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 .
Election/Restrictions
Applicant's election with traverse of invention II in the reply filed on 24 August 2026 is acknowledged.
The traversal is on the ground(s) that the record does not establish a serious search or examination burden. Applicant argues that the separate classification of the two inventions does not show that each invention has attained recognition in the art as a separate subject for inventive effort, because the two classifications identified lie within the same CPC main group. Applicant further argues that claim 1 and claim 16:
“recite the same computing architecture, namely at least one processor, a communication interface communicatively coupled to the at least one processor, and a memory device storing executable code. Both claims recite receiving, via a user device of a user, the selection of a subset of documents from a corpus of documents. Both claims recite processing documents of that subset via a neural network, and both recite, in identical language, that the neural network ‘is tuned in accordance with one or more defined rules’. Both claims recite generating a textual response”.
Applicant concludes that a search to the elected invention would therefore be conducted in the same art, over the same subject matter, and using the same search terms that a search directed to invention II would require.
This is not found persuasive because invention I and invention II, while sharing some superficial elements, are directed to analyzing distinct inputs to generate distinct results. As discussed in the restriction requirement, invention I is directed to a process for comparing a vectorized input query to a selected subset of documents from a corpus of documents in order to generate a textual response to the textual query. Invention II is directed to a process for determining a rationale for why each document in a selected subset of documents received a specific relevance score in response to a (previously performed) document query. As described by Applicant’s specification, invention II provides an auditing process to determine rationales for why certain answers were provided to users. Applicant argues that both claims recite receiving a selection of a subset of documents from a corpus of documents. However, claim 1 recites selecting a subset of documents “to be included in an electronic document search”, while claim 16 recites selecting a subset of documents “to be analyzed in response to the one or more textual inputs” where the “one or more textual inputs” are “textual inputs indicating whether one or more documents identified in a response to a document query is responsive to the document query”. These are two distinct subsets of documents, selected for two distinct processes. Applicant argues that both claims “recite processing documents of that subset via a neural network” where the neural network is “tuned in accordance with one or more defined rules”. As just discussed, “that subset” is actually two distinct subsets. Additionally, the mere presence of a “neural network” that is “tuned in accordance with one or more defined rules”, where the neural network is not further defined, and without any further specification of the “one or more defined rules” does not make inventions I and II indistinct. Rather, claim 1 recites the neural network generates a textual response “incorporating language derived from the subset of documents”, while claim 16 recites the neural network “determine[s] a rationale for why each respective document of the subset of documents received a specific relevance score in response to the document query”. Since the neural networks recited in claims 1 and 16 receive distinct input data and generate distinct outputs, there is no reason to assume that the “one or more defined rules” used to tune the neural networks are the same one or more rules in claims 1 and 16.
Invention I would require searches related to query execution using a vector-based model, classified in G06F16/3347. Invention II, which includes no limitations related to vectors, would require searches related to query formulation using relevance feedback from the user, classified in G06F16/3326. Inventions I and II are therefore distinct, and there would be a serious search and/or examination burden on the examiner if restriction is not required.
The requirement is still deemed proper and is therefore made FINAL.
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, 3, 5, 8-11, and 14-15 is/are rejected under 35 U.S.C. 103 as being unpatentable over Levinson et al. (U.S. Patent Application Pub. No. 2025/0061139, hereinafter “Levinson”), in view of Perez et al. (Training and Inference of Large Language Models using 8-bit Floating Point, hereinafter “Perez”).
In regard to claim 1, Levinson discloses an electronic file management system incorporating semantic vectors (Fig. 1, 100), the system comprising:
at least one processor (Fig. 6, processor 610, paragraph [0125]);
a communication interface communicatively coupled to the at least one processor (communications interfaces 616, paragraph [0125]); and
a memory device storing executable code (memory 620, paragraph [0125]) that, when executed, causes the at least one processor to:
receive, via a user device of a user, an indication for selection of a subset of documents from a corpus of documents, the subset of documents to be included in an electronic document search (see Fig. 5B, a user selects a subset of documents to which a subsequent search should be scoped, paragraphs [0091-0092]);
filter the subset of documents according to predefined filter rules resulting in filtered documents (a filter is applied to limit to the results to content from document IDs in the selected subset, paragraph [0068]);
apply a semantic vector analysis module to text of the filtered documents to build respective vectors for each document of the filtered documents (documents from the document corpus are semantically embedded as document vectors, paragraph [0045]);
store the respective vectors to a vector database configured for one or more vectorized queries to be performed thereon, wherein the vector database is specific to the filtered documents of the electronic document search (an embedding store stores the document vectors, paragraph [0043]; where the semantic search identifies only the embeddings associated with the subset of selected document IDs, paragraph [0099]);
receive, via the user device, a textual query and based thereon vectorize the textual query to produce a query vector that represents search criteria (a semantic search request is received from the user and converted to a semantic vector, paragraphs [0096-0097]);
search, in accordance with the search criteria of the query vector, the vector database to identify similar documents from the subset of documents that satisfy a similarity measure, the searching comparing similarities of the query vector to the respective vectors, and generate a ranked list of vectors comprising highest similarity scores relative to the query vector, the ranked list identifying the similar documents (a semantic search engine searches the documents in the embedding data store using the query vector to identify responsive semantic results, paragraphs [0098] and [0100]; then provides a collection of semantically relevant documents as a result, paragraphs [0102-0103]; the semantically relevant documents being identified using a cosine similarity between the embedding vectors, paragraph [0058]);
process, via a neural network, document data of the similar documents to generate a textual response to the textual query, the textual response incorporating language derived from the subset of documents (the documents determined in the semantic search are provided to an LLM to generate a response using the language from the semantically relevant documents, paragraphs [0103-0106]); and
transmit one or more control signals to the user device to initiate displaying, via the user device, the generated textual response and a history of question answer events, the history including the textual query and the textual response (see Figs. 5D and 5F, where the generated textual response is displayed to the user in a chat history format, paragraphs [0107], [0111], and [0123]).
Levinson does not expressly disclose the document vectors and query vector are floating point vectors. Levinson also does not expressly disclose the neural network of the LLM is tuned in accordance with one or more defined rules.
Perez discloses an LLM neural network that utilizes floating point vectors, wherein the LLM neural network is tuned in accordance with one or more defined rules (8-bit floating point formats, section 1, 2nd paragraph; where the LLM is fine-tuned utilizing the rules defined in Fig. 1, section 3.2).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to utilize floating point vectors and tune the neural network in accordance with one or more pre-defined rules, because it would provide significant efficiency gains for the training of the neural network, as taught by Perez (section 1, 1st paragraph).
In regard to claim 3, Levinson discloses the semantic vector analysis module incorporates a model configured to produce contextualized word vectors by encoding (i) each word’s position within the text of the filtered documents, and (ii) each word of the text of the filtered documents (the semantic embeddings are created by generating embeddings of normalized text chunks of each document, then creating an index mapping the document text vectors to the text chunks from which the document text vectors were generated in the document, paragraphs [0049-0050]).
In regard to claim 5, Levinson discloses the semantic vector analysis module is multi-threaded such that the semantic vector analysis module is configured to concurrently perform textual processing on multiple documents (processing is performed in parallel, paragraph [0140]).
In regard to claim 8, Levinson discloses the executable code, when executed, further causes the at least one processor to initiate displaying, via a user interface of the user device (see Fig. 5B):
textual representation of the one or more defined rules (specific tags available for filtering with textual descriptions, paragraph [0091]); and
a list indicating the similar documents and as well as one or more documents from which information for the textual response was derived (see Fig. 5F, dropdown lists of direct sources and related excerpts, paragraph [0110]).
In regard to claim 9, Levinson discloses the executable code, when executed, further causes the at least one processor to generate, via the neural network, and initiates displaying, via the user interface, a textual summary of at least one document of either the similar documents or the one or more documents from which the information for the textual response was derived (5D, generative text describing the sources 562 as well as snippets from the documents 564, paragraphs [0109-0111]).
In regard to claim 10, Levinson discloses a respective control input is associated with each of the similar documents and the one or more documents from which the information for the textual response was derived, wherein each respective control input is actionable thereby enabling a user of the user device to perform one or more user actions thereon (the user can select the direct source citations or related document citations to perform further actions with the documents, paragraphs [0112-0113]).
In regard to claim 11, Levinson discloses the executable code, when executed, further causes the at least one processor to initiate displaying a free text field configured to enable a user to provide one or more customized inputs associated with a question answer event of the history of question answer events (see Fig. 5C, query bar 556 for querying the AI assistant, paragraph [0086]; in which the user can add text from the response history, paragraphs [0112-0113]).
In regard to claim 14, Levinson discloses any known embedding models may be used to generate the document vectors, but does not expressly disclose the neural network comprises a decoder-only transformer model.
Perez discloses the neural network comprises a decoder-only transformer model (GPT model, which is a decoder-only transformer model, section 4.1).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to utilize a decoder-only transformer model, because such models are widely known to provide state-of-the-art results on a variety of natural language processing tasks.
In regard to claim 20, Levinson discloses a computer-implemented method, comprising:
receiving, via a user device of a user, an indication for selection of a subset of documents from a corpus of documents, the subset of documents to be included in an electronic document search (see Fig. 5B, a user selects a subset of documents to which a subsequent search should be scoped, paragraphs [0091-0092]);
filtering the subset of documents according to predefined filter rules resulting in filtered documents (a filter is applied to limit to the results to content from document IDs in the selected subset, paragraph [0068]);
applying a semantic vector analysis module to text of the filtered documents to build respective vectors for each document of the filtered documents (documents from the document corpus are semantically embedded as document vectors, paragraph [0045]);
storing the respective vectors to a vector database configured for one or more vectorized queries to be performed thereon, wherein the vector database is specific to the filtered documents of the electronic document search (an embedding store stores the document vectors, paragraph [0043]; where the semantic search identifies only the embeddings associated with the subset of selected document IDs, paragraph [0099]);
receiving, via the user device, a textual query and based thereon vectorize the textual query to produce a query vector that represents search criteria (a semantic search request is received from the user and converted to a semantic vector, paragraphs [0096-0097]);
searching, in accordance with the search criteria of the query vector, the vector database to identify similar documents from the subset of documents that satisfy a similarity measure, the searching comparing similarities of the query vector to the respective vectors, and generate a ranked list of vectors comprising highest similarity scores relative to the query vector, the ranked list identifying the similar documents (a semantic search engine searches the documents in the embedding data store using the query vector to identify responsive semantic results, paragraphs [0098] and [0100]; then provides a collection of semantically relevant documents as a result, paragraphs [0102-0103]; the semantically relevant documents being identified using a cosine similarity between the embedding vectors, paragraph [0058]);
processing, via a neural network, document data of the similar documents to generate a textual response to the textual query, the textual response incorporating language derived from the subset of documents (the documents determined in the semantic search are provided to an LLM to generate a response using the language from the semantically relevant documents, paragraphs [0103-0106]).
Levinson does not expressly disclose the document vectors and query vector are floating point vectors. Levinson also does not expressly disclose the neural network of the LLM is tuned in accordance with one or more defined rules.
Perez discloses an LLM neural network that utilizes floating point vectors, wherein the LLM neural network is tuned in accordance with one or more defined rules (8-bit floating point formats, section 1, 2nd paragraph; where the LLM is fine-tuned utilizing the rules defined in Fig. 1, section 3.2).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to utilize floating point vectors and tune the neural network in accordance with one or more pre-defined rules, because it would provide significant efficiency gains for the training of the neural network, as taught by Perez (section 1, 1st paragraph).
Claim(s) 2, 4, 7, and 12 is/are rejected under 35 U.S.C. 103 as being unpatentable over Levinson, in view of Perez, and further in view of Kislal et al. (U.S. Patent Application Pub. No. 2024/0012842, hereinafter “Kislal”).
In regard to claim 2, Levinson discloses any known embedding models may be used to generate the document vectors, but Levinson and Perez do not expressly disclose a transformer-based language model.
Kislal discloses a semantic vector analysis module that comprises a transformer-based language model configured to pre-train deep bidirectional representations from the unlabeled text data by jointly conditioning on both left and right context in all layers of the transformer-based language model (vectors for a semantic search engine are generated using Bidirectional Encoder Representations from Transformers (BERT), paragraph [0163]).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to utilize a transformer-based language model configured to pre-train deep bidirectional representations from the unlabeled text data by jointly conditioning on both left and right context in all layers of the transformer-based language model, because, as is widely recognized in the art, BERT models provide state-of-the-art performance on a wide range of natural language processing tasks.
In regard to claim 4, Levinson and Perez do not disclose the vector database comprises vector data categorized according to vector distance, wherein the vector distance is associated with similarity of the vector data.
Kislal discloses a vector database for querying documents, wherein the vector database comprises vector data categorized according to vector distance, wherein the vector distance is associated with similarity of the vector data (documents are clustered in a database based on distance metrics, paragraph [0103]).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to categorize the vector data according to vector distance, because it would allow the documents to be automatically classified into discovered categories, as taught by Kislal (paragraphs [0044-0045]).
In regard to claim 7, Levinson and Perez do not disclose the vector database consists only of respective floating point vectors specific to the filtered documents of the electronic document search, thereby excluding documents unrelated to the electronic document search.
Kislal discloses the vector database consists only of respective floating point vectors specific to the filtered documents of the electronic document search, thereby excluding documents unrelated to the electronic document search (an administrator selects documents from a corpus that are to be searched, and the documents are further filtered to remove, for example, sensitive content, and a vector database is built only for the filtered documents, paragraphs [0123-0124]).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to exclude documents unrelated to the electronic document search, because it would enhance security by allowing a user to search a private document repository in a secured manner inaccessible by outside parties, as taught by Kislal (paragraph [0116]).
In regard to claim 12, Levinson and Perez do not disclose to initiate displaying a control input configured to classify each question answer event of the history of question answer events.
Kislal discloses to initiate displaying a control input configured to classify each question answer event of the history of question answer events (an interactive user query interface allows a user to provide indications of concept/category values for each query and corresponding result, paragraphs [0131-0132]).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to display a control input configured to classify each question answer event of the history of question answer events, because it would allow the user to disambiguate the correct category for a question and answer, which would then allow the system to improve itself based on the category, as taught by Kislal (paragraph [0131]).
Claim(s) 6 is/are rejected under 35 U.S.C. 103 as being unpatentable over Levinson, in view of Perez, and further in view of Chan et al. (U.S. Patent Application Pub. No. 2021/0406266, hereinafter “Chan”).
In regard to claim 6, Levinson and Perez do not disclose the filtered documents comprise datasets that include multiple rows and columns of data having associated header information.
Chan discloses a vector database for datasets that include multiple rows and columns of data having associated header information (tables of information, paragraph [0043]), and where a semantic vector analysis module is configured to embed each row of the multiple rows individually to maintain context of values included in the rows and columns as defined by the associated header information (a row feature aggregation component encodes a feature vector representing each row in a table, paragraph [0055]; in context with each column header, paragraph [0058]).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to embed each row of multiple rows and columns of data having associated header information, because it would allow semantic searching of table format documents, as taught by Chan (paragraphs [0003-0005]).
Claim(s) 13 is/are rejected under 35 U.S.C. 103 as being unpatentable over Levinson, in view of Perez, and further in view of Ormerod (U.S. Patent Application Pub. No. 2022/0245350).
In regard to claim 13, while Perez discloses the one or more rules are configurable (hyperparameters, etc. are configurable, section 4.1), Levinson and Perez do not expressly disclose the one or more defined rules are configurable via a control input configured to be displayed, via a user interface of the user device, such that the one or more defined rules are capable of being modified by a user of the user device.
Ormerod discloses a system for configuring defined rules for language models, wherein the one or more defined rules are configurable via a control input configured to be displayed, via a user interface of the user device, such that the one or more defined rules are capable of being modified by a user of the user device (Fig. 1, a user interface 102 allows users to configure natural language processing neural networks according to defined rules, paragraphs [0039-0040]).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to provide a control input configured to be displayed via a user interface of the user device, such that the one or more defined rules were capable of being modified by a user of the user device, because it would allow nonskilled artisans to configure the models, as taught by Ormerod (paragraphs [0041-0042]).
Claim(s) 15 is/are rejected under 35 U.S.C. 103 as being unpatentable over Levinson, in view of Perez, and further in view of Miller (U.S. Patent Application Pub. No. 2025/0181975, hereinafter “Miller”).
In regard to claim 15, Levinson and Perez do not disclose determining whether any documents satisfy one or more privileged conditions.
Miller discloses a system for processing documents, comprising executable code, wherein the executable code, when executed, further causes the at least one processor to:
derive, via the neural network, whether any documents of the subset of documents satisfy one or more privileged conditions, the one or more privileged conditions being predefined according to one or more conditions (electronic data comprising documents are monitored for event identifiers indicating the document satisfies a privileged condition, using a machine learning model comprising a neural network, paragraphs [0028], [0032], and [0038]); and
initiate displaying, via a user interface of the user device, a visual indication marking one or more privileged documents that satisfy the one or more privileged conditions (documents determined to be privileged are marked indicating the document is privileged, paragraph [0029]).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to determine whether any documents of the subset of documents satisfied one or more privileged conditions and displaying a visual indication thereof, because it would ensure enterprises maintained legal privilege for particular documents, as taught by Miller (paragraphs [0003-0005]).
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Vargas et al., Jablokov et al., Hoppe, Nahamoo et al., and Wade et al. disclose additional vector database systems for searching documents and/or litigation document search.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to BRIAN LOUIS ALBERTALLI whose telephone number is (571)272-7616. The examiner can normally be reached M-F 8AM-3PM, 4PM-5PM.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, 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.
BLA 10/28/26
/BRIAN L ALBERTALLI/ Primary Examiner, Art Unit 2656