DETAILED ACTION
This communication is in response to the Application filed on 03/03/2025. Claims 1-21 are pending and have been examined. Claims 1, 20 and 21 are independent. This Application was published as US Pub 2026/0140978.
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 .
Information Disclosure Statement
The information disclosure statement (IDS) submitted on 07/01/2025 was filed. The submission is in compliance with the provisions of 37 CFR 1.97. Accordingly, the information disclosure statement is being considered by the examiner.
Priority
Acknowledgment is made of applicant’s claim for foreign priority based on application GB2417043.3 filed in UK Intellectual Property Office (IPO) on 11/20/2024 and receipt of a certified copy thereof.
Drawings
The drawings are objected to because Figures 1-9 include a series of numbered blocks that do not include any suitable legend for understanding the drawing.
Accordingly, these drawings fail to convey the part of invention to which they are intended to pertain without referring to the Specification.
37 C.F.R. 1.84 Standards for drawings:
(o) Legends. Suitable descriptive legends may be used subject to approval by the Office or may be required by the examiner where necessary for understanding of the drawing. They should contain as few words as possible.
To overcome the objection, refer to Figures of the cited references which are good examples of the amount of information that is considered "suitable descriptive legends" and a drawing should include.
Corrected drawing sheets in compliance with 37 CFR 1.121(d) are required in reply to the Office action to avoid abandonment of the application. Any amended replacement drawing sheet should include all of the figures appearing on the immediate prior version of the sheet, even if only one figure is being amended. The figure or figure number of an amended drawing should not be labeled as “amended.” If a drawing figure is to be canceled, the appropriate figure must be removed from the replacement sheet, and where necessary, the remaining figures must be renumbered and appropriate changes made to the brief description of the several views of the drawings for consistency. Additional replacement sheets may be necessary to show the renumbering of the remaining figures. Each drawing sheet submitted after the filing date of an application must be labeled in the top margin as either “Replacement Sheet” or “New Sheet” pursuant to 37 CFR 1.121(d). If the changes are not accepted by the examiner, the applicant will be notified and informed of any required corrective action in the next Office action. The objection to the drawings will not be held in abeyance.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1-21 are rejected under 35 U.S.C. 101 because the claimed invention is directed to abstract idea without significantly more.
Regarding Claims 1, 20 and 21,
Claim 1 recites a method, which falls under the statutory category of process. Claims 20 and 21 recite a computer program product comprising a computer-readable storage medium and a computer system, which fall under the statutory categories of manufacture and machine, respectively (Step 1: Yes).
Claims recite limitations, “(a) loading a set of questions and answers…”, “(b) generating a respective set of chunks of the documents…”, and “(c) generating a prompt for the LLM…”, “(d) providing the prompt as an input to the LLM and receiving…”, “(e) performing a comparison between the respective provisional answer…”, “(f) generating a respective overall score for…”, “(g) generating training data…”, “(h) training a machine-learning module, ML-module, using the sets…”, and “(i) performing a search for the improved set of values…”.
Except for the recitation of large language model (LLM) and machine learning module (ML-module), limitations (b), (c), (e), and (g) are concepts, which can be performed in the human mind through observation, evaluation, or judgement, or by a human using a pen and paper. The claims, under their broadest reasonable interpretation, cover the concept of an English teacher dividing documents into passages and assemble selected passages by hand, writing text that combines context and a question, comparing two answers and grading one against the other, and tabulating parameter values against scores (see MPEP 2106.04(a)(2) III).
Under its broadest reasonable interpretation when read in light of the specification, the actions recited in limitations (b), (c), (e), and (g) encompass mental processes practically performed in the human mind. In addition, limitations (f), (h), and (i) also encompass the mathematical concept or calculation since limitations involve aggregating scores (e.g., averaging), fitting a model mapping input to target (i.e., regression), and mathematical optimization, respectively. Accordingly, the claim recites an abstract idea (Step 2A, Prong one).
The judicial exception is not integrated into a practical application. In particular, limitations recites additional elements of LLM and ML-module, but they are recited at a high level of generality (i.e., generic LLM or ML-module as a combination of hardware and software of a generic computing device and generic computer components performing a generic computer functions such as processing and storing data from given input) ,for example ,without disclosing specific LLM architecture or specific model type, such that it amounts to no more than mere instructions to apply to the exception using a generic computer component.
The claim recites additional limitations (a) and (d). The additional limitations are recited at a high level of generality, and amount to mere data gathering and output, which is a form of insignificant extra-solution activity. Each of the additional limitations is no more than mere instructions to apply the exception using a generic computer component.
Accordingly, the additional element do not integrate the abstract idea into a practical application because it does not impose any meaningful limits on practicing the abstract idea and the claim is therefore directed to the judicial exception. (Step 2A: YES).
The claims do not include additional elements that are sufficient to amount to significantly more than the judicial exception because they do not include subject matter that could not be performed by a human, as discussed above with respect to integration of the abstract idea into a practical application, the additional element of using the generic computing elements to perform the claimed elements amounts to no more than mere instructions to apply the exception using a generic computer component. Mere instructions to apply an exception using a generic computer component cannot provide an inventive concept.
As noted previously, the claim as a whole merely describes how to generally link the use of the aforementioned concept to a particular technological environment or field of use. Thus, even when viewed as a whole, nothing in the claim adds significantly more (i.e., an inventive concept) to the abstract idea. The claim is not patent eligible. (Step 2B: NO).
Regarding Dependent Claims 2-19,
Claims 2-19 are dependent on supra Claim 1 and include all the limitations of the claim and further limits the elements of Claims 1. Therefore, the dependent claims recite the same abstract idea. The claim recites the additional limitations, which are no more than mere instructions to apply the exception using a generic computer component, generally linking the use of the judicial exception to a particular technological environment or field of use, insignificant extra-solution activity, or that are well understood, routine and conventional activities previously known to the industry.
No additional elements beyond the use of generic computing elements are claimed, therefore the judicial exception is not integrated into a practical application nor are the claim elements sufficient to amount to significantly more than the judicial exception. Therefore, claims are not patent eligible.
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.
Claims 1-2, 5, 7-8, 10, 12-13, 15-18 and 20-21 are rejected under 35 U.S.C. 103 as being unpatentable over Fu et al. ("AutoRag-hp: Automatic online hyper-parameter tuning for retrieval-augmented generation." Findings of the Association for Computational Linguistics: EMNLP 2024. 2024) in view of Boue et al. (US Pub 2025/0384332) further in view of Madisetti et al. (US Pat 12,405,977).
Regarding Claim 1,
Fu discloses a method for determining an improved set of values of parameters for operating a Large-Language-Model, LLM, dependent on a set of documents (Fu, 1 Introduction, "…an autonomous and self-optimizing system for RAG, termed AutoRAG...As the first step, this study focuses on hyper-parameter tuning in RAG (AutoRAG-HP)..."), the method comprising:
loading a set of questions and answers, wherein a respective answer corresponds to one of the questions and the questions relate to a content provided by the set of documents (Fu, 4.1 Experiment Setup, Dataset, "…We utilize the ALCE-ASQA and Natural Questions...Both datasets are in QA format and include candidate document chunks for each question. We use their evaluators to assess the accuracy of generated responses...");
performing repetitions, the repetitions comprising using a respective mode (Fu, 3.1 Problem Formulation, Trial, "…A trial is a single iteration in which the algorithm selects an arm, observes the associated reward, and updates its estimation...Optimal settings may be determined after a predetermined number of iterations, T, or upon meeting a specific stopping criterion..."):
wherein the generating of the set of chunks and the input content is performed according to the respective mode and the respective mode differs for a respective repetition and is specified by a respective set of values of the parameters (Fu, 3.1 Problem Formulation, Arm, "…an arm refers to a specific combination of hyper-parameters that we aim to optimize...When optimizing multiple hyper-parameters simultaneously, an arm corresponds to a combination of these hyper-parameters..."),
a value of a respective parameter specifying how the generating of the set of chunks or the input content is performed (Fu, 3.1 Problem Formulation, Fig.1, "…Since excessive compression may also remove relevant information, leading to decreased performance, it is crucial to find an optimal compression ratio (C)..."; 4.1 Experiment Setup, Search Space, "…we consider two optimization tasks based on the number of hyper-parameters: Joint optimization of (K,C): They take discrete values from K Î [1, 3, 5, 7, 9] and C Î [0.3, 0.5, 0.7, 0.9, 1]...");
generating a prompt for the LLM dependent on the input content and a question of the questions (Fu, 3.1 Problem Formulation, Fig.1, "…an example RAG system... a prompt construction module (not shown) that assembles the final prompt sent to LLMs for answer generation..."; Appendix F Prompt, "…Table 4: Prompt for ASQA. The prompt consists of Instruction, Question, and K retrieved Documents...");
providing the prompt as an input to the LLM and receiving a respective provisional answer in response from the LLM (Fu, 4.1 Experiment Setup, Base LLMs, "…We adopt GPT3.5-Turbo and GPT-4 models as base LLMs..."); and
performing a comparison between the respective provisional answer and the respective answer corresponding to that question dependent on the prompt that was generated, resulting in a score for the question (Fu, 4.1 Experiment Setup, Dataset, "…We use their evaluators to assess the accuracy of generated responses..."; 3.1 Problem Formulation, Reward, "…we formulate the reward function as a linear combination of response accuracy and input token length...");
the method further comprising: generating a respective overall score for the respective mode dependent on the resulting scores for the questions (Fu, 4.1 Experiment Setup, Ground-Truth and Evaluation Metric, "…The ground-truth parameter combinations are determined using the Grid Search method (Lecunet al., 1998), which exhaustively evaluates all hyper-parameter combinations on the entire dataset…");
Fu does not explicitly disclose the training data set pairing values of the parameters and overall score, a ML-module training, nor search for the improved set of parameter values using the trained ML-module.
Boue, in the analogous field of endeavor, discloses generating training data sets comprising for the modes the respective set of values of the parameters specifying the respective mode and the respective overall score (Boue, par [019], "…telemetry data is represented as a combination of data elements with the following schema: ( eqi, esi, ftsi, hpragi, feedbacki), where the index "i" runs overall the ( qi, ai) pairs for which the users provided a feedback value ( e.g., user feedback 228)..."; "…hyperparameter generation process 10 provides about 15-25% exploration with different hyperparameter values in the calls to generative AI model 206 (i.e., generating varied values) ..");
training a machine-learning module, ML-module, using the sets of values of the parameters of the training data sets as input and the overall scores of the training data sets as target output of the ML-module (Boue, par [021], "…hyperparameter generation process 10 trains 110 supervised machine learning model 300 to generate a feedback result or score for a given set of query embeddings, content portion embeddings, user context information, and hyperparameters...if the user feedback value is numerical, hyperparameter generation process 10 trains a regression machine learning model..."; par [022], "…with a given question embedding ( eq), source document embedding (es), user context information (fts) and RAG hyperparameter values (hprag), the trained supervised machine learning model (M) acts as a function as shown in Equation 1 which returns the expected feedback value (f,)..."); and
performing a search for the improved set of values of the parameters using the trained ML-module, wherein inputting the improved set of values of the parameters into the trained ML-module results in an improved score greater than a greatest overall score of the training data sets (Boue, par [031], "…hyperparameter generation process 10 leverages the trained surrogate model ( e.g., supervised machine learning model 300) ( (M) to determine what the optimal set of hyperparameters (hprag) are to maximize the expected user feedback...hyperparameter generation process 10 solves the following optimization problem as described in Equation 2..."; par [032], "…optimization may be carried out with gradient-free methods (e.g., Nelder-Mead, Powell that require only function calls (i.e., inference calls to M), etc.)...").
Therefore, it would have been obvious to a person of ordinary skill in the art to modify Fu's AutoRag-HP, which scores each RAG hyper-parameter combination by LLM answer accuracy against ground truth, with Boue's known technique of training a regression model on logged RAG hyperparameter values and their scores and search the trained model for the best-predicted hyperparameter while the surrogate optimization model drastically reduces computation time and provides scalable and domain-flexible model (Boue, par [030]).
Neither Fu nor Boue explicitly discloses generation of the chunks of the documents for each mode's parameter. Madisetti, in the same field of RAG optimization, discloses generating a respective set of chunks of the documents dependent on the set of documents and an input content dependent on the set of chunks of the documents (Madisetti, Fig.40, col.24:62-64, "…A document chunking module 3806 is operable to break down the received document into semantically meaningful segments…"; Fig.41, col.26:36-44, "…If unsatisfactory, the system initiates a fine-tuning process at step 3936. In the fine-tuning process, the system can tweak the models and parameters used in the topic modeling, document chunking and tagging, citation analysis, chunk selection, meta-data tagging, indexing and hybrid search stages, to improve the overall performance. The system may then iterate steps 3902 through 3912 and 3918 through 3932...").
Therefore, it would have been obvious to a person of ordinary skill in the art to add Madisetti's known score-RAG technique of adjusting document-chunking parameter and re-running the chunking step to the RAG hyperparameter optimization of Fu in view of Boue, so that each mode regenerates the chunks and input content under its own parameter values. Because Madisetti's re-chunking performs the same function of improving RAG output through how documents are chunked, the combination predictably yields modes that differ in chunk generation.
Regarding Claim 2,
Fu in view of Boue further in view of Madisetti discloses the method of claim 1, the method further comprising: receiving a new question related to the set of documents (Boue, Fig.4, par [023], "…Referring also to FIG.4, a new user query is issued (e.g., query 400)...");
generating a new set of chunks of the documents dependent on the set of documents and a new input content dependent on the new set of chunks of the documents, wherein the generating of the new set of chunks and the new input content is performed according to a new mode, and the new mode is specified by the improved set of values of the parameters, the respective value of the improved set of values of the parameters specifying how the generating of the new set of chunks or the new input content is performed (Boue, Fig.4, par [026], "…hyperparameter generation process 10 generates 114 a vector embedding (eq) (e.g., content portion embedding 408) from content portion 404 which is matched for similarity with the content so that the top-k most related source content portions are returned and aggregated into an embedding (es)..."; Madisetti, Fig.41, col.26:34-35, "…The generated response undergoes an evaluation process at step 3930 to assess its quality and relevance. The system determines if the generated response meets the required performance criteria at step 3932. If satisfactory, a final response is produced at step 3934...");
generating a new prompt for the LLM dependent on the new input content and the new question (Boue, par [033], "…the generated hyperparameters are forwarded to the generative AI model 206 along with the initial query (e.g., query 400) (q) and the top-k retrieved source content (e.g., content portion 404) in a prompt (e.g., prompt 410) as in a traditional RAG architecture...");
providing the new prompt as a new input to the LLM and receiving a new answer in response from the LLM (Boue, Fig.4, par [034], "…using prompt 410, user context information, and hyperparameters 406, generative AI model 206 generates a result ( e.g., result 412) for query 400..."); and
providing the new answer as a response to the new question (Boue, Fig.4, par [034], "…Hyperparameter generation process 10 provides 122 result 412 to user 200...").
Regarding Claim 5,
Fu in view of Boue further in view of Madisetti discloses the method of claim 1, wherein the performing of the search comprises performing a grid search, a random search, or performing the search based on a genetic algorithm (Boue, par [029], "…Optimization algorithms, such as gradient-based methods or genetic algorithms, are employed to find the input parameters that minimize the objective function. The surrogate model allows these algorithms to evaluate many candidate solutions quickly...").
Regarding Claim 7,
Fu in view of Boue further in view of Madisetti discloses the method of claim 1, the method further comprising for the respective repetition: selecting a chunk size for the chunks of the set of chunks, wherein a value of a parameter of the set of parameters specifies the chunk size (Madisetti, Fig.38, col.22:28-30, "…At step 3602, the document is split into smaller, manageable blocks/chunks. The block/chunk size is configurable, typically ranging between 10,000 to 50,000 tokens..."; Fig.47, col.29:9-15, "…chunking process 4500, which can vary based on specific needs 4502, for example a ten-page PDF can be broken into 5 three-page chunks (with overlapping text) 4506, the same PDF could be divided into 5 two-page chunks 4510..")
Regarding Claim 8,
Fu in view of Boue further in view of Madisetti discloses the method of claim 1, the method further comprising for the respective repetition: selecting chunks of the set of chunks and generating the input content dependent on the selected chunks, wherein a value of a parameter of the set of parameters specifies a number of the selected chunks (Fu, 3.1 Problem Formulation, Fig.1, "…the top-k (K) document chunks retrieved from an external knowledge base..."; 4.1 Experiment Setup, Search Space, "…we consider two optimization tasks based on the number of hyper-parameters: Joint optimization of (K,C): They take discrete values from K Î [1, 3, 5, 7, 9]..."; Appendix F Prompt, "…Table 4: Prompt for ASQA. The prompt consists of Instruction, Question, and K retrieved Documents...").
Regarding Claim 10,
Fu in view of Boue further in view of Madisetti discloses the method of claim 1, the method further comprising for the respective repetition: selecting a type of an embedding model for generating a respective embedding vector dependent on the respective chunk of the set of chunks, wherein the respective embedding vector represents the respective chunk, wherein a value of a parameter of the set of parameters specifies the type of the embedding model (Fu, 3.1 Problem Formulation, Fig.1, "…two tunable hyper-parameters: the top-k (K) document chunks retrieved from an external knowledge base and the embedding model (E) used for ranking these retrieved chunks..."; 4.1 Experiment Setup, Search Space, "… we evaluate the effects of the top-k hyper-parameter (K) and the embedding model (E) by considering three different choices, i.e., “mpnet” (Song et al., 2020), “ada_002” (OpenAI, 2022), and “contriever” (Izacard et al., 2021)...").
Regarding Claim 12,
Fu in view of Boue further in view of Madisetti discloses the method of claim 1, the method further comprising for the respective repetition: selecting a size of the provisional answer, wherein a value of a parameter of the set of parameters specifies the size of the provisional answer (Boue, par [024], "…For hyperparameters concerning inference, hyperparameter generation process 10 can customize temperature, top-k sampling, top-p sampling, a maximum number of tokens...The maximum token hyperparameter defines the maximum number of tokens to generate during inference...").
Regarding Claim 13,
Fu in view of Boue further in view of Madisetti discloses the method of claim 1, the method further comprising for the respective repetition: selecting a value of a parameter of the set of parameters, wherein the value of the parameter indicates a degree of repetition of similar context in the provisional answer (Boue, par [024], "…The repetition penalty penalizes the generative AI model for generating the same token repeatedly...").
Regarding Claim 15,
Fu in view of Boue further in view of Madisetti discloses the method of claim 1, the method further comprising for the respective repetition: selecting a value of a parameter of the set of parameters, wherein the value of the parameter indicates a degree of modification of a probability distribution of the LLM (Boue, par [024], "…temperature controls the randomness of predictions. Higher values lead to a more random output while lower values lead to more deterministic values. The top-k sampling limits the sampling pool of the top-k highest probability tokens. The top-p sampling limits the sampling pool to the smallest set of tokens with a cumulative probability above a threshold (p)...").
Regarding Claim 16,
Fu in view of Boue further in view of Madisetti discloses the method of claim 1, the method further comprising for the respective repetition: selecting a type of retrieval method, the generating of the input content comprising performing the retrieval method for obtaining the input content, wherein a value of a parameter of the set of parameters specifies the selected type of retrieval method (Madisetti, col.25:30-41, "…The search dispatcher implements different search and retrieval techniques such as vector similarity search, maximal marginal relevance (MMR) search, keyword-based search, cluster search (which groups similar documents/ chunks into clusters based on their similarity and the query is match with the most relevant clusters) and ensemble search...").
Regarding Claim 17,
Fu in view of Boue further in view of Madisetti discloses the method of claim 1, the method further comprising for the respective repetition: selecting a chunking method from a set of chunking methods for generating chunks dependent on documents and generating the set of chunks according to the selected chunking method, wherein a value of a parameter of the set of parameters specifies the selected chunking method (Madisetti, Fig.47, col.29:11-17, "…chunking process 4500, which can vary based on specific needs 4502...a ten-page PDF can be broken into 5 three-page chunks (with overlapping text) 4506, the same PDF could be divided into 5 two-page chunks 4510, the same PDF could be trimmed and split into just three 2-page chunks 4508, and/or, alternatively, the same PDF could be condensed by summarization into two 3-page chunks 4504..."; Fig.40, col.24:62-66, "…A document chunking module 3806 is operable to break down the received document into semantically meaningful segments. It includes a semantic segmenter for intelligent text division, a topic tagger operable to associate relevant topics with each chunk...").
Regarding Claim 18,
Fu in view of Boue further in view of Madisetti discloses the method of claim 1, the method further comprising for the respective repetition:
generating embedding vectors, wherein the respective embedding vector represents the respective chunk of the set of chunks (Boue, par [014], "…RAG system 208 generates a content portion embedding (e.g., content portion embedding 218) from content portion 212 using a language model (e.g., language model 220)...");
generating a further embedding vector which represents the question or a set of further embedding vectors which represent the question (Boue, par [014], "…the query text is transformed into a vector of embeddings (e.g., query embedding 216) by passing each query 204 through a language model (e.g., language model 214)
generate a vector of numbers corresponding to the dimensions for the vector embedding (e.g., query embedding 216)...");
selecting a type of similarity metric, wherein a value of a parameter of the set of parameters specifies the selected type of similarity metric (Boue, par [015], "…The score is based on the similarity, or distance, between the vector embeddings of the content portion and query. It will be appreciated that any distance metric can be used...determines the content portion similarity score using cosine similarity..."; Madisetti, col.16:40-43, "…The documents 2800 are stored in the form of vector embeddings in the vector database which allows fast searching, retrieval and comparison of documents using different similarity measures...");
performing a comparison between the embedding vectors and the further embedding vector using the similarity metric or a comparison between the embedding vectors and the set of further embedding vectors using the similarity metric (Boue, par [015], "…The score is based on the similarity, or distance, between the vector embeddings of the content portion and query. It will be appreciated that any distance metric can be used within the scope of the present disclosure. In one example, hyperparameter generation process 10 determines the content portion similarity score using cosine similarity...");
selecting a subset of the chunks dependent on a result of the comparison; and generating the input content dependent on the selected subset of the chunks (Boue, par [015], "…RAG system 208 generates a content portion embedding (e.g., content portion embedding 218) from content portion 212 using a language model (e.g., language model 220)...").
Claim 20 is a computer program product claim with limitations similar to the limitations of Claim 1 and is rejected under similar rationale. Additionally,
Boue discloses a computer program product comprising a computer-readable storage medium having computer-readable program code embodied therewith (Fig.5, paras [035-048], "…within computing system 500, are executed by one or more processors (not shown) and one or more memory architectures (not shown) included within computing system 500. Storage device 504 may include…the present disclosure may take the form of a computer program product on a computer-usable storage medium having computer-usable program code embodied in the medium...").
…
Rationale for combination is similar to that provided for Claim 1.
Claim 21 is a system claim with limitations similar to the limitations of Claim 1 and is rejected under similar rationale. Additionally,
Boue discloses a computer system for determining an improved set of values of parameters for generating an improved prompt for a Large-Language-Model, LLM, dependent on a set of documents, comprising: a processor set; a computer-readable storage media; and program instructions stored on the computer-readable storage media to cause the processor set to perform operations (Fig.5, paras [035-048], "…within computing system 500, are executed by one or more processors (not shown) and one or more memory architectures (not shown) included within computing system 500. Storage device 504 may include…the present disclosure may take the form of a computer program product on a computer-usable storage medium having computer-usable program code embodied in the medium...").
…
Rationale for combination is similar to that provided for Claim 1.
Claims 3 and 4 are rejected under 35 U.S.C. 103 as being unpatentable over Fu in view of Boue further in view of Madisetti further in view of Moharrer et al. (US Pub 2020/0380378).
Regarding Claim 3,
Fu in view of Boue further in view of Madisetti discloses the method of claim 1.
Fu discloses a regression machine learning model but does not explicitly teach training the decision tree.
Moharrer, in the analogous field of endeavor, discloses wherein the ML-module comprises a decision tree, the training of the ML-module comprising training the decision tree (Moharrer, Fig.1A, par [044], "… regressor 150 is, or is accompanied by, a random forest that learns correlations between quality scores, model hyperparameters, and/or dataset metafeatures..."; par [137], "…A random forest or random decision forest is an ensemble of learning approaches that construct a collection of randomly generated nodes and decision trees during a training phase...").
Therefore, it would have been obvious to one of ordinary skill in the art, before effective filing date of the claimed invention, to have substituted Moharrer's random-forest regressor for the unspecified model of the RAG-tuning system of Fu in view of Boue further in view of Madisetti, with the reasonable expectation of success, to yield the predictable result of a decision-tree ML-module predicting mode scores since both perform the same score-prediction function.
Regarding Claim 4,
Fu in view of Boue further in view of Madisetti further in view of Moharrer discloses the method of claim 3, wherein the ML-module comprises a set of decision trees comprising the decision tree (Moharrer, par [137], "…Different decision trees of a forest are constructed to be each randomly restricted to only particular subsets of feature dimensions of the data set, such as with feature bootstrap aggregating (bagging)"),
the training of the ML-module comprising generating proper subsets of the training data sets, wherein the proper subsets are random samples of the training data sets, and training the respective decision tree using the respective subset of the training data sets (Moharrer, par [138], "…Random forest hyper-parameters may include: number-of-trees-in-the-forest, maximum-number-of-features-considered-for-splitting-a-node, number-of-levels-in-each-decision-tree, minimum-number-of-data-points-on-a-leaf-node, method-for-sampling-data-points, etc...").
Claim 6 is rejected under 35 U.S.C. 103 as being unpatentable over Fu in view of Boue further in view of Madisetti further in view of Moharrer further in view of Breiman ("Random forests." Machine learning 45.1 (2001): 5-32).
Regarding Claim 6,
Fu in view of Boue further in view of Madisetti discloses the method of claim 1, wherein the parameters of the set of parameters represent input features of the ML-module (Boue, par [022], "…with a given question embedding ( eq), source document embedding (es), user context information (fts) and RAG hyperparameter values (hprag), the trained supervised machine learning model (M) acts as a function...") but does not explicitly teach a feature importance score indicating a relative impact of the input feature on prediction.
Moharrer discloses the training of the ML-module comprising determining a feature importance score for a respective input feature, wherein a feature importance score indicates a relative impact of the respective input feature on a prediction of the ML-module (Moharrer, Fig.1A, par [044], "…The random forest may naturally provide ML explainability ( MLX ) for the model and / or the dataset . For example , a random forest may reveal which hyperparameters and / or metafeatures are most or least significant for affecting quality scores...").
But Moharrer does not explicitly disclose a numeric importance score. Breiman, in the analogous field of endeavor, discloses the per-variable importance score (Breiman, 10. Exploring the random forest mechanism, "…After each tree is constructed, the values of the mth variable in the out-of-bag examples are randomly permuted and the out-of-bag data is run down the corresponding tree…The output is the percent increase in misclassification rate as compared to the out-of-bag rate (with all variables intact) (i.e., score)…").
Therefore, it would have been obvious to one of ordinary skill in the art, before effective filing date of the claimed invention, to have modified the RAG-tuning system of Fu-Boue-Madisetti with Moharrer's forest surrogate and Breiman's permutation importance, as suggested by Moharrer's teaching that forests reveal which hyperparameters most affect score, with a reasonable expectation of success, to yield the predictable result of a feature-importance score per RAG parameter, guiding tuning.
Claim 9 is rejected under 35 U.S.C. 103 as being unpatentable over Fu in view of Boue further in view of Madisetti further in view of Lyu (“CRUD-RAG: A Comprehensive Chinese Benchmark for Retrieval-Augmented Generation of Large Language Models”, Art. no. arXiv:2401.17043, 2024).
Regarding Claim 9,
Fu in view of Boue further in view of Madisetti discloses the method of claim 1 but does not explicitly disclose the common part of the chunk, which is comprised by another chunk (i.e., overlap).
Lyu, in the analogous field of endeavor, discloses the method further comprising for the respective repetition: selecting a size of common parts of the set of chunks (Lyu, 4.1 Experimental Settings, "…Chunk size: The RAG system splits the external knowledge into chunks of a certain length and stores them in a vector database…Chunk overlap: Chunk overlap refers to the shared tokens between two consecutive text chunks and is used to ensure semantic coherence when chunking…Chunk overlap: 0%, 10%, 30%, 50%, 70%..."; ); and
performing the generating of the set of chunks such that the respective chunk of the set of chunks comprises at least one common part, which is comprised by another chunk of the set of chunks, wherein a value of a parameter of the set of parameters specifies the size of the at least one common part (Lyu, 4.3 Analyzing the Impact of Chunk Overlap on RAG Performance in Different Tasks, "…Chunk overlap is the number of tokens that two adjacent chunks share. To keep the text semantics coherent, adjacent chunks have some overlap. Chunk overlap determines the size of this overlap…").
Therefore, it would have been obvious to one of ordinary skill in the art, before effective filing date of the claimed invention, to have tried Lyu's chunk overlap as a mode parameter of the Fu-Boue-Madisetti RAG-tuning system, with a reasonable expectation of success, to yield the predictable result of modes varying overlap size since overlap preserves semantic coherence and changes answer accuracy (Lyu, Section 4.3).
Claims 11 and 19 are rejected under 35 U.S.C. 103 as being unpatentable over Fu in view of Boue further in view of Madisetti further in view of Kusupati et al. ("Matryoshka representation learning." Advances in Neural Information Processing Systems 35 (2022): 30233-30249).
Regarding Claim 11,
Fu in view of Boue further in view of Madisetti discloses the method of claim 1 but does not explicitly disclose the parameter values specifying the respective size of the embedding vectors.
Kusupati, in the analogous field of representation learning for retrieval, discloses the method further comprising for the respective repetition: selecting a respective size of embedding vectors, wherein a respective embedding vector represents the respective chunk of the set of chunks (Kusupati, Fig.1, 1 Introduction, "…We introduce Matryoshka Representation Learning (MRL) to induce flexibility in the learned representation...The first m-dimensions, m ∈ [d], of the Matryoshka Representation is an information-rich low-dimensional vector, at no additional training cost, that is as accurate as an independently trained m-dimensional representation..."; 4.3.1 Adaptive Retrieval, "…using a lower-dimensional representation, e.g. Ds = 16 followed by reranking with a higher capacity representation..."), and
wherein a respective value of a parameter of the set of parameters specifies the respective size of the embedding vectors (Kusupati, Appendix E, "…We generated the database and query sets, containing N and Q samples respectively...We specify the representation size at which we retrieve a shortlist of k-nearest neighbors (k-NN) by Ds...").
Therefore, it would have been obvious to one of ordinary skill in the art, before effective filing date of the claimed invention, to have applied Kusupati's size-selectable Matryoshka embeddings to the RAG-tuning system of Fu-Boue-Madisetti because it already tunes the chunk-ranking embedding (Fu, section 3.1), with a reasonable expectation of success, to yield the predictable result of modes varying embedding-vector size. Smaller embeddings cut retrieval compute and memory cost at comparable accuracy (Kusupati, section 4.3.1).
Regarding Claim 19,
Fu in view of Boue further in view of Madisetti discloses the method of claim 1 but does not explicitly teach a selectable embedding-model value.
Kusupati, in the analogous field of representation learning for retrieval, discloses the method further comprising for the respective repetition: performing a selection of a value of a parameter of an embedding model for generating a respective embedding vector dependent on the respective chunk of the set of chunks, wherein the respective embedding vector represents the respective chunk (Kusupati, 4.1 Representation Learning, "…We use M={8,16,32,64,128,256,512,1024,2048} and M = {12,24,48,96,192,384,768} as the explicitly optimized nested dimensions respectively..."; Appendix A, Algorithm 2, "…self.nesting_list=nesting_list # set of m in M (Eq. 1)..."),
wherein a value of a parameter of the set of parameters specifies the value of the parameter of the embedding model (Kusupati, Appendix E, "…We generated the database and query sets, containing N and Q samples respectively...We specify the representation size at which we retrieve a shortlist of k-nearest neighbors (k-NN) by Ds...").
Rationale for combination is similar to that provided for Claim 11.
Claims 14 are rejected under 35 U.S.C. 103 as being unpatentable over Fu in view of Boue further in view of Madisetti further in view of Wang et al. ("Cost-effective hyperparameter optimization for large language model generation inference." International Conference on Automated Machine Learning. PMLR, 2023).
Regarding Claim 14,
Fu in view of Boue further in view of Madisetti discloses the method of claim 1 but does not explicitly teach selecting a type of the LLM.
Wang, in the analogous field of LLM inference-hyperparameter tuning, discloses the method further comprising for the respective repetition: selecting a type of the LLM, wherein a value of a parameter of the set of parameters specifies the selected type of the LLM (Wang, 2.2 How Do Hyperparameters Affect Text Generation Performance?, "…(1) model- this is a required input, specifying the model ID to use..."; Table 1: Default search space for the optimization framework, "…GPT models with diverse cost-quality trade-off...").
Therefore, it would have been obvious to one of ordinary skill in the art, before effective filing date of the claimed invention, to have modified the RAG-tuning system of Fu-Boue-Madisetti with a model selection hyperparameter of Wang, as suggested by Fu's statement , citing Wang, that its LLM API parameters are tunable (Fu, section 4.1), with a reasonable expectation of success, to yield the predictable result of modes varying LLM type (Wang, Table1).
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to JANGWOEN LEE whose telephone number is (703)756-5597. The examiner can normally be reached Monday-Friday 8:00 am - 5:00 pm ET.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, BHAVESH MEHTA can be reached at (571)272-7453. 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.
/JANGWOEN LEE/ Examiner, Art Unit 2656
/BHAVESH M MEHTA/ Supervisory Patent Examiner, Art Unit 2656
Prosecution Insights