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 .
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 1-20 are rejected under 35 U.S.C. 101.
Claims 1, 4 and 17 are rejected under 101 Abstract Idea. All three claims are directed to collecting information, organizing it into searchable form, mathematically representing it as embeddings, searching/ranking it, using the retrieved information to form an answer, and providing the answer. Applying the Alice/Mayo framework by asking whether the claim adds “significantly more”.; The Office treats claims focused on collecting, analyzing, and displaying information as abstract when they do not claim a specific technological improvement.
Claim 1 describes information processing: chunking code, summarizing it, embedding it, searching it, ranking search results, augmenting a language-mode prompt, and returning a response in an IDE chat.
Claim 4 describes a generic RAG workflow: receive an NLQ, embed it, search a vector database, augment the query, predict an answer, and provide it.
Claim 17 describes adds a system form with computing devices, a storage service, and a natural-language-query service in a multi-tenant provider network, but that mainly places the same abstract information processing steps into a generic service architecture. Merely using a computer as a tool, or limiting an abstract process to a technological environment, is not enough to integrate the abstract idea into a practical application.
The claims also lack an inventive concept because the additional elements use known computer/AI components for their ordinary functions: indexes, vector databases, embeddings, lexical search, semantic search, hybrid scoring, prompts, language models, chat interfaces, storage services and provider network services. The claims do not clearly recite a new data structure, new embedding technique, new search algorithm, improved computer performance, or specific technical mechanism for improving the IDE or provider network.
The claim(s) does/do not include additional elements that are sufficient to amount to significantly more than the judicial exception because the claims are (i) mere instructions to implement the idea on a computer, and/or (ii) recitation of generic computer structure that serves to perform generic computer functions that are well-understood, routine, and conventional activities previously known to the pertinent industry. Viewed as a whole, these additional claim element(s) do not provide meaningful limitation(s) to transform the abstract idea into a patent eligible application of the abstract idea such that the claim(s) amounts to significantly more than the abstract idea itself. Therefore, the claim(s) are rejected under 35 U.S.C. 101 as being directed to non-statutory subject matter. There is further no improvement to the computing device.
Dependent claims 2-3, 5-16 and 18-20 further recite an abstract idea performable by a human and do not amount to significantly more than the abstract idea as they do not provide steps other than what is conventionally known.
Claims 2-3: claims merely add more information processing steps or task labels to the abstract idea of claim 1. Rewriting a query using context, and identifying the query as a request for code explanation, refactoring, knowledge gathering, code analysis, code generation, or debugging, do not add a specific technological improvement or inventive computer functionality beyond using generic computer/AI tools to process and present information.
Claims 5-16: the claims only add more information processing details to the abstract idea of claim 4. The claims cover rewriting a query using context, performing lexical and/or sematic search, generating or explaining code, debugging, receiving the query through an IDE, ingesting and chunking code/documents, summarizing chunks, creating embeddings, indexing information in a vector database, using a search/database service, using a fine-tuned model, and defining enterprise context as code plus related documents. These are all steps of collecting, organizing, searching, analyzing, and presenting information using generic computer and AI tools.
The add limitations do not recite a specific improvement to computer technology, the IDE, the vector database, the search system, or the language model itself. Instead, they use conventional components, ML models, embeddings, lexical/semantic search, vector databases, services, and IDEs, as tools to carry out the same abstract information retrieval and response generation process.
Claims 18-20: they only identify the type of user request, code refactoring, knowledge gathering, or code analysis, within the abstract NLQ processing system of Claim 17. These limitations do not add a specific technological improvement; they merely use generic computer/AI tools to process and respond to information requests.
Claim Rejections - 35 USC § 102
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 the appropriate paragraphs of 35 U.S.C. 102 that form the basis for the rejections under this section made in this Office action:
A person shall be entitled to a patent unless –
(a)(1) the claimed invention was patented, described in a printed publication, or in public use, on sale, or otherwise available to the public before the effective filing date of the claimed invention.
Claim(s) 4 and 8 is/are rejected under 35 U.S.C. 102(a)(1) as being anticipated by Boiko (“Method of Parallel Hybrid Search for Large-Scale Code Repositories”; Radio Electronics, Computer Science, Control; Sept. 22, 2025).
Claim 4,
Boiko teaches a computer-implemented method comprising: receiving a request to answer a natural language query (NLQ) ([3.] Boiko states that the retrieval phase identifies relevant code blocks “based on user-supplied natural language queries”; [3.] [Fig. 4] Boiko further states: “Upon receiving a natural language query, the system performs keyword-based retrieval by submitting the query to a full-text search engine, such as Elasticsearch”; [Fig. 4] Fig. 4 also shows the steps “Accept user natural language code search query”);
generating one or more embedding vectors for the NLQ ([3.] [Eq. 11] Boiko states: “In parallel, the system conducts semantic retrieval by encoding the user query into a high dimensional embedding using the same embedding model employed during indexing”; Boiko further states: “The resulting query embedding is then compared against the stored file embeddings in a vector database using cosine similarity as the distance metric”);
performing a search of a vector database having information regarding one or more of enterprise context, natural language summaries of enterprise context, embedding vectors for the enterprise context, and embedding vectors for natural language summaries of the enterprise context according to the request to generate a search result ([3.] Boiko states: “The system employs Qdrant as the underlying vector database to support high-performance semantic code search.”; Boiko further states that Qdrant stores and queries “dense embedding vectors derived from code summaries”; [3.] Boiko states: “The system stores not only the normalized embeddings of summarized code chunks, but also custom metadata (payload) including the full file path, the name and type of the code block (e.g. class, method), and the corresponding line range within the file”; see Fig. 2; [3.] Boiko also states: “After this phase, this indexed codebase contain structured, searchable entries for each file, consisting of its metadata, summaries and their dense vectors, and keyword-indexed content”);
augmenting the NLQ using the search result ([3.] Boiko states: “To enable this refinement, each retrieved code block is further analyzed using an LLM, which operates on the full content of it, its summary, and the original user query”; Boiko also states: “The model receives as input the user’s natural language query, the code block summary that was previously generated and indexed, and the source code in a line range”; Fig. 5 shows the LLM prompt including “summary”; “source code”; Line range; see Fig. 5);
predicting a response for the NLQ using the search result augmented NLQ ([3.] [eq. 14] Boiko states that “every code block retrieved by the initial hybrid search is further refined using LLM,” and that “the resulting set contains the final, context-aware, and human-readable answers that the system returns to the user.” Boiko further states that the LLM determines “which parts of the code are most likely to satisfy the sematic intent of the query” by prompting the model with “both the query and the block-level context”); and
providing the predicted response ([3.] [Fig. 4] Boiko states that the LLM-refined result set contain “human-readable answers that the system returns to the user”; Fig. 4 shows the retrieval phase ending with “Display results”).
Claim 8,
Boiko further teaches the computer-implemented method of claim 4, wherein the performing a search of a vector database having information regarding one or more of enterprise context, natural language summaries of the enterprise context, embedding vectors for the enterprise context, and embedding vectors for natural language summaries of the enterprise context according to the request to generate a search result comprises performing a lexical search and a semantic search ([Abstract] Boiko states in the Abstract: “A two-path search method with post-processing is proposed, where textual keyword search and embedding-based semantic search are executed in parallel”; [3.] Boiko further states: “This phase integrates two parallel retrieval processes: traditional keyword-based search and semantic vector-based search”; For lexical search Boiko states: “Upon receiving a natural language query, the system performs keyword-based retrieval by submitting the query to a full-text search engine, such as Elasticsearch”; For semantic search, Boiko states: “In parallel, the system conducts semantic retrieval by encoding the user query into a high-dimensional embedding,” and comparing the query embedding against stored embeddings in a vector database; [3.] Boiko also states that the “proposed hybrid code retrieval method integrates structured indexing, dual-mode retrieval, and LLM-assisted post-processing,” and uses “parallel keyword-based and vector-based search mechanisms”).
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) 5 is/are rejected under 35 U.S.C. 103 as being unpatentable over Boiko (“Method of Parallel Hybrid Search for Large-Scale Code Repositories”; Radio Electronics, Computer Science, Control; Sept. 22, 2025) and further in view of Wu et al. (“CONQRR: Conversational Query Rewriting for Retrieval with Reinforcement Learning”; Dec. 2022; Association for Computational Linguistics).
Claim 5,
Boiko teaches all the limitations in claim 4. The difference between the prior art and the claimed invention is that Boiko does not explicitly teach using contextual information to re-write the received natural language query.
Wu teaches using contextual information to re-write the received natural language query ([Introduction] Wu states: that “query rewriting (QR) in a conversation … is to convert a context-dependent question into a self-contained question”; [Abstract] Wu further states: “we develop a query rewriting model CONQRR that rewrites a conversational question in the context into a standalone question.”; [3.] Wu defines the task inputs as “a dialogue context x consisting of a sequence of previous utterances … and the current user question,” and states that the task “aims to rewrite x into a query q such that R can take q such that R can take q as the input query to retrieve passages”; [3.1] Wu also teaches that T5 is fine-tuned “to rewrite a conversational query with the input as the concatenation of utterances in the dialogue context x,” producing “a self-contained version of the query”).
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to modify the teachings of Boiko with teachings of Wu by modifying the method of parallel hybrid search for large-scale code repositories as taught by Boiko to include using contextual information to re-write the received natural language query as taught by Wu for the benefit of understanding user questions within the dialogue context (Wu [Introduction]).
Claim(s) 6 and 16 is/are rejected under 35 U.S.C. 103 as being unpatentable over Boiko (“Method of Parallel Hybrid Search for Large-Scale Code Repositories”; Radio Electronics, Computer Science, Control; Sept. 22, 2025) and further in view of Jain et al. (US 2024/0256582).
Claim 6,
Boiko teaches all the limitations in claim 4. The difference between the prior art and the claimed invention is that Boiko does not explicitly teach wherein the performing a search of a vector database having information regarding one or more of enterprise context, natural language summaries of the enterprise context, embedding vectors for the enterprise context, and embedding vectors for natural language summaries of the enterprise context according to the request to generate a search result comprises performing a lexical search on the code and related documents of the enterprise context and natural language summaries of the enterprise context.
Jain teaches wherein the performing a search of a vector database having information regarding one or more of enterprise context, natural language summaries of the enterprise context, embedding vectors for the enterprise context, and embedding vectors for natural language summaries of the enterprise context according to the request to generate a search result comprises performing a lexical search on the code and related documents of the enterprise context and natural language summaries of the enterprise context ([0002] Jain states: “Individuals associated with an organization (e.g., a company or business entity) may have restricted access to electronic documents and data that are stored across various repositories and data stores, such as enterprise databases and cloud-based data storage services”; Jain teaches summaries and lexical-search fields: “The search index may include portions of text, text summaries, unique words, terms, and term frequency information per indexed document”; [0048] Jain teaches enterprise code and related documentation: “the particular set of data may comprise source code and technical documentation for a particular product”; [0026] Jain teaches summaries and lexical-search fields: “The search index may include portions of text, text summaries, unique words, terms, and term frequency information per indexed document”; [0058] Jain teaches related/referenced documents: “Text associated with other documents linked to or referenced by a searchable document … may also be stored within the search index 204 to provide context for the searchable content”; [0059] Jain teaches lexical matching: “literal matches or lexical matches of search query terms within the body of a message or document may correspond with a first weighting”).
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to modify the teachings of Boiko with teachings of Jain by modifying the method of parallel hybrid search for large-scale code repositories as taught by Boiko to include wherein the performing a search of a vector database having information regarding one or more of enterprise context, natural language summaries of the enterprise context, embedding vectors for the enterprise context, and embedding vectors for natural language summaries of the enterprise context according to the request to generate a search result comprises performing a lexical search on the code and related documents of the enterprise context and natural language summaries of the enterprise context as taught by Jain for the benefit of reduced energy consumption and cost of computing resources, reduced search system downtime, increased quality of search results, increased reliability of information provided to search users, and improved search system performance (Jain [0004]).
Claim 16,
Boiko teaches all the limitations in claim 4. The difference between the prior art and the claimed invention is that Boiko does not explicitly teach wherein the enterprise context at least includes code and documents related to the code.
Jain teaches wherein the enterprise context at least includes code and documents related to the code ([0002] Jain teaches an enterprise/organization setting: “Individuals associated with an organization (e.g., a company or business entity) may have restricted access to electronic documents and data that are stored across various repositories and data stores”; [0032] Jain further teaches that enterprise data sources include “distributed version control and source code management tools,” and electronic files may include “source code files” and “electronic documents”; [0048] Jain teaches “the particular set of data may comprise source code and technical documentation for a particular product”; [0058] Jain also teaches that the search index stores “text, title strings, chat message bodies, metadata, and access rights,” and that “document summaries” and text associated with linked or reference documents may be stored to provide context).
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to modify the teachings of Boiko with teachings of Jain by modifying the method of parallel hybrid search for large-scale code repositories as taught by Boiko to include wherein the enterprise context at least includes code and documents related to the code as taught by Jain for the benefit of reduced energy consumption and cost of computing resources, reduced search system downtime, increased quality of search results, increased reliability of information provided to search users, and improved search system performance (Jain [0004]).
Claim(s) 7 is/are rejected under 35 U.S.C. 103 as being unpatentable over Boiko (“Method of Parallel Hybrid Search for Large-Scale Code Repositories”; Radio Electronics, Computer Science, Control; Sept. 22, 2025) and further in view of Cherednichenko et al. (“AI Agent for conversational Q&A over SaaS codebase using large language models”; May 15-16, 2025; IWS, thereinafter, Chered).
Claim 7,
Boiko teaches all the limitations in claim 4. The difference between the prior art and the claimed invention is that Boiko does not explicitly teach wherein the performing a search of a vector database having information regarding one or more of enterprise context, natural language summaries of the enterprise context, embedding vectors for the enterprise context, and embedding vectors for natural language summaries of the enterprise context according to the request to generate a search result comprises performing a semantic search on the embedding vectors for the enterprise context and embedding vectors for natural language summaries of the enterprise context.
Chered teaches wherein the performing a search of a vector database having information regarding one or more of enterprise context, natural language summaries of the enterprise context, embedding vectors for the enterprise context, and embedding vectors for natural language summaries of the enterprise context according to the request to generate a search result comprises performing a semantic search on the embedding vectors for the enterprise context and embedding vectors for natural language summaries of the enterprise context ([3.] Chered teaches that OpenSearch was selected “to efficiently manage and search the vector embeddings representing code snippets” and as a “vector storage and search engine” for “semantic code search and retrieval in large SaaS codebases”; Chered further states that OpenSearch “allows us to efficiently index and query the vector embeddings generated by text-embedding-3-large, enabling sematic search capabilities that go beyond simple keyword matching,” and supports cosine similarity to find “code snippets that are semantically similar to user queries.”; For both raw context and summaries, Chered states: “To generate meaningful vector representations of our text data – including both raw code and the summaries produced by Claude 3.5 Haiku – we utilized the text-embedding-3-large model”; [4.] Chered also states that the system uses text-embedding-3-large “to create vector embeddings” and that theses embeddings are stored in OpenSearch and “can be efficiently queried, using vector search”).
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to modify the teachings of Boiko with teachings of Chered by modifying the method of parallel hybrid search for large-scale code repositories as taught by Boiko to include wherein the performing a search of a vector database having information regarding one or more of enterprise context, natural language summaries of the enterprise context, embedding vectors for the enterprise context, and embedding vectors for natural language summaries of the enterprise context according to the request to generate a search result comprises performing a semantic search on the embedding vectors for the enterprise context and embedding vectors for natural language summaries of the enterprise context as taught by Chered for the benefit of using LLMs to streamline information retrieval, accelerate onboarding, and enhance overall productivity (Chered [Abstract]).
Claim(s) 9 is/are rejected under 35 U.S.C. 103 as being unpatentable over Boiko (“Method of Parallel Hybrid Search for Large-Scale Code Repositories”; Radio Electronics, Computer Science, Control; Sept. 22, 2025) and further in view Arora et al. (US 2025/0244960).
Claim 9,
Boiko teaches all the limitations in claim 4. The difference between the prior art and the claimed invention is that Boiko does not explicitly teach wherein the NLQ is a request to generate code using contextual information.
Arora teaches wherein the NLQ is a request to generate code using contextual information ([0041] Arora teaches that the code editor interface includes a generative-model interface that enables “queries such as prompts” to be provided to generative models, and that the model output, “such as executable code generated in response to a prompt,” can be provided in the code-cell interface; [0042] Arora further teaches that the prompt editor allows users to submit “users queries such as prompts,” and that “the generative model(s) can generate code cells and populate the code cells with executable code generated in response to user queries”; [0048-0049] For contextual information, Arora teaches that the prompt editor includes controls enabling a user “to select a file, file location, or other input to be provided as part of a prompt,” and gives an example where the user asks, “Create a pipeline that will create an animal classifier based on this data”; Arora further states that, after the prompt, “the generative model writes executable code”).
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to modify the teachings of Boiko with teachings of Arora by modifying the method of parallel hybrid search for large-scale code repositories as taught by Boiko to include wherein the NLQ is a request to generate code using contextual information as taught by Arora for the benefit of, for a large-language model to generate an output that includes executable code in a text file format (Arora [0002]).
Claim(s) 10 and 12 is/are rejected under 35 U.S.C. 103 as being unpatentable over Boiko (“Method of Parallel Hybrid Search for Large-Scale Code Repositories”; Radio Electronics, Computer Science, Control; Sept. 22, 2025) and further in view of Nam et al. (“Using an LLM to Help With Code Understanding”; ICES 2024).
Claim 10,
Boiko teaches all the limitations in claim 4. The difference between the prior art and the claimed invention is that Boiko does not explicitly teach wherein the NLQ is a request to explain a segment of code using contextual information.
Nam teaches wherein the NLQ is a request to explain a segment of code using contextual information ([Abstract] Nam states that their IDE plugin queries GPT-3.5-turbo “with four high-level requests without the user having to write explicit prompts: to explain a highlighted section of code, provide details of API calls used in the code, explain key domain-specific terms, and provide usage examples for an API”; [3.] Nam further states that “GILT is a plugin for the VS code IDE … that considers user context (the code selected by the user) when querying an LLM”; [3.1] Name also states that user can select code and trigger “AI Explanation” to “receive a summary description of the highlighted code,” and that users may “ask a specific question directly to the LLM via the input box”; For contextual information, Name states: “If no code is selected, the entire source code is used as context … alternatively, the relevant code highlighted by the user is used” and “the model will then answer the question with that code as context”; [3.2] Name further states that for direct prompts, “GILT will automatically add any selected code as context for the query” and “request the LLM to respond based on the code context”).
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to modify the teachings of Boiko with teachings of Nam by modifying the method of parallel hybrid search for large-scale code repositories as taught by Boiko to include wherein the NLQ is a request to explain a segment of code using contextual information as taught by Nam for the benefit of completing tasks more than a web search (Nam [Abstract]).
Claim 12,
Boiko teaches all the limitations in claim 4. The difference between the prior art and the claimed invention is that Boiko does not explicitly teach wherein the NLQ is received via an integrated development environment (IDE).
Nam teaches wherein the NLQ is received via an integrated development environment (IDE) ([Abstract] Nam states: “we provide a first investigation of an LLM-based conversational UI built directly in the IDE that is geared towards code understanding”; [3.] Nam further states: “GILT is a plugin for the VS Code IDE”; [3.1] Nam also states: “Users can also ask a specific question directly to the LLM via the input box”).
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to modify the teachings of Boiko with teachings of Nam by modifying the method of parallel hybrid search for large-scale code repositories as taught by Boiko to include wherein the NLQ is received via an integrated development environment (IDE) as taught by Nam for the benefit of completing tasks more than a web search (Nam [Abstract]).
Claim(s) 11 is/are rejected under 35 U.S.C. 103 as being unpatentable over Boiko (“Method of Parallel Hybrid Search for Large-Scale Code Repositories”; Radio Electronics, Computer Science, Control; Sept. 22, 2025) and further in view Carrara et al. (US 2025/0076836).
Claim 11,
Boiko teaches all the limitations in claim 4. The difference between the prior art and the claimed invention is that Boiko does not explicitly teach wherein the NLQ is a request to perform code debug.
Carrara teaches wherein the NLQ is a request to perform code debug ([0135-0136] Carrara states that “FIG. 17 illustrates an example methodology 1700 for using generative AI to troubleshoot or debug software components of an industrial control system project (e.g., control programming, HMI applications, etc.)”; Carrara further states: “At 1704, a plain language instruction to troubleshoot industrial control code is received via a chat interface of an industrial IDE system”; Carrara then states that, in response, “ the control code is analyzed using the generative AI model … to identify presence of control programming errors,” deviations, security-protocol deviations, inefficiencies, or deviations from functional specifications; Carrara further states that, if an issue is discovered, the IDE “at least one of corrects the discovered issues or renders a plain language summary of the discovered issue and recommendations for modifying the control code to correct the issue.”).
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to modify the teachings of Boiko with teachings of Carrara by modifying the method of parallel hybrid search for large-scale code repositories as taught by Boiko to include wherein the NLQ is a request to perform code debug as taught by Carrara for the benefit of improving one or more performance metrics of the automation system to be controlled by the code (e.g., speed, energy consumption, machine uptime, product throughput, cycle time, etc.), or otherwise cause a performance metric to satisfy a criterion indicative of improved performance (Carrara [0100]).
Claim(s) 13 and 14 is/are rejected under 35 U.S.C. 103 as being unpatentable over Boiko (“Method of Parallel Hybrid Search for Large-Scale Code Repositories”; Radio Electronics, Computer Science, Control; Sept. 22, 2025) and further in view Madan et al. (US 12,222,898).
Claim 13,
Boiko further teaches the computer-implemented method of claim 4, further comprising: ingesting the code repository ([3.] Boiko states that the indexing phrase transforms “raw source code files” into “structured, querable data”; Boiko also states: “Each file in the codebase is processed independently. The content of the file is read and passed through a natural language models”);
chunking the ingested code repository into function and/or class-level chunks ([pg. 52] [NOMENCLATURE] Boiko’s nomenclature defines B as “a set of source code blocks,” and defines a “chunking function of a particular code block.”; [3.] Boiko further teaches a chunking function and states that “each chunk is summarized individually”; For class/method units, Boiko states that the system stores metadata including “the name and type of the code block (e.g. class, method)”);
generating natural language summaries of the chunks using a first machine learning model ([3.] Boiko states that “each chunk is summarized individually using LLM”; Boiko also states that summarization is performed using “the LLM of the GPT-4o model,” and that a “concise, high-level summary is returned”);
generating embedding vectors for the chunks using a second machine learning model ([3.] Boiko states: “Following the natural language summary generation, the system computes a semantic embedding of the summary using a pre-trained embedding model”; Boiko identifies the selected embedding model as “OpenAI’s text-embedding-ada=002”; Boiko also states that Qdrant stores dense embedding vectors “derived from code summaries”);
generating embedding vectors for the chunks using the second using a machine learning model ([pg. 52] Boiko’s nomenclature identifies M as “an embedding model text-embedding-ada-002,” and E as “a set of summary embeddings.”; [3.] Boiko further states that the same embedding model is used during indexing and query embedding generation (see pg. 57));
indexing the chunks with their associated generated embedding vectors into a vector database ([3.] Boiko states: “The system employs Qdrant as the underlying vector database to support high-performance semantic code search”; Boiko further states that Qdrant is suitable for “storing and querying dense embedding vectors derived from code summaries,” and that the system stores “the normalized embeddings of summarized code chunks with payload metadata).
The difference between the prior art and the claimed invention is that Boiko does not explicitly teach further ingesting related documents; indexing the chunks with their associated summaries into the vector database.
Madan teaches ingesting related documents ([col. 4 lines 13-29] Madan states that a user may specify or upload “a set of data objects” or “a set of files” for processing, including files from cloud storage or files uploaded from a drive or local area network.; Madan further states that the set of files may include “a group of engineering documents”);
indexing the chunks with their associated summaries into the vector database ([col. 7 lines 17-28] Madan states: “data chunk summarization system 60 creates textual summaries 62 of each data chunk 48”; Madan further states “Each textual summary 62 is indexed and linked to a node of a corresponding data chunk”; Madan then states: “Data chunks 48 and summaries 62 may be stored in vector database 32 using object-based structures”).
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to modify the teachings of Boiko with teachings of Madan by modifying the method of parallel hybrid search for large-scale code repositories as taught by Boiko to include ingesting related documents; indexing the chunks with their associated summaries into the vector database as taught by Madan for the benefit of providing an artificial intelligence (AI) platform that processes a specified set of data objects and provides an AI querying service to perform queries associated with the specified set of data objects (Madan [Summary]).
Claim 14,
Madan further teaches the computer-implemented method of claim 13, wherein the vector database is maintained by a search and database service ([col. 3 lines 33-52] Madan states that the Ai platform includes “an AI processing and query server (“server”) 12 that can be accessed by one or more clients 18 via a gateway service 16”; Madan further states that “Server 12 includes an App service 22” that allows upload/specification of data objects and “provides an interface for querying information” using a natural-language interface, and that “processing of objects and performing queries of those resources is managed by the App service 22,” supported by “AI model service 24”; [col. 3 lines 53-67] Mandan also states that “Server 12 also includes … an index database such as mongo database 30 … and a vector database 32,” and operates with the mongo database 30 to index and store vector embeddings for fast retrieval and similarity search”).
Claim(s) 15 is/are rejected under 35 U.S.C. 103 as being unpatentable over Boiko (“Method of Parallel Hybrid Search for Large-Scale Code Repositories”; Radio Electronics, Computer Science, Control; Sept. 22, 2025) and further in view Penta et al. (US 12,373,506) .
Claim 15,
Boiko teaches all the limitations in claim 4. The difference between the prior art and the claimed invention is that Boiko does not explicitly teach wherein predicting a response for the NLQ using the search result augmented NLQ is to be performed using a model fine-tuned on aspects of the enterprise context.
Penta teaches wherein predicting a response for the NLQ using the search result augmented NLQ is to be performed using a model fine-tuned on aspects of the enterprise context ([col. 3 line 6 to col. 4 line 14] Penta states that its personalized RAG system “trains and utilizes” a RAG model “personalized on a per-entity basis by fine-tuning large language models to specific data contexts relating to user accounts, data stored for user accounts, and/or specific software applications.”; Penta further states that the entity may be “an enterprise, or an organization”; Penta teaches generating the response by providing the query and data context to the LLM and causing the LLM to generate a response “specific to the data context”; Penta also states that the system can “fine-tune the query, the embedding model, the vector database, the data contexts, and/or the large language model” so the system generates a response considering “the context and environment of the entity”).
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to modify the teachings of Boiko with teachings of Penta by modifying the method of parallel hybrid search for large-scale code repositories as taught by Boiko to include wherein predicting a response for the NLQ using the search result augmented NLQ is to be performed using a model fine-tuned on aspects of the enterprise context as taught by Penta for the benefit of providing the personalized response for display on a client device associated with the entity (Penta [Abstract]).
Claim(s) 1 and 3 is/are rejected under 35 U.S.C. 103 as being unpatentable over Boiko (“Method of Parallel Hybrid Search for Large-Scale Code Repositories”; Radio Electronics, Computer Science, Control; Sept. 22, 2025) in view of Nam et al. (“Using an LLM to Help With Code Understanding”; ICES 2024).
Claim 1,
Boiko teaches a computer-implemented method comprising ([4.] Boiko states: “The system was implemented in a .NET-based environment, incorporating components for indexing, vector storage, text search, and OpenAI API integration for both summarization and post-retrieval code refinement”; [3.] Boiko also states that “the method consists of 3 phrases: indexing, retrieval, and post-processing”):
generating an index of a codebase by at least ([3.] Boiko states: “The indexing phase servers as the preparatory stage in the hybrid code search system, where raw source code files are transformed into structured, queryable data suitable for both keyword based and embedding based retrieval”; Boiko further states: “Within the context of this work, indexing involves parsing source code files, extracting metadata, generating natural language summaries, and creating vector representations of these summaries for storage and subsequent retrieval”)
(i) chunking source code of the codebase to extract function-level and class-level code chunks ([1.] Boiko states: “Suppose we have a set of source code chunks B = {b1, b2, …, bn} from a software codebase”; Boiko further states: “Each chunk bi ɛ B contains one or more code blocks”; [3.] [Fig. 2] For class/method-level extraction, Boiko states that the stored metadata include “the name and type of the code block (e.g. class, method), and the corresponding line range within the file”),
(ii) generating respective natural language summaries of the code chunks ([3.] [Eq. 4] Boiko states: “Each chunk is summarized individually using LLM”; [3.] [Fig. 1] Boiko also states that the prompt should cause the model to “return summaries of each meaningful block of code in the file and represent it in a JSON format”),
(iii) creating respective semantic vector embeddings of the code chunks and the natural language summaries ([3.] Boiko states: “Following the natural language summary generation, the system computes a semantic embedding of the summary using a pre-trained embedding model”; Boiko further states: “This embedding is a high-dimensional vector that captures the semantic meaning of the text and enables efficient similarity comparisons with query embeddings during retrieval”; Boiko also states that the system stores “the normalized embeddings of summarized code chunks”), and
(iv) storing the code chunks and the respective natural language summaries and respective semantic vector embeddings in a search index ([3.] Boiko states: “Each vector is associated with the corresponding file and stored in a vector search database to facilitate cosine similarity queries during semantic retrieval”; [3.] [Fig. 2] Boiko further states: “The system stores not only the normalized embeddings of summarized code chunks, but also custom metadata (payload) including the full file path, the name and type of code block (e.g. class, method), and the corresponding line range within the file”; Boiko also states: “Summaries of code blocks, as well as additional metadata, are indexed using Elasticsearch”; Boiko states: “After this phase, the indexed codebase contains structured, searchable entries for each file, consisting of its metadata, summaries and their dense vectors, and keyword-indexed content”);
generating one or more embedding vectors for the NLQ ([3.] [Eq. 11] Boiko states: “In parallel, the system conducts semantic retrieval by encoding the user query into a high dimensional embedding using the same embedding model employed during indexing”; Boiko further states: “The resulting query embedding is then compared against the stored file embeddings in a vector database using cosine similarity as the distance metric”);
performing multi-strategy retrieval on the index of the codebase using the one or more embedding vectors for the NLQ, wherein the multi-strategy retrieval includes performing lexical search on the code chunks and natural language summaries and semantic search on the vector embeddings, and combining results from the lexical and semantic searches using hybrid scoring ([3.] Boiko states: “The retrieval phase constitutes the core of the hybrid search mechanism and is responsible for identifying relevant source code blocks based on user-supplied natural language queries”; Boiko further states: “This phase integrates two parallel retrieval processes: traditional keyword-based search and sematic vector-based search”; for lexical search, Boiko states: “Upon receiving a natural language query, the system performs keyword-based retrieval by submitting the query to a full-text search engine, such as Elasticsearch”, and that the engine operates over indexed textual content, “particularly focusing on the code block summaries and the generated sematic tags”; [3.] [eq. 11] For semantic search, Boiko states that the query is encoded into an embedding and compared against stored embeddings in a vector database; for hybrid scoring, Boiko states: “The results from both retrieval paths are processed independently and can be returned as separate ranked list or integrated into a unified ranking,” and “these results can be merged, with ranking adjusted based on predefined weighting or scoring strategies”; [eq. 12] Boiko’s formula combines BM25 and cosine similarity: see eq. 12);
retrieving particular ones of the code chunks and their respective natural language summaries based on the hybrid scoring ([3.] [eq. 12] Boiko states that the hybrid score selects top results using eq. 12 (“Y = TopK….”); Boiko further states: “to optimize performance and reduce inference time, only the most relevant results (typically the top five according to the hybrid scoring function) are selected for automatic refinement”; Boiko also states: “Each retrieved file entry includes associated metadata such as the file name, path, summary, and optionally the top-matching tags or score explanations from both retrieval methods”);
generating a response to the NLQ using a prompt to a language model that has been augmented with the retrieved code chunks and their respective natural language summaries ([3.] Boiko states: “each retrieved code block is further analyzed using an LLM, which operates on the full content of it, its summary, and the original user query”; Boiko further states: “The model receives as input the user’s natural language query, the code block summary that was previously generated and indexed, and the source code in line range.”; Boiko also states that the resulting set contains “context-aware, and human-readable answers that they system returns to the user”); and
The difference between the prior art and the claimed invention is that Boiko does not explicitly teach receiving, from a user, a natural language query (NLQ) about the codebase in a chat interface an integrated development environment; providing the response to the user via the chat interface.
Nam teaches receiving, from a user, a natural language query (NLQ) about the codebase in a chat interface an integrated development environment ([Abstract] Name states: “we provide a first investigation of an LLM-based conversational UI built directly in the IDE that is geared towards code understanding”; [3.] Nam further states: “GILT is a plugin for the VS Code IDE”; [3.1] [Fig. 1] Nam also states: “Users can also ask a specific question directly to the LLM via the input box”);
providing the response to the user via the chat interface ([3.1] Nam states: “The method will then answer the question with that code as context”; Nam further states: “GILT also allows users to probe the LLM by supporting conversational interaction”; [Abstract] Nam also describes the interface as “an LLM-based conversational UI built directly in the IDE”).
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to modify the teachings of Boiko with teachings of Nam by modifying the method of parallel hybrid search for large-scale code repositories as taught by Boiko to include receiving, from a user, a natural language query (NLQ) about the codebase in a chat interface an integrated development environment; providing the response to the user via the chat interface as taught by Nam for the benefit of completing tasks more than a web search (Nam [Abstract]).
Claim 3,
Nam further teaches the computer-implemented method of claim 1, wherein the NLQ is a request to perform one or more of providing explanation of a segment of code using contextual information, performing code refactoring, performing knowledge gathering, performing code analysis, generating code using contextual information, and/or performing code debugging ([Abstract] Nam states that its IDE plugin supports requests “to explain a highlighted section of code, provide details of API calls used in the code, explain key domain-specific terms, and provide usage examples for an API”; [3.1] Name also states that users may select code and trigger the tool “to receive a summary description of the highlighted code,” and then use buttons for “API,” “domain-specific concepts,” and “usage examples,” which provide “more detailed explanations with preset prompts”; [3.1] Nam further teaches contextual information: “If no code is selected, the entire source code is used as context; alternatively, the relevant code highlighted by the user is used,” and “the mode will then answer the question with that code as context”; [3.2] Nam additionally states that “GILT will automatically add any selected code as context for the query” and “requests the LLM to respond based on the code context”).
Claim(s) 2 is/are rejected under 35 U.S.C. 103 as being unpatentable over Boiko (“Method of Parallel Hybrid Search for Large-Scale Code Repositories”; Radio Electronics, Computer Science, Control; Sept. 22, 2025) in view of Nam et al. (“Using an LLM to Help With Code Understanding”; ICES 2024) and further in view of Wu et al. (“CONQRR: Conversational Query Rewriting for Retrieval with Reinforcement Learning”; Dec. 2022; Association for Computational Linguistics).
Claim 2,
Boiko and Nam teach all the limitations in claim 1. The difference between the prior art and the claimed invention is that Boiko nor Nam explicitly teach using contextual information to re-write the received natural language query.
Wu teaches using contextual information to re-write the received natural language query ([Introduction] Wu states: that “query rewriting (QR) in a conversation … is to convert a context-dependent question into a self-contained question”; [Abstract] Wu further states: “we develop a query rewriting model CONQRR that rewrites a conversational question in the context into a standalone question.”; [3.] Wu defines the task inputs as “a dialogue context x consisting of a sequence of previous utterances … and the current user question,” and states that the task “aims to rewrite x into a query q such that R can take q such that R can take q as the input query to retrieve passages”; [3.1] Wu also teaches that T5 is fine-tuned “to rewrite a conversational query with the input as the concatenation of utterances in the dialogue context x,” producing “a self-contained version of the query”).
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to modify the teachings of Boiko with teachings of Wu by modifying the method of parallel hybrid search for large-scale code repositories as taught by Boiko to include using contextual information to re-write the received natural language query as taught by Wu for the benefit of understanding user questions within the dialogue context (Wu [Introduction]).
Claim(s) 17 is/are rejected under 35 U.S.C. 103 as being unpatentable over Boiko (“Method of Parallel Hybrid Search for Large-Scale Code Repositories”; Radio Electronics, Computer Science, Control; Sept. 22, 2025) in view of Lan et al. (US 12,265,528).
Claim 17,
Boiko teaches receive a request to answer a natural language query (NLQ) ([pg. 52] [NOMENCLATURE] Boiko defines q as “a user-supplied natural language query”; [3.] Boiko further states that the retrieval phase identifies relevant source-code blocks “based on user-supplied natural language queries” and that, “upon receiving a natural language query,” the system performs retrieval);
generate one or more embedding vectors for the NLQ ([3.] Boiko sates: “In parallel, the system conducts semantic retrieval by encoding the user query into a high-dimensional embedding using the same embedding model employed during indexing”; Boiko further states that the resulting query embedding is compared against stored embeddings in a vector database);
perform a search of a vector database having information regarding one or more of enterprise context, natural language summaries of enterprise context, embedding vectors for the enterprise context, and embedding vectors for natural language summaries of the enterprise context according to the request to generate a search result ([3.] Boiko states: “The system employs Qdrant as the underlying vector database to support high-performance semantic code search”; Boiko further states that Qdrant is suitable for “storing and querying dense embedding vectors derived from code summaries”; Boiko also states that Qdrant stores “the normalized embeddings of summarized code chucks” and payload metadata including file path, code-block name/type, and line range);
augment the NLQ using the search result ([3.] Boiko states: “each retrieved code block is further analyzed using an LLM, which operates on the full content of it, its summary, and the original user query”; Boiko further states that the model receives “the user’s natural language query,” the “code block summary”, and “the source code in a line range”);
predict a response for the NLQ using the search result augmented NLQ ([3.] Boiko states that the model is prompted with “both the query and block-level context,” and that the LLM-refined result contains “the final, context-aware, and human-readable answers that the system returns to the user”); and
provide the predicted response ([3.] Boiko states that the resulting set contains “human-readable answers that the system returns to the user”; [4.] Boiko also states that retrieved results were “displayed to the user” with cosine similarity scores and matched summaries).
The difference between the prior art and the claimed invention is that Boiko does not explicitly teach a system comprising: a first one or more computing devices to implement a storage service in a multi-tenant provider network; and a second one or more computing devices to implement a natural language query service in the multi-tenant provider network, the natural language query service including instructions that upon execution cause the natural language query service.
Lan teaches a system comprising ([claim 15] Lan states in claim 15: “a first one or more electronic devices to implement a database storage service in a multi-tenant provider network; and a second one or more electronic devices to implement a natural language query processing service in the multi-tenant provider network”):
a first one or more computing devices to implement a storage service in a multi-tenant provider network ([claim 15] Lan claim 15 states that “a first one or more electronic devices to implement a database storage service in a multi-tenant provider network”; [col. 3 line 26 to col. 4 line 3] Lan also teaches that datasets are “stored by a storage service 160 of a provider network” and that the provider network includes “data/storage resources,” including “object storage, block-level storage, data archival storage, databases and database tables”); and
a second one or more computing devices to implement a natural language query service in the multi-tenant provider network ([claim 15] Lan’s claim 15 states “a second one or more electronic device to implement a natural language query processing service in the multi-tenant provider network”; [col. 5 lines 7-18] Lan also states that the “NLQ processing service 110,” answers questions and has a “a topic modeling phase and a question answering (interface) phase”),
the natural language query service including instructions that upon execution cause the natural language query service ([claim 15] Lan’s claim 15 states that the natural language query processing service includes “instructions that upon execution cause the natural language query processing service to” perform recited operations; [col. 30 lines 40-52] Lan also states that “program instructions and data” are stored as “NLQ processing service code 1525,” executable to implement the NLQ processing service).
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to modify the teachings of Boiko with teachings of Lan by modifying the method of parallel hybrid search for large-scale code repositories as taught by Boiko to include a system comprising: a first one or more computing devices to implement a storage service in a multi-tenant provider network; and a second one or more computing devices to implement a natural language query service in the multi-tenant provider network, the natural language query service including instructions that upon execution cause the natural language query service as taught by Lan for the benefit of making structured query language (SQL) call, where to make such a call, how to interpret the results, etc. (Lan [col. 2 lines 1-14]).
Claim(s) 18 is/are rejected under 35 U.S.C. 103 as being unpatentable over Boiko (“Method of Parallel Hybrid Search for Large-Scale Code Repositories”; Radio Electronics, Computer Science, Control; Sept. 22, 2025) in view of Lan et al. (US 12,265,528) and further in view of AlOmar et al. (“How to Refactor this code? An Exploratory Study on Developer-ChatGPT Refactoring Conversations”; April 2024; ACM).
Claim 18,
Boiko and Lan teaches all the limitations in claim 17. The difference between the prior art and the claimed invention is that Boiko does not explicitly teach wherein the NLQ is a request to perform code refactoring.
AlOmar wherein the NLQ is a request to perform code refactoring ([Abstract] AlOmar states that their study concerns “conversations between developers and ChatGPT related to refactoring” and “developers’ explicit refactoring intention.”; [Introduction] AlOmar further states that developers “request ChatGPT to refactor their code” and “request the refactoring of a given input code.”; [3.3] AlOmar also reports that developer prompts include code fragments “that need to be refactored” along with textual descriptions/instruction, and that many developers provide “code fragments that require refactoring” plus “a textual description of how they want them to be addressed”; [4.] Finally, AlOmar states: “the prompt asks the model to refactor the code to eliminate duplication of an input source code”).
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to modify the teachings of Boiko with teachings of AlOmar by modifying the method of parallel hybrid search for large-scale code repositories as taught by Boiko to include wherein the NLQ is a request to perform code refactoring as taught by AlOmar for the benefit of understanding of the collaboration between developers and AI models (AlOmar [Abstract]).
Claim(s) 19 is/are rejected under 35 U.S.C. 103 as being unpatentable over Boiko (“Method of Parallel Hybrid Search for Large-Scale Code Repositories”; Radio Electronics, Computer Science, Control; Sept. 22, 2025) in view of Lan et al. (US 12,265,528) and further in view of Nam et al. (“Using an LLM to Help With Code Understanding”; ICES 2024).
Claim 19,
Boiko and Lan teach all the limitations in claim 17. The difference between the prior art and the claimed invention is that Boiko does not explicitly teach wherein the NLQ is a request to perform knowledge gathering.
Nam teaches wherein the NLQ is a request to perform knowledge gathering ([Introduction] Nam states that developers need information about “the intricacies of the APIs used, and the relevant domain-specific concepts”; [Abstract] Nam further states that their IDE plugin supports high-level requests “to explain a highlighted section of code, provide details of API calls used in the code, explain key domain-specific terms, and provide usage examples for an API”; [3.1] Nam also teaches buttons for “API,” “domain-specific concepts,” and “usage examples,” where the API button give “detailed explanations about the API calls used in the code,” the Concept button provides “domain-specific concepts”, and the Usage button provides “a code example involving API calls”).
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to modify the teachings of Boiko with teachings of Nam by modifying the method of parallel hybrid search for large-scale code repositories as taught by Boiko to include wherein the NLQ is a request to perform knowledge gathering as taught by Nam for the benefit of completing tasks more than a web search (Nam [Abstract]).
Claim(s) 20 is/are rejected under 35 U.S.C. 103 as being unpatentable over Boiko (“Method of Parallel Hybrid Search for Large-Scale Code Repositories”; Radio Electronics, Computer Science, Control; Sept. 22, 2025) in view of Lan et al. (US 12,265,528) and further in view of Acharya et al. (US 2025/0045148).
Claim 20,
Boiko and Lan teach all the limitations in claim 17. The difference between the prior art and the claimed invention is that Boiko does not explicitly teach wherein the NLQ is a request to perform code analysis.
Acharya teaches herein the NLQ is a request to perform code analysis ([0014] Acharya states that the generative AI system receives “a user request to evaluate software code,” including software code not associated with a detected issue; [0074] Acharya further sates that a prompt may indicate a request to analyze “the following software code in light of the context following the software code,” and that the generative AI provides the prompt, code lines, and context to the language model; [0075] Acharya also states that the language model outputs a response for the user request after processing “the received context” and “the lines of software code”).
Therefore, it would have been obvious to one of ordinary skill in the art, before the effective filing date of the claimed invention, to modify the teachings of Boiko with teachings of Acharya by modifying the method of parallel hybrid search for large-scale code repositories as taught by Boiko to include herein the NLQ is a request to perform code analysis as taught by Acharya for the benefit of providing the output to the requestor and/or performs one or more actions to further evaluate or repair the detected issue (Acharya [0004]).
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Duan et al. (US 2023/0359441) – A retrieval-augmented code completion system uses the context of a partially-formed source code snippet of a source code program and a hint to predict the source code tokens needed to complete the partially-formed source code snippet. The hint is a source code segment that completes a semantically-similar source code segment of the partially-formed source code snippet. The hint is found in a retrieval source code database using a hybrid retrieval technique. A deep learning decoder model uses the context of the partially-formed source code snippet and the hint to predict the most likely candidate sequence of source code tokens to complete the partially-formed source code snippet.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to SHREYANS A PATEL whose telephone number is (571)270-0689. The examiner can normally be reached Monday-Friday 8am-5pm PST.
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, Pierre Desir can be reached at 571-272-7799. 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.
SHREYANS A. PATEL
Primary Examiner
Art Unit 2653
/SHREYANS A PATEL/ Examiner, Art Unit 2659
Prosecution Insights