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 .
Response to Amendment
The Amendment filed on 01/30/2026 has been entered. Claims 1-20 remain pending in this application.
Response to Arguments
Applicant’s arguments filed 01/30/2026 have been fully considered but are not persuasive.
With respect to the 35 U.S.C. 101 rejection, on pages 12-13, the Applicant asserts that claim 1 is not directed to a mental process, and that claim 1, as amended, contains limitations that integrate any alleged abstract idea into a practical application. The Applicant cites Ex Parte Desjardins and states that claim 1 describes generating a custom machine-learning model context that improves the accuracy and functioning of machine-learning models, such as large language models, by constraining the set of data provided to a machine learning model. They also assert that claim 1, as amended, recites a custom context that reduces the amount of data provided to the machine learning model, resulting in reduced storage and streamlined processing.
The Examiner respectfully disagrees. The original claims, and the claims as amended, are merely utilizing computational devices, in this case “a large language model (LLM)”, as tools to perform a method which is directed to an abstract idea. The claim, under its broadest reasonable interpretation, recites a method of extracting, processing, and analyzing data (i.e. digital files) that is then input into an LLM. This is an abstract idea in the form of certain methods of organizing human activity (i.e. mental processes such as observation, evaluation, judgement, and opinion). The steps of extracting and processing the data from the digital files could be performed by a human using pen and paper or by purely mental reasoning, save for the recitation of generic computing components. Further, the claims do not integrate the judicial exception into a practical application. The recitation of “a large language model (LLM)” is a generic instruction to perform the abstract idea on a computer or using a computing device and does not impose a meaningful limit on the judicial exception. The “a large language model (LLM)” is recited at such a high-level of generality and is used merely as a tool to perform the abstract idea faster or more efficiently. There is no reasonable improvement to the functioning of the data extraction, processing, generation, or analyzation, the LLMs, the computational system as a whole, nor to any other technology or technical field. The claims do not include any additional elements that amount to significantly more than the judicial exception. The claims, as written and amended, do not include more than mere instructions to perform the abstract method using generic computer components. Hence, the Applicant’s arguments are not persuasive.
With respect to the 35 U.S.C. rejection, on pages 10-12, of claims 1-7, 9, and 13 under Karandish et al. (US Patent Application Publication No. 2021/0056150), hereinafter referred to as Karandish, in view of Lewis et al. ("Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks", 2020), hereinafter referred to as Lewis, claims 8 and 14-20 under Karandish, in view of Lewis, and further in view of Mol et al. ("Review on knowledge extraction from text and scope in agriculture domain", 09/29/2022), hereinafter referred to as Mol, and claims 10-12 under Karandish, in view of Lewis, and further in view of Leary et al. (US Patent Application Publication No. 2024/0095463), hereinafter referred to as Leary, the Applicant asserts that both Karandish and Lewis fail to disclose the amended limitations of the claims, specifically “determining whether an element of the text query includes identification data; in response to a determination that the element includes identification data, determining whether a candidate entry of the set of entries is associated with the identification data; in response to a determination that the candidate entry is not associated with the identification data, marking the candidate entry as irrelevant to the text query”.
In response to the argument that neither Karandish or Lewis discloses the amended limitations of “determining whether an element of the text query includes identification data; in response to a determination that the element includes identification data, determining whether a candidate entry of the set of entries is associated with the identification data”, Lewis pg. 2 states “The retriever provides latent documents conditioned on the input, and the seq2seq model then conditions on these latent documents together with the input to generate the output”. The input contains information (i.e., identification data) that allows the retriever to obtain documents that are related to it. Karandish has a similar disclosure in paragraph [0035]: “As an example, user 350 can use user device 340 to send a question to chat agent 321, which can send the question to question answering system 310, and question answering system 310 can determine an answer to be returned through chat agent 321 to user device 340 in response to the question. In several embodiments, question answering system 310 can perform an ingestion process, such as method 400 (FIGS. 4A and 4B, described below) and/or method 600 (FIGS. 6A and 6B, described below), which can be followed by a retrieval and presentment process, such as method 500 (FIGS. 5A and 5B, described below) and/or method 700 (FIGS. 7A and 7B, described below)”. In order for successful document retrieval to take place, the input query must have some sort of identification data or information that allows the system to retrieve documents from the set related to it, as shown in both Lewis and Karandish. In response to the argument that neither Karandish or Lewis discloses the amended limitation of “in response to a determination that the candidate entry is not associated with the identification data, marking the candidate entry as irrelevant to the text query”, Karandish paragraph [0054] states “For example, a pre-determined and/or dynamic relevance threshold can be applied, such that the content sections (e.g., sections of text) are either extracted or not extracted, depending on the score for the content section determined by the model. The extracted content sections (e.g., 432-436) can meet or exceed the relevance threshold, and the non-extracted content sections (e.g., 437-439) do not meet the relevance threshold”. This states that Karandish deems sections of text, or documents, irrelevant utilizing a relevance score under a certain threshold. This choice to not extract documents or sections of text under a particular relevance score threshold is thereby marking a document or section of text as being irrelevant to the text query. Hence, the 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.
Claim(s) 1-20 rejected under 35 U.S.C. 101 because the claimed invention is directed to an abstract idea without significantly more. Independent claims 1, 16, and 20 recite “generating a custom machine-learning model context”, “extracting data”, “processing the extracted data”, “generating a vector embedding”, “ingesting the vector embeddings”, “receiving a text query”, “identifying, based on the vector embeddings, a set of entries among the entries in the database, wherein each entry of the set of entries is semantically similar to the text query”, “retrieving the set of entries”, “determining whether an element of the text query includes identification data”, “determining whether a candidate entry of the set of entries is associated with the identification data”, “marking the candidate entry as irrelevant to the text query”, “formulating a prompt”, “submitting the prompt”, “obtaining a response to the prompt”, and “outputting the response”. These limitations, as drafted, are a process that, under a broadest reasonable interpretation, covers the abstract idea of “mental processes” because they cover concepts performed in the human mind, including observation, evaluation, judgement, and opinion. See MPEP 2106.04(a)(2). That is, other than reciting “a large language model (LLM)”, nothing in the claimed elements preclude the steps from practically being performed by a person reading a page of text, parsing the text into coherent segments, embedding that text, and then using said extracted and embedded text to answer a question.
This judicial exception is not integrated into a practical application because the additional elements “a large language model (LLM)” are recited at such a high level of generality. Accordingly, these additional elements do not integrate the abstract idea into a practical application because they do not impose any meaningful limits on practicing the abstract idea. Thus, the claims as a whole are directed to an abstract idea (Step 2A, prong two).
Claims 1, 16, and 20 do not include any additional elements that are sufficient to amount to significantly more than the judicial exception because, as discussed above with respect to integration of the abstract idea into a practical applications, the additional elements of “a large language model (LLM)” amount to no more than mere instructions to apply the exception using generic computer components. Mere instructions to apply an exception using a generic computer component cannot provide an inventive concept (Step 2B).
Dependent claims 2-15 and 17-19 are directed to describing the method of text extraction, as well as the prompt and the contents of the document database. These limitations are also related to the abstract idea of “mental processes.” That is, nothing in the claimed elements preclude the steps from practically being performed by a person reading a page of text, parsing the text into coherent segments, embedding that text, and then using said extracted and embedded text to answer a question.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claim(s) 1-7, 9, 13, and 16-18 is/are rejected under 35 U.S.C. 103 as being unpatentable over Karandish et al. (US Patent Application Publication No. 2021/0056150), hereinafter referred to as Karandish, in view of Lewis et al. ("Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks", 2020), hereinafter referred to as Lewis.
Regarding claim 1, Karandish discloses a computer-implemented method comprising: generating a custom machine-learning model context by:- (“Question answering systems, such as chatbots, generally provide automated mechanisms for users to ask questions in a natural language form and receive answers to those questions. Many question answering systems cache a copy of source documents so that the cached information will be available when answering questions,” Karandish para [0002]);
extracting data from a digital file, wherein the extracted data includes text ("In a number of embodiments, as shown in FIG. 4B, method 400 additionally can include a block 430 of text mining," Karandish para [0053]);
processing the extracted data to generate semantically coherent text segments ("Block 430 of text mining can include receiving the pre-processed document (e.g., 421 (FIG. 4A)) as input and automatically outputting extracted content sections that can be used as answers," Karandish para [0053]);
processing the extracted data to generate question-answer pairs (Karandish Fig. 4B reference character 440);
generating a vector embedding for each semantically coherent text segment ("In many embodiments, the extracted content section (e.g., 432, 433) can be transformed into vector embeddings using a suitable machine learning-based transformation algorithm," Karandish para [0086]);
generating a vector embedding for each question-answer pair ("In many embodiments, the extracted content section (e.g., 432, 433) can be transformed into vector embeddings using a suitable machine learning-based transformation algorithm," Karandish para [0086]);
and ingesting the vector embeddings into a database as entries (Karandish Fig. 3 shows Ingestion System 312 and Index Database 316 within the Question Answering System);
identifying, based on the vector embeddings, a set of entries among the entries in the database, wherein each entry of the set of entries is semantically similar to the text query ("The system can advantageously store location metadata about each question-answer pair, which can beneficially allow finding and retrieving the relevant answer and information from within the source file," Karandish para [0128] and “In some embodiments, the machine learning algorithm used in text mining can use a number of engineered features for each section of text, such as TF-IDF (term frequency-inverse document frequency) keyword rank, word frequency count, section character length, and/or sentence vector semantic similarity (e.g., using BERT (Bidirectional Encoder Representations from Transformations) vector embeddings,” Karandish para [0055]);
set of entries ("The system can advantageously store location metadata about each question-answer pair, which can beneficially allow finding and retrieving the relevant answer and information from within the source file," Karandish para [0128]);
in response to a determination that the candidate entry is not associated with the identification data, marking the candidate entry as irrelevant to the text query (“For example, a pre-determined and/or dynamic relevance threshold can be applied, such that the content sections (e.g., sections of text) are either extracted or not extracted, depending on the score for the content section determined by the model. The extracted content sections (e.g., 432-436) can meet or exceed the relevance threshold, and the non-extracted content sections (e.g., 437-439) do not meet the relevance threshold,” Karandish para [0054]).
However, Karandish fails to disclose receiving a text query; determining whether an element of the text query includes identification data; in response to a determination that the element includes identification data, determining whether a candidate entry of the set of entries is associated with the identification data; and formulating a prompt for a large language model (LLM) based on the text query, wherein the formulating incorporates data from relevant entries of the set of entries into the prompt
Lewis teaches receiving a text query (Lewis Figure 1 shows receiving a query on the left);
determining whether an element of the text query includes identification data (“The retriever provides latent documents conditioned on the input, and the seq2seq model then conditions on these latent documents together with the input to generate the output,” Lewis pg. 2, the input contains information (i.e., identification data) that allows the retriever to obtain documents that are related to it);
in response to a determination that the element includes identification data, determining whether a candidate entry of the set of entries is associated with the identification data (“The retriever provides latent documents conditioned on the input, and the seq2seq model then conditions on these latent documents together with the input to generate the output,” Lewis pg. 2);
and formulating a prompt for a large language model (LLM) based on the text query, wherein the formulating incorporates data from relevant entries of the set of entries into the prompt (Lewis Figure 1 shows the generator, which is BART-large (an LLM) per 2.3 pg. 3 and "To combine the input x with the retrieved content z when generating from BART, we simply concatenate them," Lewis 2.3 pg. 3).
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to have modified Karandish’s method of text extraction and question-answer pair generation by including Lewis’ method of utilizing the information within the prompt and then utilizing an LLM to formulate a prompt. This would result in improved accuracy for the output of the LLM, as well as reduced hallucination frequency, as the model is pulling documents and information from a database of that domain based upon the prompt itself.
Regarding claim 2, Karandish, in view of Lewis, discloses all of the limitations of claim 1. Karandish further discloses structured examples of questions and answers ("The one or more databases can include data used in ingesting and retrieving source documents for question answering, for example," Karandish para [0040] and "Block 440 of question generation can include using the extracted content sections (e.g., 432-436) as input and outputting question-answer pairs that are added to an index, such as in index database 316 (FIG. 3)," Karandish para [0056]).
However, Karandish fails to disclose wherein the prompt comprises.
Lewis teaches wherein the prompt includes ("To combine the input x with the retrieved content z when generating from BART, we simply concatenate them," Lewis 2.3 pg. 3).
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to have modified Karandish’s method of text extraction and question-answer pair generation by including Lewis’ method of utilizing an LLM to formulate a prompt. This would result in improved accuracy for the output of the LLM, as well as reduced hallucination frequency, as the model is pulling documents and information from a database of that domain.
Regarding claim 3, Karandish, in view of Lewis, discloses all of the limitations of claim 1. However, Karandish fails to disclose further comprising: submitting the prompt to the LLM; obtaining a response to the prompt from the LLM; and outputting the response as an answer to the query.
Lewis teaches further comprising: submitting the prompt to the LLM (Lewis Figure 1 shows the generator, which is BART-large (an LLM) per 2.3 pg. 3);
obtaining a response to the prompt from the LLM (Lewis Figure 1 shows the LLM's response on the right side);
and outputting the response as an answer to the text query (Lewis Figure 1 shows the LLM's response on the right side).
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to have modified Karandish’s method of text extraction and question-answer pair generation by including Lewis’ method of utilizing an LLM to formulate a prompt and then submit that prompt to an LLM. This would result in improved accuracy for the output of the LLM, as well as reduced hallucination frequency, as the model is pulling documents and information from a database of that domain.
Regarding claim 4, Karandish, in view of Lewis, discloses all of the limitations of claim 1. Karandish further discloses wherein processing the extracted data to generate the semantically coherent text segments includes applying a natural language processing algorithm to group the text into the semantically coherent text segments, and (Karandish Fig. 4B reference character 440 (reference para [0057]) and 450).
Regarding claim 5, Karandish, in view of Lewis, discloses all of the limitations of claim 4. Karandish further discloses wherein applying the natural language processing algorithm to group the text into the semantically coherent text segments includes: detecting formatting in the extracted data ("In a number of embodiments, method 400 additionally can include a block 450 of adding location metadata," Karandish para [0059] and “In some embodiments, the location delimiters for an answer can include keywords in the answer, identification of formatting tags associated with the answer (e.g., bold text, etc.), or other suitable location indicators,” Karandish para [0060]);
detecting metadata tags in the extracted data ("In a number of embodiments, method 400 additionally can include a block 450 of adding location metadata," Karandish para [0059]);
grouping the text into a set of preliminary text segments based on the detected formatting ("For example, as shown in FIG. 4B, location metadata 451-455 can be generated and added to the index in association with question-answer pairs 441-445, respectively," Karandish para [0059] and “In some embodiments, the location delimiters for an answer can include keywords in the answer, identification of formatting tags associated with the answer (e.g., bold text, etc.), or other suitable location indicators,” Karandish para [0060]);
grouping the text into a set of preliminary text segments based on the detected metadata tags ("For example, as shown in FIG. 4B, location metadata 451-455 can be generated and added to the index in association with question-answer pairs 441-445, respectively," Karandish para [0059]);
identifying text segments for recombination among the set of preliminary text segments ("For example, as shown in FIG. 5A, question 511 can be matched with question-answer pair 441, which is stored in the index in association with location metadata 451," Karandish para [0069]);
recombining the identified text segments and adding the recombined text segments to the set of preliminary text segments ("The question received can be matched to a question-answer pair, even if the question received does not exactly match, word-for-word, the question stored in the question-answer pair," Karandish para [0069]);
determining that a selected preliminary text segment requires additional context from an adjacent portion of the text ("For example, the techniques can include searching for text in the pre-processed document (e.g., 531) that matches the location delimiters (e.g., 562-564)," Karandish para [0077]).
However, Karandish fails to disclose and concatenating the selected preliminary text segment with the additional context.
Lewis teaches and concatenating the selected preliminary text segment with the additional context ("To combine the input x with the retrieved content z when generating from BART, we simply concatenate them," Lewis 2.3 pg. 3).
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to have modified Karandish’s method of text extraction and question-answer pair generation by including Lewis’ method of utilizing an LLM to formulate a prompt and then submit that prompt to an LLM. Concatenating the query with retrieved documents, e.g. additional context, would result in improved accuracy for the output of the LLM, as well as reduced hallucination frequency, as the model is pulling documents and information from a database of that domain.
Regarding claim 6, Karandish, in view of Lewis, discloses all of the limitations of claim 1. Karandish further discloses wherein generating the vector embedding for each semantically coherent text segment includes encoding semantic information associated with the semantically coherent text segment into a fixed-length numeric vector ("In several embodiments, block 812 of performing secure ingestion can include, for each question-answer pair of the question-answer pairs, a block 910 of transforming an answer of the question-answer pair into a first numeric vector representation," Karandish para [0109]).
Regarding claim 9, Karandish, in view of Lewis, discloses all of the limitations of claim 1. Karandish further discloses set of entries, an entry that is most relevant to the text query ("The system can advantageously store location metadata about each question-answer pair, which can beneficially allow finding and retrieving the relevant answer and information from within the source file," Karandish para [0128] and Karandish Fig. 10 reference character 1020 for similarity score, i.e. most relevant).
Regarding claim 7, Karandish, in view of Lewis, discloses all of the limitations of claim 9. Karandish further discloses wherein: identifying the entry that is most relevant to the text query includes initiating a retrieval process in which entries that are semantically similar to the text query and that originated from a second digital file that references the element are retrieved from the database ("The system can advantageously store location metadata about each question-answer pair, which can beneficially allow finding and retrieving the relevant answer and information from within the source file," Karandish para [0128] and "The portions of text in the extracted content sections can be extracted from the pre-processed document based on using conventional or customized machine learning algorithms trained on keywords, frequency, semantics," Karandish para [0053] and Karandish Fig. 10 reference character 1020 for similarity score, i.e. most relevant).
However, Karandish fails to disclose the text query includes a reference to the element.
Lewis teaches the text query includes a reference to the element (Lewis Figure 1 shows examples of queries on the left).
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to have modified Karandish’s method of text extraction and question-answer pair generation by including Lewis’ method of utilizing an LLM to formulate a prompt. This would result in improved accuracy for the output of the LLM, as well as reduced hallucination frequency, as the model is pulling documents and information from a database of that domain.
Regarding claim 13, Karandish, in view of Lewis, discloses all of the limitations of claim 7. Karandish further discloses wherein: the digital file is in Portable Document Format (PDF), and digital file includes at least one of: performing Optical Character Recognition (OCR) on the digital file, or applying a Computer Vision Model to the digital file ("Source documents can be files, webpages, or other suitable sources of content. Content can be any suitable type of information," Karandish para [0044] and "In several embodiments, an image capturing tool can be used to produce screenshots of the content and use optical character recognition (OCR) to convert it to the simplified text format," Karandish para [0052]).
As to claim 16, system claim 16 and method claim 1 are related as method and system of using same, with each claimed element’s function corresponding to the method step. Accordingly, claim 16 is similarly rejected under the same rationale as applied above with respect to the method claim.
Regarding claim 17, Karandish, in view of Lewis, discloses all of the limitations of claim 16. Karandish further discloses further comprising an internal representation of a natural language processing algorithm configured to group the data extracted from the plurality of digital files into semantically coherent text segments, wherein the vector embeddings comprise vector embeddings associated with the semantically coherent text segments (Karandish Fig. 4B reference character 440 (reference para [0057]) and 450).
Regarding claim 18, Karandish, in view of Lewis, discloses all of the limitations of claim 16. However, Karandish fails to disclose further comprising a Retrieval-Augmented Generation (RAG) architecture in which receipt of the text query triggers
Lewis teaches further comprising a Retrieval-Augmented Generation (RAG) architecture in which receipt of the text query triggers (Lewis Figure 1 shows a RAG architecture and Lewis Figure 1 shows a query being input, which then leads into a retriever, detailed in section 2.2).
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to have modified Karandish’s method of text extraction and question-answer pair generation by including Lewis’ method of utilizing a retrieval-augmented generation (RAG) architecture. This would result in improved accuracy and reliability of the output of the LLM, as well as a reduced hallucination frequency, as the model is pulling documents and information from a database of that particular domain.
Claim(s) 8, 14-15, and 19-20 is/are rejected under 35 U.S.C. 103 as being unpatentable over Karandish, in view of Lewis, and further in view of Mol et al. ("Review on knowledge extraction from text and scope in agriculture domain", 09/29/2022), hereinafter referred to as Mol.
Regarding claim 8, Karandish, in view of Lewis, disclose all of the limitations of claim 7. However, Karandish fails to disclose wherein the element is selected from a group consisting of a product, a chemical, a crop, and a pest. Mol teaches a review on knowledge extraction from text in the agricultural domain.
Mol teaches wherein the element is selected from a group consisting of a product, a chemical, a crop, and a pest ("The subdomains crop, disease, soil and region are taken into consideration," Mol sec. 7 pg. 19 and "Information regarding fertilizers, seeds, pesticides, weather, soil, crop varieties and results of various scientific research in the agricultural domain is crucial for the farmers, scientists and researchers," Mol sec. 7 pg. 26).
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to have modified Karandish’s method of text extraction and question-answer pair generation by including Mol’s method of knowledge extraction in the agricultural domain. Enormous information in various domains, including agriculture, is available in the natural language from several resources (Mol Abstract). This research in the agricultural domain will be helpful in assisting farmers in decision making by providing knowledge about the recent advancements in agriculture and other associated technologies based on the latest research results (Mol sec. 1 pg. 2).
Regarding claim 14, Karandish, in view of Lewis, disclose all of the limitations of claim 13. However, Karandish fails to disclose wherein: the element is an agricultural product, and digital file includes identification data for the agricultural product.
Mol teaches wherein: the element is an agricultural product, and digital file includes identification data for the agricultural product ("Information regarding fertilizers, seeds, pesticides, weather, soil, crop varieties and results of various scientific research in the agricultural domain is crucial for the farmers, scientists and researchers. This domain-specific structural knowledge will be beneficial for various purposes like getting the direct answers to the questions, automating farming activities, and promoting research and education in the agriculture field. Dissemination of advanced knowledge will improve productivity in the agriculture sector, and at the same time, ensuring the validity of the knowledge is also essential," Mol sec. 7 pg. 26).
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to have modified Karandish’s method of text extraction and question-answer pair generation by including Mol’s method of knowledge extraction in the agricultural domain. Enormous information in various domains, including agriculture, is available in the natural language from several resources (Mol Abstract). This research in the agricultural domain will be helpful in assisting farmers in decision making by providing knowledge about the recent advancements in agriculture and other associated technologies based on the latest research results (Mol sec. 1 pg. 2).
Regarding claim 15, Karandish, in view of Lewis, and further in view of Mol, disclose all of the limitations of claim 14. However, Karandish fails to disclose wherein the identification data includes usage instructions for the agricultural product.
Mol teaches wherein the identification data includes usage instructions for the agricultural product ("Information regarding fertilizers, seeds, pesticides, weather, soil, crop varieties and results of various scientific research in the agricultural domain is crucial for the farmers, scientists and researchers. This domain-specific structural knowledge will be beneficial for various purposes like getting the direct answers to the questions, automating farming activities, and promoting research and education in the agriculture field. Dissemination of advanced knowledge will improve productivity in the agriculture sector, and at the same time, ensuring the validity of the knowledge is also essential," Mol sec. 7 pg. 26).
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to have modified Karandish’s method of text extraction and question-answer pair generation by including Mol’s method of knowledge extraction in the agricultural domain. Enormous information in various domains, including agriculture, is available in the natural language from several resources (Mol Abstract). This research in the agricultural domain will be helpful in assisting farmers in decision making by providing knowledge about the recent advancements in agriculture and other associated technologies based on the latest research results (Mol sec. 1 pg. 2).
Regarding claim 19, Karandish, in view of Lewis, discloses all of the limitations of claim 16. However, Karandish fails to disclose wherein the databaseincludes: a public agronomic information category; a semi-public agronomic information category; and a proprietary agronomic information category.
Mol teaches wherein the databaseincludes: a public agronomic information category; a semi-public agronomic information category; and a proprietary agronomic information category ("Recently, some research has been conducted on agricultural knowledge extraction to represent the agricultural information in a structured way in knowledge bases or as an ontology for the semantic web," Mol sec. 7 pg. 19 and "The outcome of these researches is available in various natural language documents in human-readable form. Once this knowledge is represented in a structured, machine-readable form, then this can be used to get the direct answers to the questions and will be helpful for the end-users like the farmers, agricultural researchers, or the people interested in farming," Mol sec. 7 pg. 19).
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to have modified Karandish’s method of text extraction and question-answer pair generation by including Mol’s method of knowledge extraction in the agricultural domain. Enormous information in various domains, including agriculture, is available in the natural language from several resources (Mol Abstract). This research in the agricultural domain will be helpful in assisting farmers in decision making by providing knowledge about the recent advancements in agriculture and other associated technologies based on the latest research results (Mol sec. 1 pg. 2).
Regarding claim 20, Karandish discloses a non-transitory computer-readable medium comprising computer-executable instructions (Karandish Fig. 2 reference character 208) the instructions including: generating a custom machine-learning model context by (“Question answering systems, such as chatbots, generally provide automated mechanisms for users to ask questions in a natural language form and receive answers to those questions. Many question answering systems cache a copy of source documents so that the cached information will be available when answering questions,” Karandish para [0002]):
populating a database with data extracted from digital files ("In many embodiments, the extracted content section (e.g., 432, 433) can be transformed into vector embeddings using a suitable machine learning-based transformation algorithm," Karandish para [0086] and Karandish Fig. 3 reference character 316), ("Block 430 of text mining can include receiving the pre-processed document (e.g., 421 (FIG. 4A)) as input and automatically outputting extracted content sections that can be used as answers," Karandish para [0053]), generating a vector embedding for each semantically coherent text segment and question-answer pair ("In many embodiments, the extracted content section (e.g., 432, 433) can be transformed into vector embeddings using a suitable machine learning-based transformation algorithm," Karandish para [0086]), and ingesting the vector embeddings into the database as entries (Karandish Fig. 3 shows Ingestion System 312 and Index Database 316 within the Question Answering System);
identifying, based on the vector embeddings, a set of entries among the entries in the database, wherein each entry of the set of entries is semantically similar to the text query ("The system can advantageously store location metadata about each question-answer pair, which can beneficially allow finding and retrieving the relevant answer and information from within the source file," Karandish para [0128]);
set of entries ("The system can advantageously store location metadata about each question-answer pair, which can beneficially allow finding and retrieving the relevant answer and information from within the source file," Karandish para [0128]);
in response to a determination that the candidate entry is not associated with the identification data, marking the candidate entry as irrelevant to the text query (“For example, a pre-determined and/or dynamic relevance threshold can be applied, such that the content sections (e.g., sections of text) are either extracted or not extracted, depending on the score for the content section determined by the model. The extracted content sections (e.g., 432-436) can meet or exceed the relevance threshold, and the non-extracted content sections (e.g., 437-439) do not meet the relevance threshold,” Karandish para [0054]).
However, Karandish fails to disclose including identification data for a plurality of agricultural products; plurality of agricultural products; determining whether an element of the text query includes identification data; in response to a determination that the element includes identification data, determining whether a candidate entry of the set of entries is associated with the identification data; formulating a prompt for a large language model (LLM) based on the text query, wherein the prompt incorporates data from relevant entries of the set of entries text query.
Lewis teaches receiving a text query (Lewis Figure 1 shows receiving a query on the left);
determining whether an element of the text query includes identification data (“The retriever provides latent documents conditioned on the input, and the seq2seq model then conditions on these latent documents together with the input to generate the output,” Lewis pg. 2, the input contains information (i.e., identification data) that allows the retriever to obtain documents that are related to it);
in response to a determination that the element includes identification data, determining whether a candidate entry of the set of entries is associated with the identification data (“The retriever provides latent documents conditioned on the input, and the seq2seq model then conditions on these latent documents together with the input to generate the output,” Lewis pg. 2);
formulating a prompt for a large language model (LLM) based on the text query, wherein the prompt incorporates data from relevant entries of the set of entries (Lewis Figure 1 shows the generator, which is BART-large (an LLM) per 2.3 pg. 3 and "To combine the input x with the retrieved content z when generating from BART, we simply concatenate them," Lewis 2.3 pg. 3);
submitting the prompt to the LLM (Lewis Figure 1 shows the generator, which is BART-large (an LLM) per 2.3 pg. 3);
obtaining a response to the prompt from the LLM (Lewis Figure 1 shows the LLM's response on the right side);
and outputting the response as an answer to the text query (Lewis Figure 1 shows the LLM's response on the right side).
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to have modified Karandish’s method of text extraction and question-answer pair generation by including Lewis’ method of utilizing an LLM to formulate a prompt and then submit that prompt to an LLM. This would result in improved accuracy for the output of the LLM, as well as reduced hallucination frequency, as the model is pulling documents and information from a database of that domain.
Mol teaches including identification data for a plurality of agricultural products ("Information regarding fertilizers, seeds, pesticides, weather, soil, crop varieties and results of various scientific research in the agricultural domain is crucial for the farmers, scientists and researchers. This domain-specific structural knowledge will be beneficial for various purposes like getting the direct answers to the questions, automating farming activities, and promoting research and education in the agriculture field. Dissemination of advanced knowledge will improve productivity in the agriculture sector, and at the same time, ensuring the validity of the knowledge is also essential," Mol sec. 7 pg. 26);
plurality of agricultural products ("Information regarding fertilizers, seeds, pesticides, weather, soil, crop varieties and results of various scientific research in the agricultural domain is crucial for the farmers, scientists and researchers. This domain-specific structural knowledge will be beneficial for various purposes like getting the direct answers to the questions, automating farming activities, and promoting research and education in the agriculture field. Dissemination of advanced knowledge will improve productivity in the agriculture sector, and at the same time, ensuring the validity of the knowledge is also essential," Mol sec. 7 pg. 26).
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to have modified Karandish’s method of text extraction and question-answer pair generation by including Mol’s method of knowledge extraction in the agricultural domain. Enormous information in various domains, including agriculture, is available in the natural language from several resources (Mol Abstract). This research in the agricultural domain will be helpful in assisting farmers in decision making by providing knowledge about the recent advancements in agriculture and other associated technologies based on the latest research results (Mol sec. 1 pg. 2).
Claim(s) 10-12 is/are rejected under 35 U.S.C. 103 as being unpatentable over Karandish, in view of Lewis, and further in view of Leary et al. (US Patent Application Publication No. 2024/0095463), hereinafter referred to as Leary.
Regarding claim 10, Karandish, in view of Lewis, discloses all of the limitations of claim 7. However, Karandish fails to disclose wherein: the LLM is a first LLM, identifying the entry that is most relevant to the text query is performed by applying a third LLM. Leary discloses approaches to utilizing large language models for a multitude of tasks.
Lewis teaches wherein: the LLM is a LLM, wherein the retrieval process is performed by applying a LLM (Lewis 2.3 pg. 3 describes DPR, which utilizes LLMs (BERT)), and identifying the entry that is most relevant to the text query is performed by applying a LLM (Lewis 2.3 pg. 3 describes DPR, as well as how the documents with highest probability are determined).
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to have modified Karandish’s method of text extraction and question-answer pair generation by including Lewis’ method of utilizing an LLM to formulate a prompt and then submit that prompt to an LLM. This would result in improved accuracy for the output of the LLM, as well as reduced hallucination frequency, as the model is pulling documents and information from a database of that domain.
Leary teaches a first LLM, a second LLM, a third LLM (Leary Fig. 1 LLM service 102 shows multiple LLMs 104 being used).
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to have modified Karandish’s method of text extraction and question-answer pair generation by including Leary’s method of utilizing multiple LLMs for different tasks. Using multiple LLMs would improve accuracy and efficiency, as each model would have been trained on its respective tasks. This would also allow for LLMs to essentially run in tandem, reducing run-times and being more time effective.
Regarding claim 11, Karandish, in view of Lewis, and further in view of Leary, discloses all of the limitations of claim 10. Karandish further discloses wherein processing the extracted data to generate the question-answer pairs includes applying a LLM to generate the question-answer pairs from the text ("Block 430 of text mining can include receiving the pre-processed document (e.g., 421 (FIG. 4A)) as input and automatically outputting extracted content sections that can be used as answers," Karandish para [0053] and "In some embodiments, the machine learning algorithm used in text mining can use a number of engineered features for each section of text, such as TF-IDF (term frequency-inverse document frequency) keyword rank, word frequency count, section character length, and/or sentence vector semantic similarity (e.g., using BERT (Bidirectional Encoder Representations from Transformations) vector embeddings, as described in Jacob Devlin et al., BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding, 2018, available at https://arxiv.org/abs/1810.04805, or other suitable sentence embedding techniques)," Karandish para [0055]).
However, Karandish fails to disclose a fourth LLM.
Leary teaches a fourth LLM (Leary Fig. 1 LLM service 102 shows multiple LLMs 104 being used).
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to have modified Karandish’s method of text extraction and question-answer pair generation by including Leary’s method of utilizing multiple LLMs for different tasks. Using multiple LLMs would improve accuracy and efficiency, as each model would have been trained on its respective tasks. This would also allow for LLMs to essentially run in tandem, reducing run-times and being more time effective.
Regarding claim 12, Karandish, in view of Lewis, and further in view of Leary, discloses all of the limitations of claim 10. Karandish further discloses text query includes: identifying, among the set of entries, two entries with a highest relevance to the text query ("A machine learning algorithm (such as neural network, random forest, or another suitable algorithm) can be used to combine the sentences into groupings, and score and/or rank their relevance or importance," Karandish para [0054] and "One or more of the location delimiters (e.g., 662-664) can be used in any suitable technique to determine the answer text (e.g., 541). In many embodiments, the vector embeddings of the pre-processed document (e.g., 531) can be searched and ranked by similarity (e.g., approximate string matching) with the location delimiters (e.g., 662-664) and/or the transformed answer (e.g., 647) using a suitable similarity measure, such as cosine similarity or Manhattan distance," Karandish para [0100]);
generating, for each of the two entries, a score indicating similarity of the entry to the text query ("A machine learning algorithm (such as neural network, random forest, or another suitable algorithm) can be used to combine the sentences into groupings, and score and/or rank their relevance or importance," Karandish para [0054] and "One or more of the location delimiters (e.g., 662-664) can be used in any suitable technique to determine the answer text (e.g., 541). In many embodiments, the vector embeddings of the pre-processed document (e.g., 531) can be searched and ranked by similarity (e.g., approximate string matching) with the location delimiters (e.g., 662-664) and/or the transformed answer (e.g., 647) using a suitable similarity measure, such as cosine similarity or Manhattan distance," Karandish para [0100]);
and selecting, among the two entries, the entry with a higher text query ("For example, similarity scores can be generated for the portions of the vector embeddings of the pre-processed document (e.g., 531), and the highest ranked portion can be identified as the matching representation," Karandish para [0100] and "One or more of the location delimiters (e.g., 662-664) can be used in any suitable technique to determine the answer text (e.g., 541). In many embodiments, the vector embeddings of the pre-processed document (e.g., 531) can be searched and ranked by similarity (e.g., approximate string matching) with the location delimiters (e.g., 662-664) and/or the transformed answer (e.g., 647) using a suitable similarity measure, such as cosine similarity or Manhattan distance," Karandish para [0100]).
However, Karandish fails to disclose wherein the third LLM comprises a cross-encoder model, and the third LLM.
Lewis teaches wherein the LLM includes a cross-encoder model ("RAG compares favourably to the DPR QA system, which uses a BERT-based “crossencoder” to re-rank documents, along with an extractive reader," Lewis 4.1 pg. 5).
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to have modified Karandish’s method of text extraction and question-answer pair generation by including Lewis’ method of utilizing a cross-encoder model. Cross-encoder models offer a high accuracy due to their ability to analyze the interaction between a query and a document together. This would result in more precise relevance judgements, which makes them a top choice for re-ranking in search or retrieval systems.
Leary teaches the third LLM, and the third LLM (Leary Fig. 1 LLM service 102 shows multiple LLMs 104 being used).
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to have modified Karandish’s method of text extraction and question-answer pair generation by including Leary’s method of utilizing multiple LLMs for different tasks. Using multiple LLMs would improve accuracy and efficiency, as each model would have been trained on its respective tasks. This would also allow for LLMs to essentially run in tandem, reducing run-times and being more time effective.
Conclusion
THIS ACTION IS MADE FINAL. Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to ADAM MICHAEL WEAVER whose telephone number is (571)272-7062. The examiner can normally be reached Monday-Friday, 8AM-5PM EST.
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, Richemond Dorvil can be reached at (571) 272-7602. 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.
/ADAM MICHAEL WEAVER/Examiner, Art Unit 2658
/RICHEMOND DORVIL/Supervisory Patent Examiner, Art Unit 2658
Prosecution Insights