader. For example, the generative AI model 728 may implement a language model such as a generative pre-trained transformer (GPT) language model, among others, [Wherein the search model 704 comprised of BERT and generative AI model 728 comprised of GPT of Fig. 7 correspond to respective first and second transformer-based language models]). Matson and He are considered analogous art within semantic and lexical parsing of input queries. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the teachings of Matson to incorporate the teachings of He, because of the novel way to introduce advances to intra-document search capabilities (consisting of at least lexical and semantic searching, see [0036]) that return relevant results based on semantic and lexical similarity while creating abstractive summaries of relevant snippets, improving the search experience over current keyword based lexicon search (He, [0032]-[0036]). Matson in view of He does not disclose: wherein the first transformer-based language model is trained by: for each sentence of the plurality of sentences in the knowledge resources that is randomly masked based on a first predetermined percentage to generate first masked sentences, randomly masking a second predetermined percentage of second words included in the same sentence to generate second masked sentences. Fei discloses: wherein the first transformer-based language model is trained by: for each sentence of the plurality of sentences in the knowledge resources that is randomly masked based on a first predetermined percentage to generate first masked sentences ([Fig. 8, S805 “Sample a batch of cross-lingual query and document pairs…”], [0011] sampling a batch of cross-lingual query and document pairs, in which the relevance ranking batch comprises (1) at least one positive sample cross-lingual query and document pair, in which the query and the document are in different languages and the document is relevant to the query and (2) at least one negative sample cross-lingual query and document pair, in which the query and the document are in different languages and the document is not relevant to the query, [A batch of queries/documents indicates a plurality of knowledge resources in view of the previously disclosed knowledge resources of Matson in view of the masking of Fei cited below, indicating at least first and second masking probabilities for first and second masked sentences]), randomly masking a second predetermined percentage of second words included in the same sentence to generate second masked sentences ([0150] Masking query with 30% probability was regarded as a baseline. Two model embodiments were trained with different query masking probabilities (15% and 45%). The other two model embodiments were trained with a masking probability of 30%, [Different model embodiments (without disclosing different input data) indicates the same set of data applied to different masking probabilities/embodiments]). Matson, He, and Fei are considered analogous art within masked text prediction. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the teachings of Matson in view of He to incorporate the teachings of Fei, because of the novel way to develop transformers which are capable of accepting longer input text sequences, significantly improving upon general multilingual language model in both cross-lingual retrieval and cross-lingual transfer (Fei, [0009]). He further discloses: providing the first masked sentences to the first transformer-based language model ([0126] encourages the model to learn meaningful representations for each token that take into account the context in which it appears, [The ability of the model to learn meaningful representations for tokens indicates the necessary reception of the tokens to be learned by the model]), wherein the first transformer-based language model is configured to predict the masked first words based on unmasked words included in the sentences and generate a sentence embedding for each of the first masked sentences ([0126] the model is trained to predict the original values of the masked tokens based on their surrounding context, [0138] the search manager 124 may use the search model 704 to encode a search query 144 using BERT to generate a contextualized embedding for the search query 144, [Predicting masked tokens based on context indicates the context to be unmasked tokens]), and for each of the first set of candidate resources and the second set of candidate resources, determine a level of confidence indicating a relevance of the candidate resource to the query ([0070] The search manager 124 may implement a set of improved search tools and algorithms to perform lexical searching, semantic searching, or a combination of both, [0138] search model 704 can compute the attention scores between the query embedding and each document token embedding using the dot product or the cosine similarity, [0139] The search model 704 may normalize the scores using a softmax function. Softmax is a mathematical function that is commonly used in machine learning and neural networks. It is used to convert a vector of real numbers into a probability distribution, where each element in the vector represents the probability of a particular outcome, [0143] the search manager 124 may use the lexical search generator 732 to perform a lexical full-text search to produce and rank a first set of search results 146, [0214] Another example is the Latent Dirichlet Allocation (LDA) algorithm, which is a probabilistic model that identifies topics within a corpus and assigns probabilities to documents based on their relevance to these topics, [Performing ranking based on a normalized probability distribution, wherein the probability relates to a similarity, indicates the probability to be a confidence as related to being able to solve the input query for lexical search results in view of the LDA algorithm applied to semantic search results. Further, generation of a ranked list necessarily has a related “confidence” to the ordering of the list in terms of correctness]). Matson (in view of the confidence score related to relevance of He) further discloses: rank the first set of candidate resources and the second set of candidate resources based on at least the level of confidence determined for each of the first set of candidate resources and the second set of candidate resources to generate a ranked list of candidate resources ([0093] In embodiments, the search results may correspond with a weight or a rank indicating an extent of relevance or relatedness to the search query. For instance, a semantic similarity or distance may be used to identify or determine a weight or rank associated with a search result, [0099] the search results may correspond with a weight or a rank indicating an extent of relevance or relatedness to the search query. For instance, a lexical similarity or distance may be used to identify or determine a weight or rank associated with a search result, [Wherein a similarity/relevance/relatedness in the context of Matson will clearly correspond to a confidence of relevance as disclosed in the instant application in view of the probabilities, i.e. confidences, of He]); and, provide a response to the query comprising at least a subset of the ranked list candidate resources to a graphical user interface for display thereby ([0113] a comprehensive set of search results is provided in response to the search query. The comprehensive search results can include a combination of the first set of search results and the second set of search results. In this way, the comprehensive set of search results includes search results generated from performing a semantic search and performing a lexical search. In some cases, the search results from the semantic search and lexical search may be interleaved with one another based on ranking of relevance to the search query). Regarding claim 13, Matson in view of He, further in view of Fei, further in view of Farmer discloses: the system of claim 9. He further discloses: wherein the second predetermined percentage is greater than the first predetermined percentage ([0138] the search manager 124 may use the search model 704 to encode a search query 144 using BERT to generate a contextualized embedding for the search query 144, [0147] The search manager 124 may implement or access a generative AI model 728 for the model inferencer 208. The generative AI model 728 uses a large language module (LLM) to assist in summarizing the search results 146 to produce an Abstractive summary 148, [Using masking in the context of BERT and generative AI necessarily indicates the percentage of masks of a generative task, i.e. to generate a summary, to be greater than that used in BERT, i.e. single/span token masking. Consider the differences in size between generating a sentence embedding of a search query ([Fig. 14, 1406[Wingdings font/0xE0]1408]) against prompting generation of an entire summary of multiple sentences based on multiple search results (see [Fig. 10]) to be generated ([Fig. 14, 1414[Wingdings font/0xE0]1416])]). Regarding claim 17, Matson discloses: a non-transitory computer-readable device having instructions stored thereon that ([0146] Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer-readable instructions), when executed by at least one computing device ([0146] Computing device 900 typically includes a variety of computer-readable media. Computer-readable media can be any available media that can be accessed by computing device 900 and includes both volatile and nonvolatile), causes the at least one computing device to perform operations, the operations comprising: receiving a query for knowledge resources from one or more knowledge bases ([0022] each time a search query is issued, the information must be searched for and located at a particular computer storage address of a storage device, [The information being searched, i.e. knowledge resources, of a storage device, i.e. knowledge base]), wherein the query comprises one or more search terms ([0088] generate a text embedding, for example in the form of a vector, to represent the search query (e.g., search query 252), [The ability to generate a text embedding based on a received search query indicates that query to be comprised of one or more search terms, i.e. words]), and wherein the knowledge resources comprise a plurality of sentences ([0059] Using logical breaks can ensure a more coherent separation of content (e.g., not in the middle of a sentence or word). In this regard, content may be separated into portions based on word structure, sentence structure, [Summarizing text (which is also semantically and lexically analyzed) based on separation into sentences indicates the knowledge resources to be comprised of sentences]); obtaining, based on the query, a first set of candidate resources from the one or more knowledge bases having a lexical similarity to the one or more search terms ([Fig. 2, Lexical Search Manager 236], [0094] The lexical search manager 236 is generally configured to manage a lexical search to identify search results related to or relevant to search queries. A lexical search generally refers to a search in which a character string is recognized by search for a sequence of segmented patterns that fits a string in a lexicon, [0099] Based on similarities or matches between a search query (e.g., lexical search query) and lexical search data representing various content items, a set of content items or search results can be identified as relevant to a search query, [0101] search results manager 238 provides search results relevant to a search query, [Identifying content items, i.e. candidate resources, by the lexical search manager to be provided to the search results manager indicates a required obtaining before being provided to the search results manager]); obtaining, based on first embeddings representative of the knowledge resources and a second embedding representative of the one or more search terms of the query ([0019] a text embedding model may be used to produce text embeddings for content. When a search query is obtained, the text embedding model produces a text embedding for the search query and, thereafter, the search query text embedding and the content text embeddings are compared), a second set of candidate resources from the one or more knowledge bases having a semantic similarity to the one or more search terms ([Fig. 2, Semantic Search Manager 234], [0086] The semantic search manager 234 is generally configured to perform semantic searches. As described, using a semantic search enables a search with semantic meaning, thereby enhancing search accuracy by understanding a searcher's intent and contextual meaning of terms as they appear in a searchable dataspace, [0102] the search results manager 238 aggregates or integrates search results identified via the semantic search manager 234, [Aggregating semantic search results using the search results manager indicates a required obtaining of search results, i.e. candidate resources, to be provided to the search results manager for aggregation/integration]). Matson does not disclose: wherein the first embeddings and the second embedding are generated by a first transformer-based language model and a second transformer-based language model, wherein the first transformer-based language model is trained by: for each sentence of the plurality of sentences in the knowledge resources, randomly masking a first predetermined percentage of first words included in the sentence to generate first masked sentences; and providing the first masked sentences to the first transformer-based language model, wherein the first transformer-based language model is configured to predict the masked first words based on unmasked words included in the sentences and generate a sentence embedding for each of the first masked sentences. He discloses: wherein the first embeddings and the second embeddings are generated by a first transformer-based language model ([0121] the ML model 312 may comprise a search model 704 trained to transform document content contained within an electronic document 706 into semantically searchable document content. For example, the search model 704 may implement an artificial neural network 400, such as a recurrent neural network (RNN) for an Embeddings from Language Models (ELMo), Bidirectional Encoder Representations from Transformers (BERT), a BERT variant, and so forth) and a second transformer-based language model ([0121] the ML model 312 may comprise a generative AI model 728 to implement generative AI techniques to assist in summarizing some or all of the search results in a natural language such as a human language for better readability and understanding by a human reader. For example, the generative AI model 728 may implement a language model such as a generative pre-trained transformer (GPT) language model, among others, [Wherein the search model 704 comprised of BERT and generative AI model 728 comprised of GPT of Fig. 7 correspond to respective first and second transformer-based language models]), wherein the first transformer-based language model is trained by: for each sentence of the plurality of sentences in the knowledge resources, randomly masking a first predetermined percentage of first words included in the sentence to generate first masked sentences ([0126] BERT is pre-trained on a large corpus of unlabeled text data using two objectives: (1) masked language modeling; and (2) next sentence prediction. During the masked language modeling objective, a certain percentage of the input tokens are randomly masked); and, providing the first masked sentences to the first transformer-based language model ([0126] encourages the model to learn meaningful representations for each token that take into account the context in which it appears, [The ability of the model to learn meaningful representations for tokens indicates the necessary reception of the tokens to be learned by the model]), wherein the first transformer-based language model is configured to predict the masked first words based on unmasked words included in the sentences and generate a sentence embedding for each of the first masked sentences ([0126] the model is trained to predict the original values of the masked tokens based on their surrounding context, [0138] the search manager 124 may use the search model 704 to encode a search query 144 using BERT to generate a contextualized embedding for the search query 144, [Predicting masked tokens based on context indicates the context to be unmasked tokens]). Matson and He are considered analogous art within semantic and lexical parsing of input queries. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the teachings of Matson to incorporate the teachings of He, because of the novel way to introduce advances to intra-document search capabilities (consisting of at least lexical and semantic searching, see [0036]) that return relevant results based on semantic and lexical similarity while creating abstractive summaries of relevant snippets, improving the search experience over current keyword based lexicon search (He, [0032]-[0036]). Matson in view of He does not disclose: wherein the second transformer-based language model is trained by: for each sentence of the plurality of sentences in the knowledge resources that was randomly masked based on the first predetermined percentage to generate the first masked sentences, randomly masking a second predetermined percentage of second words included in the same sentence to generate second masked sentences. Fei discloses: wherein the second transformer-based language model ([0091] pretraining a cross-lingual language model with global+sliding window attention using a query language modeling task, according to embodiments of the present disclosure. Given a pair of cross-lingual query 510 and document 520, a percentage of query tokens 515 are masked (605), which may be masked at random [In view of the first and second transformer-based language models of He, a cross-lingual LM containing transformers (see Fig. 2) indicates the cross-lingual LM to be a second transformer-based model]) is trained by: for each sentence of the plurality of sentences in the knowledge resources that was randomly masked based on the first predetermined percentage to generate the first masked sentences ([Fig. 8, S805 “Sample a batch of cross-lingual query and document pairs…”], [0011] sampling a batch of cross-lingual query and document pairs, in which the relevance ranking batch comprises (1) at least one positive sample cross-lingual query and document pair, in which the query and the document are in different languages and the document is relevant to the query and (2) at least one negative sample cross-lingual query and document pair, in which the query and the document are in different languages and the document is not relevant to the query, [A batch of queries/documents indicates a plurality of knowledge resources in view of the previously disclosed knowledge resources of Matson which were masked using the first method of He]), randomly masking a second predetermined percentage of second words included in the same sentence to generate second masked sentences ([0150] Masking query with 30% probability was regarded as a baseline. Two model embodiments were trained with different query masking probabilities (15% and 45%). The other two model embodiments were trained with a masking probability of 30%, [Different model embodiments (without disclosing different input data) indicates the same set of data applied to different masking probabilities/embodiments]). Matson, He, and Fei are considered analogous art within masked text prediction. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the teachings of Matson in view of He to incorporate the teachings of Fei, because of the novel way to develop transformers which are capable of accepting longer input text sequences, significantly improving upon general multilingual language model in both cross-lingual retrieval and cross-lingual transfer (Fei, [0009]). He further discloses: providing the sentence embedding for each of the first masked sentences to the second transformer-based language model ([0147] generative AI model 728 uses a large language module (LLM) to assist in summarizing the search results 146 to produce an Abstractive summary 148. The generative AI may provide an Abstractive summary 148 of the search results 146 relevant to a given search query 144, [Generating an Abstractive summary based on search results, i.e. embeddings of first previously masked sentences gathered using BERT search model 704, indicates the AI model, i.e. second transformer-based model, must necessarily be provided the sentence embeddings for the search results to know what to generate/vary summaries for]), wherein the second transformer-based language model is configured to predict the masked second words of the second masked sentences based on the sentence embedding for each of the first masked sentences ([0122] search manager 124 may prepare a prompt with both the search query 144 and some or all of the search results 146 (e.g., the top k sections) from the electronic document 706, and send it to the generative AI model 728 to create an Abstractive summary 148, [Generating an abstractive summary from search results, i.e. those gathered through the masking operation performed by search model 704, see previously cited [0121], wherein the AI model generating the summary is a GPT model (necessarily performing masking operations), indicting the generated summary is a prediction of masked words based on the sentence embeddings, i.e. search results, from first masked sentences, i.e. the search results]); and, for each of the first set of candidate resources and the second set of candidate resources, determining a level of confidence indicating a relevance of the candidate resource to the query ([0070] The search manager 124 may implement a set of improved search tools and algorithms to perform lexical searching, semantic searching, or a combination of both, [0138] search model 704 can compute the attention scores between the query embedding and each document token embedding using the dot product or the cosine similarity, [0139] The search model 704 may normalize the scores using a softmax function. Softmax is a mathematical function that is commonly used in machine learning and neural networks. It is used to convert a vector of real numbers into a probability distribution, where each element in the vector represents the probability of a particular outcome, [0143] the search manager 124 may use the lexical search generator 732 to perform a lexical full-text search to produce and rank a first set of search results 146, [0214] Another example is the Latent Dirichlet Allocation (LDA) algorithm, which is a probabilistic model that identifies topics within a corpus and assigns probabilities to documents based on their relevance to these topics, [Performing ranking based on a normalized probability distribution, wherein the probability relates to a similarity, indicates the probability to be a confidence as related to being able to solve the input query for lexical search results in view of the LDA algorithm applied to semantic search results. Further, generation of a ranked list necessarily has a related “confidence” to the ordering of the list in terms of correctness]). Matson (in view of the confidence score related to relevance of He) further discloses: ranking the first set of candidate resources and the second set of candidate resources based on at least the level of confidence determined for each of the first set of candidate resources and the second set of candidate resources to generate a ranked list of candidate resources ([0093] In embodiments, the search results may correspond with a weight or a rank indicating an extent of relevance or relatedness to the search query. For instance, a semantic similarity or distance may be used to identify or determine a weight or rank associated with a search result, [0099] the search results may correspond with a weight or a rank indicating an extent of relevance or relatedness to the search query. For instance, a lexical similarity or distance may be used to identify or determine a weight or rank associated with a search result, [Wherein a similarity/relevance/relatedness in the context of Matson will clearly correspond to a confidence of relevance as disclosed in the instant application in view of the probabilities, i.e. confidences, of He]); and, providing a response to the query comprising at least a subset of the ranked list candidate resources to a graphical user interface for display thereby ([0113] a comprehensive set of search results is provided in response to the search query. The comprehensive search results can include a combination of the first set of search results and the second set of search results. In this way, the comprehensive set of search results includes search results generated from performing a semantic search and performing a lexical search. In some cases, the search results from the semantic search and lexical search may be interleaved with one another based on ranking of relevance to the search query). Regarding claim 21, Matson in view of He, further in view of Fei discloses: the computer-implemented method of claim 1. He further discloses: performing lexical pre-processing and semantic processing before obtaining the first set of candidate resources and before obtaining the second set of candidate resources ([Fig. 3, Pre-processor 306, Model Trainer 204, Data 304, New Data 332], [0090] The model trainer 204 may receive the pre-processed data 318 as input 310 or via the database 308. The model trainer 204 may implement a suitable ML algorithm 326 to train an ML model 312 on a set of training data 328 from the pre-processed data 318, [Wherein a pre-processor to collect data before training a model indicates the training step to be a processing step after a pre-processing step. Further, applying new data 332 to the same ML model 312 after training indicates first and second sets of candidate resources in view of those previously disclosed of Matson based on lexical and semantic similarity (see [0088] of Matson which discloses processing in advance of operations). Further still, consider lexical search generator 732 and semantic search generator 702 of He indicating lexical and semantic processing operations are within the scope of He]). Regarding claim 23, Matson in view of He, further in view of Fei discloses: the system of claim 9. He further discloses: wherein the at least one processor is further configured to: perform lexical pre-processing and semantic processing before obtaining the first set of candidate resources and before obtaining the second set of candidate resources ([Fig. 3, Pre-processor 306, Model Trainer 204, Data 304, New Data 332], [0090] The model trainer 204 may receive the pre-processed data 318 as input 310 or via the database 308. The model trainer 204 may implement a suitable ML algorithm 326 to train an ML model 312 on a set of training data 328 from the pre-processed data 318, [Wherein a pre-processor to collect data before training a model indicates the training step to be a processing step after a pre-processing step. Further, applying new data 332 to the same ML model 312 after training indicates first and second sets of candidate resources in view of those previously disclosed of Matson based on lexical and semantic similarity (see [0088] of Matson which discloses processing in advance of operations). Further still, consider lexical search generator 732 and semantic search generator 702 of He indicating lexical and semantic processing operations are within the scope of He]). Regarding claim 25, Matson in view of He, further in view of Fei discloses: the non-transitory computer-readable medium of claim 17. He further discloses: performing lexical pre-processing and semantic processing before obtaining the first set of candidate resources and before obtaining the second set of candidate resources ([Fig. 3, Pre-processor 306, Model Trainer 204, Data 304, New Data 332], [0090] The model trainer 204 may receive the pre-processed data 318 as input 310 or via the database 308. The model trainer 204 may implement a suitable ML algorithm 326 to train an ML model 312 on a set of training data 328 from the pre-processed data 318, [Wherein a pre-processor to collect data before training a model indicates the training step to be a processing step after a pre-processing step. Further, applying new data 332 to the same ML model 312 after training indicates first and second sets of candidate resources in view of those previously disclosed of Matson based on lexical and semantic similarity (see [0088] of Matson which discloses processing in advance of operations). Further still, consider lexical search generator 732 and semantic search generator 702 of He indicating lexical and semantic processing operations are within the scope of He]). Claim(s) 2, 6-7, 10, 14-15, 18 is/are rejected under 35 U.S.C. 103 as being unpatentable over Matson in view of He, further in view of Fei, further in view of Farmer et al. (US-20210089719-A1), hereinafter Farmer. Regarding claim 2, Matson in view of He, further in view of Fei discloses: the computer-implemented method of claim 1. Matson further discloses: wherein obtaining the second set of candidate resources comprises: determining a respective measure of similarity between the second embedding and each of the first embeddings ([0070] text embeddings are generated to enable semantic similarity analysis. A text embedding is generally in the form of a vector. As such, a semantic search includes comparing text embeddings or vectors and determining a similarity distance between the embeddings or vectors). Matson in view of He, further in view of Fei does not disclose: determining one or more candidate resources from the knowledge resources having a respective measure of similarity that meets a predetermined threshold, the second set of candidate resources comprising the one or more candidate resources. Farmer discloses: determining one or more candidate resources from the knowledge resources having a respective measure of similarity that meets a predetermined threshold, the second set of candidate resources comprising the one or more candidate resources ([0041] the concept extraction logic 46 may ignore semantic concepts identified by the natural language processing kit that have a relevancy score below a predetermined threshold and only select the identified semantic concepts that have a relevancy score above the predetermined threshold). Matson, He, Fei, and Farmer are considered analogous art within semantic analysis of received input queries. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the teachings of Matson in view of He, further in view of Fei to incorporate the teachings of Farmer, because of the novel way to identify relevance of documents based on vector similarity comparisons to predefined semantic concept vectors, improving natural language searches for complex tasks where keyword searching may be inefficient (Farmer, [0003]-[0005]). Regarding claim 6, Matson in view of He, further in view of Fei discloses: the computer-implemented method of claim 1. He further discloses: wherein ranking the first set of candidate resources and the second set of candidate resources based on at least the level of confidence determined for each of the first set of candidate resources and the second set of candidate resources to generate the ranked list of candidate resources comprises: obtaining metadata associated with the first set of candidate resources and the second set of candidate resources ([Fig. 5, Document Corpus 508 “for use by search manager 124”], [0108] a document corpus 508 suitable for use by the search manager 124… The corpus may be created for a specific domain or purpose, and it may be annotated with metadata or labels to facilitate analysis. Document corpora are commonly used in research and industry to train machine learning models and to develop NLP applications, [Documents to be used by the search manager, i.e. to be compared for relevance, indicates the document corpus to be containing first and second candidate resource sets in view of the previously disclosed lexical and semantic searching disclosed in Matson and/or He]). Matson in view of He, further in view of Fei does not disclose: ranking the first set of candidate resources and the second set of candidate resources based on the level of confidence determined for each of the first set of candidate resources and the second set of candidate resources and the metadata. Farmer discloses: ranking the first set of candidate resources and the second set of candidate resources based on the level of confidence determined for each of the first set of candidate resources and the second set of candidate resources and the metadata ([0054] the search logic 44 searches the document database 38a based on the initial search query rather than the concepts from the concept mapping logic 48. In some examples, the search logic 44 searches the document database 38a using the Okapi BM25 ranking function, [0055] At step 310, the server computing device 12b returns the result set of documents to the user computing device 12a. In some examples, metadata associated with the result set of documents, such as publication date, may also be returned, which may be used to rank the documents, [Ranking documents based on an original search query necessarily indicates the ranking to be involving a confidence related to ability to respond to the query (in view of the previously disclosed probability/confidence of relevance of He), wherein metadata is also clearly considered as a factor]). Matson, He, Fei, and Farmer are considered analogous art within semantic analysis of received input queries. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the teachings of Matson in view of He, further in view of Fei to incorporate the teachings of Farmer, because of the novel way to identify relevance of documents based on vector similarity comparisons to predefined semantic concept vectors, improving natural language searches for complex tasks where keyword searching may be inefficient (Farmer, [0003]-[0005]). Regarding claim 7, Matson in view of He, further in view of Fei, further in view of Farmer discloses: the computer-implemented method of claim 6. Farmer further discloses: wherein the metadata comprises at least one of: an age of each candidate resource of the first set of candidate resources and the second set of candidate resources ([0055] metadata associated with the result set of documents, such as publication date, may also be returned, which may be used to rank the documents, [Publication date necessarily discloses an age of the candidate resource in view of when the resource is considered, i.e. ordering by publication date is an equivalent operation to order by “age”]); and a usage frequency of each candidate resource of the first set of candidate resources and the second set of candidate resources ([The examiner would like to note that this element does not require a mapping due to the disjunctive construction of the claim]). Regarding claim 10, Matson in view of He, further in view of Fei discloses: the system of claim 9. Matson further discloses: wherein, to obtain the second set of candidate resources, the at least one processor is configured to: determine a respective measure of similarity between the second embedding and each of the first embeddings ([0070] text embeddings are generated to enable semantic similarity analysis. A text embedding is generally in the form of a vector. As such, a semantic search includes comparing text embeddings or vectors and determining a similarity distance between the embeddings or vectors). Matson in view of He, further in view of Fei does not disclose: determine one or more candidate resources from the knowledge resources having a respective measure of similarity that meets a predetermined threshold, the second set of candidate resources comprising the one or more candidate resources. Farmer discloses: determine one or more candidate resources from the knowledge resources having a respective measure of similarity that meets a predetermined threshold, the second set of candidate resources comprising the one or more candidate resources ([0041] the concept extraction logic 46 may ignore semantic concepts identified by the natural language processing kit that have a relevancy score below a predetermined threshold and only select the identified semantic concepts that have a relevancy score above the predetermined threshold). Matson, He, Fei, and Farmer are considered analogous art within semantic analysis of received input queries. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the teachings of Matson in view of He, further in view of Fei to incorporate the teachings of Farmer, because of the novel way to identify relevance of documents based on vector similarity comparisons to predefined semantic concept vectors, improving natural language searches for complex tasks where keyword searching may be inefficient (Farmer, [0003]-[0005]). Regarding claim 14, Matson in view of He, further in view of Fei discloses: the system of claim 9. He further discloses: wherein, to rank the first set of candidate resources and the second set of candidate resources based on at least the level of confidence determined for each of the first set of candidate resources and the second set of candidate resources to generate the ranked list of candidate resources, the at least one processor is configured to: obtain metadata associated with the first set of candidate resources and the second set of candidate resources ([Fig. 5, Document Corpus 508 “for use by search manager 124”], [0108] a document corpus 508 suitable for use by the search manager 124… The corpus may be created for a specific domain or purpose, and it may be annotated with metadata or labels to facilitate analysis. Document corpora are commonly used in research and industry to train machine learning models and to develop NLP applications, [Documents to be used by the search manager, i.e. to be compared for relevance, indicates the document corpus to be containing first and second candidate resource sets in view of the previously disclosed lexical and semantic searching disclosed in Matson and/or He]). Matson in view of He, further in view of Fei does not disclose: rank the first set of candidate resources and the second set of candidate resources based on the level of confidence determined for each of the first set of candidate resources and the second set of candidate resources and the metadata. Farmer discloses: rank the first set of candidate resources and the second set of candidate resources based on the level of confidence determined for each of the first set of candidate resources and the second set of candidate resources and the metadata ([0054] the search logic 44 searches the document database 38a based on the initial search query rather than the concepts from the concept mapping logic 48. In some examples, the search logic 44 searches the document database 38a using the Okapi BM25 ranking function, [0055] At step 310, the server computing device 12b returns the result set of documents to the user computing device 12a. In some examples, metadata associated with the result set of documents, such as publication date, may also be returned, which may be used to rank the documents, [Ranking documents based on an original search query necessarily indicates the ranking to be involving a confidence related to ability to respond to the query (in view of the previously disclosed probability/confidence of relevance of He), wherein metadata is also clearly considered as a factor]). Matson, He, Fei, and Farmer are considered analogous art within semantic analysis of received input queries. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the teachings of Matson in view of He, further in view of Fei to incorporate the teachings of Farmer, because of the novel way to identify relevance of documents based on vector similarity comparisons to predefined semantic concept vectors, improving natural language searches for complex tasks where keyword searching may be inefficient (Farmer, [0003]-[0005]). Regarding claim 15, Matson in view of He, further in view of Fei, further in view of Farmer discloses: the system of claim 14. Farmer further discloses: wherein the metadata comprises at least one of: an age of each candidate resource of the first set of candidate resources and the second set of candidate resources ([0055] metadata associated with the result set of documents, such as publication date, may also be returned, which may be used to rank the documents, [Publication date necessarily discloses an age of the candidate resource in view of when the resource is considered, i.e. ordering by publication date is an equivalent operation to order by “age”]); and a usage frequency of each candidate resource of the first set of candidate resources and the second set of candidate resources ([The examiner would like to note that this element does not require a mapping due to the disjunctive construction of the claim]). Regarding claim 18, Matson in view of He, further in view of Fei discloses: the non-transitory computer-readable device of claim 17. Matson further discloses: wherein obtaining the second set of candidate resources comprises: determining a respective measure of similarity between the second embedding and each of the first embeddings ([0070] text embeddings are generated to enable semantic similarity analysis. A text embedding is generally in the form of a vector. As such, a semantic search includes comparing text embeddings or vectors and determining a similarity distance between the embeddings or vectors). Matson in view of He, further in view of Fei does not disclose: determining one or more candidate resources from the knowledge resources having a respective measure of similarity that meets a predetermined threshold, the second set of candidate resources comprising the one or more candidate resources. Farmer discloses: determining one or more candidate resources from the knowledge resources having a respective measure of similarity that meets a predetermined threshold, the second set of candidate resources comprising the one or more candidate resources ([0041] the concept extraction logic 46 may ignore semantic concepts identified by the natural language processing kit that have a relevancy score below a predetermined threshold and only select the identified semantic concepts that have a relevancy score above the predetermined threshold). Matson, He, Fei, and Farmer are considered analogous art within semantic analysis of received input queries. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the teachings of Matson in view of He, further in view of Fei to incorporate the teachings of Farmer, because of the novel way to identify relevance of documents based on vector similarity comparisons to predefined semantic concept vectors, improving natural language searches for complex tasks where keyword searching may be inefficient (Farmer, [0003]-[0005]). Claim(s) 8, 16 is/are rejected under 35 U.S.C. 103 as being unpatentable over Matson in view of He, further in view of Fei, further in view of Felbah et al. (US-20250284720-A1), hereinafter Felbah. Regarding claim 8, Matson in view of He, further in view of Fei discloses: the computer-implemented method of claim 1. Matson in view of He, further in view of Fei does not disclose: wherein each of the knowledge resources comprises at least one of: a set of instructions for rectifying issues in a computing system; or knowledge base articles comprising solutions for rectifying the issues in the computing system. Felbah discloses: wherein each of the knowledge resources comprises at least one of: a set of instructions for rectifying issues in a computing system ([0138] An entity 601, such as a business, has a knowledge base comprising numerous text documents including, for example, technical support documentation, service manuals, user manuals, troubleshooting guidelines, etc., such as text document 501); or knowledge base articles comprising solutions for rectifying the issues in the computing system ([0054] Knowledge Base (KB): An exhaustive compendium of an entity's technical support documentation including, but not limited to, user manuals and troubleshooting guidelines, devised to address customer inquiries and to furnish in-depth information on the entity's products and services). Matson, He, Fei, and Felbah are considered analogous art within query response generation through hybrid query analysis. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the teachings of Matson in view of He, further in view of Fei to incorporate the teachings of Felbah, because of the novel way to implement a hybrid scoring function which combines both keyword (lexical) match scores and embedding (semantic) similarity scores in a balanced formula, improving the relevance and accuracy of return search results (Felbah, [0120]). Regarding claim 16, Matson in view of He, further in view of Fei discloses: the system of claim 9. Matson in view of He, further in view of Fei does not disclose: wherein each of the knowledge resources comprises at least one of: a set of instructions for rectifying issues in a computing system; or knowledge base articles comprising solutions for rectifying the issues in the computing system. Felbah discloses: wherein each of the knowledge resources comprises at least one of: a set of instructions for rectifying issues in a computing system ([0138] An entity 601, such as a business, has a knowledge base comprising numerous text documents including, for example, technical support documentation, service manuals, user manuals, troubleshooting guidelines, etc., such as text document 501); or knowledge base articles comprising solutions for rectifying the issues in the computing system ([0054] Knowledge Base (KB): An exhaustive compendium of an entity's technical support documentation including, but not limited to, user manuals and troubleshooting guidelines, devised to address customer inquiries and to furnish in-depth information on the entity's products and services). Matson, He, Fei, and Felbah are considered analogous art within query response generation through hybrid query analysis. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the teachings of Matson in view of He, further in view of Fei to incorporate the teachings of Felbah, because of the novel way to implement a hybrid scoring function which combines both keyword (lexical) match scores and embedding (semantic) similarity scores in a balanced formula, improving the relevance and accuracy of return search results (Felbah, [0120]). Claim(s) 22, 24, 26 is/are rejected under 35 U.S.C. 103 as being unpatentable over Matson in view of He, further in view of Fei, further in view of Anzenberg et al. (US-20240062571-A1), hereinafter Anzenberg. Regarding claim 22, Matson in view of He, further in view of Fei discloses: the computer-implemented method of claim 21. Matson in view of He, further in view of Fei does not disclose: wherein performing the lexical pre-processing comprises: normalizing non-standard terms included in the query and the knowledge resources; and merging relatively frequent co-occurring phrases in at least one of the query or the knowledge resources into single terms. Anzenberg discloses: wherein performing the lexical pre- processing comprises: normalizing non-standard terms included in the query and the knowledge resources ([0045] an “indication” refers to one or more words within the document with the same semantic meaning, but with differing syntax (e.g., different letters, and/or added/subtracted words) relative to the keyword. For example, a keyword may be “amount,” but no such string with the exact match is found, but a similar phrase can be found in a document, such as “total.”, [In this example, “amount” is a non-standard term as no exact match is found, and the term is ‘normalized’ to “total”]); and merging relatively frequent co-occurring phrases in at least one of the query or the knowledge resources into single terms ([0043] Each value can be individual letters, numbers, words, symbols and/or a combination of numbers, words, and/or symbols depending on how the OCR component 102 parses characters. For example, two words, such as “Mulberry's Market,” although two separate words may be combined or formatted as one single value since it refers to the same entity (e.g., a place of business as determined via NER or other NLP component)). Matson, He, Fei, and Anzenberg are considered analogous art within semantic processing of user-input queries. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the teachings of Matson in view of He, further in view of Fei to incorporate the teachings of Anzenberg, because of the novel way to employ BERT models or encoders for field extraction technologies for user queries, reducing the amount of parameters required to distill knowledge for the field extraction functionality, improving memory consumption relative to existing technologies for field extraction from documents related to input user queries (Anzenberg, [0029]). Regarding claim 24, Matson in view of He, further in view of Fei discloses: the system of claim 23. Matson in view of He, further in view of Fei does not disclose: wherein, to perform the lexical pre- processing, the at least one processor is configured to: normalize non-standard terms included in the query and the knowledge resources; and merge relatively frequent co-occurring phrases in at least one of the query or the knowledge resources into single terms. Anzenberg discloses: wherein, to perform the lexical pre- processing, the at least one processor is configured to: normalize non-standard terms included in the query and the knowledge resources ([0045] an “indication” refers to one or more words within the document with the same semantic meaning, but with differing syntax (e.g., different letters, and/or added/subtracted words) relative to the keyword. For example, a keyword may be “amount,” but no such string with the exact match is found, but a similar phrase can be found in a document, such as “total.”, [In this example, “amount” is a non-standard term as no exact match is found, and the term is ‘normalized’ to “total”]); and merge relatively frequent co-occurring phrases in at least one of the query or the knowledge resources into single terms ([0043] Each value can be individual letters, numbers, words, symbols and/or a combination of numbers, words, and/or symbols depending on how the OCR component 102 parses characters. For example, two words, such as “Mulberry's Market,” although two separate words may be combined or formatted as one single value since it refers to the same entity (e.g., a place of business as determined via NER or other NLP component)). Matson, He, Fei, and Anzenberg are considered analogous art within semantic processing of user-input queries. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the teachings of Matson in view of He, further in view of Fei to incorporate the teachings of Anzenberg, because of the novel way to employ BERT models or encoders for field extraction technologies for user queries, reducing the amount of parameters required to distill knowledge for the field extraction functionality, improving memory consumption relative to existing technologies for field extraction from documents related to input user queries (Anzenberg, [0029]). Regarding claim 26, Matson in view of He, further in view of Fei discloses: the computer-implemented method of claim 21. Matson in view of He, further in view of Fei does not disclose: wherein performing the lexical pre-processing comprises: normalize non-standard terms included in the query and the knowledge resources; and merge relatively frequent co-occurring phrases in at least one of the query or the knowledge resources into single terms. Anzenberg discloses: wherein performing the lexical pre-processing comprises: normalize non-standard terms included in the query and the knowledge resources ([0045] an “indication” refers to one or more words within the document with the same semantic meaning, but with differing syntax (e.g., different letters, and/or added/subtracted words) relative to the keyword. For example, a keyword may be “amount,” but no such string with the exact match is found, but a similar phrase can be found in a document, such as “total.”, [In this example, “amount” is a non-standard term as no exact match is found, and the term is ‘normalized’ to “total”]); and merge relatively frequent co-occurring phrases in at least one of the query or the knowledge resources into single terms ([0043] Each value can be individual letters, numbers, words, symbols and/or a combination of numbers, words, and/or symbols depending on how the OCR component 102 parses characters. For example, two words, such as “Mulberry's Market,” although two separate words may be combined or formatted as one single value since it refers to the same entity (e.g., a place of business as determined via NER or other NLP component)). Matson, He, Fei, and Anzenberg are considered analogous art within semantic processing of user-input queries. Therefore, it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the teachings of Matson in view of He, further in view of Fei to incorporate the teachings of Anzenberg, because of the novel way to employ BERT models or encoders for field extraction technologies for user queries, reducing the amount of parameters required to distill knowledge for the field extraction functionality, improving memory consumption relative to existing technologies for field extraction from documents related to input user queries (Anzenberg, [0029]). Conclusion The prior art made of record and not relied upon is considered pertinent to applicant's disclosure. Wagner et al. (US-20220382978-A1) discloses “systems and methods for training masked language models based on partial sequences of tokens. A sequence of tokens for training a transformer model is received. A defined proportion of the sequence of tokens is selected. Each value of the defined proportion of the sequence of tokens is replaced with a defined value. The transformer model is trained by using the sequence of tokens to train the transformer model during a forward pass and using a subset of the sequence of tokens that includes the defined the proportion of the sequence of tokens to train the transformer model during a backward pass” (abstract). See entire document. Any inquiry concerning this communication or earlier communications from the examiner should be directed to THEODORE JOHN WITHEY whose telephone number is (703)756-1754. The examiner can normally be reached Monday - Friday, 8am-5pm. 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, Andrew Flanders can be reached at (571) 272-7516. 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. /THEODORE WITHEY/Examiner, Art Unit 2655 /ANDREW C FLANDERS/Supervisory Patent Examiner, Art Unit 2655
Read full office action

Prosecution Timeline

Show 1 earlier event
Jan 30, 2026
Non-Final Rejection mailed — §103, §112
Mar 03, 2026
Examiner Interview Summary
Mar 03, 2026
Applicant Interview (Telephonic)
Mar 12, 2026
Response Filed
May 01, 2026
Final Rejection mailed — §103, §112
Jun 25, 2026
Request for Continued Examination
Jun 29, 2026
Response after Non-Final Action
Nov 03, 2026
Non-Final Rejection mailed — §103, §112 (current)

Precedent Cases

Applications granted by this same examiner with similar technology

Patent 12724985
MACHINE-LEARNING BASED IRRELEVANT SENTENCE CLASSIFIER
3y 11m to grant Granted Sep 01, 2026
Patent 12657389
TECHNOLOGIES FOR ERROR REDUCTION IN INTENT CLASSIFICATION
3y 5m to grant Granted Jun 16, 2026
Patent 12646499
METHOD, DEVICE, AND COMPUTER PROGRAM PRODUCT FOR PROCESSING INFORMATION
3y 3m to grant Granted Jun 02, 2026
Patent 12632670
Natural Language Processing for Identifying Bias in a Span of Text
3y 2m to grant Granted May 19, 2026
Patent 12591744
METHOD FOR TRAINING SEMANTIC REPRESENTATION MODEL, DEVICE AND STORAGE MEDIUM
4y 0m to grant Granted Mar 31, 2026
Study what changed to get past this examiner. Based on 5 most recent grants.

Strategy Recommendation AI-generated — please review before filing

Get a prosecution strategy drawn from examiner precedents, rejection analysis, and claim mapping.
Typically takes 5-10 seconds — AI-generated, attorney review required before filing

Prosecution Projections

3-4
Expected OA Rounds
44%
Grant Probability
80%
With Interview (+36.1%)
3y 0m (~6m remaining)
Median Time to Grant
High
PTA Risk
Based on 34 resolved cases by this examiner. Grant probability derived from career allowance rate.

Sign in with your work email

Enter your email to receive a magic link. No password needed.

Personal email addresses (Gmail, Yahoo, etc.) are not accepted.

Free tier: 3 strategy analyses per month