DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Response to Amendment
This action is in response to applicant’s arguments and amendments filed 5/22/2026, which are in response to USPTO Office Action mailed 2/25/2026. Applicant’s arguments have been considered with the results that follow: THIS ACTION IS 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) 1-2, 6-11 and 13-20 is/are rejected under 35 U.S.C. 103 as being unpatentable over KUANG et al. (WIPO Invention Application Publication No.: WO 2024/067593 A1; Published On: Apr. 4, 2024) in view of SONG (US PGPUB No. 2023/0297374; Pub. Date: Sep. 21, 2023) and Deep et al. (US PGPUB No. 2024/0320951; Pub. Date: Sep. 26, 2024).
Regarding independent claim 1,
KUANG discloses a system comprising: a non-transitory memory; See Pg. 8, Paragraphs 8-10, (Disclosing a vector retrieval method for improving retrieval speed of vector-based data. The system may comprise a device including a processor and memory. The memory is used to store programs or instructions to be executed by the processor, i.e. a system comprising: a non-transitory memory.)
and one or more hardware processors coupled with the non-transitory memory and configured to execute instructions from the non-transitory memory to cause the system to: in response to receiving an input, generate an input vector based on the input; See Pg. 12, Paragraph 10, (Disclosing a vector retrieval method for improving retrieval speed of vector-based data. The system may utilize a feature extraction model wherein a set of sample images may be input into a convolutional neural network (CNN) for feature extraction such that the CNN model outputs a vector corresponding to each sample image to be stored in a vector base library, i.e. in response to receiving an input (e.g. computing device 20 may input a set of sample images into the CNN model), generate an input vector based on the input (e.g. the sample image vectors are generated in response to the inputting);)
query a vector database using the input vector, See Pg. 10, Paragraph 9, (Vectors may be formed and stored in a vector base library to be queried after feature extraction by searching the vector base library, i.e. query a vector database using the input vector.)
wherein the vector database comprises (ii) a vector portion including a plurality of vectors stored in a non-volatile memory of the system, See Pg. 6, Paragraph 6, (The process of vector retrieval comprises constructing a vector base library containing vectors obtained by feature extraction of a large amount of data.) See Pg. 12, Paragraph 10, (Vector base library is located in memory 202 of computing device 20 or in storage device 30. Storage device 30 may be embodied as an independent storage medium or memory, i.e. (ii) a vector portion including a plurality of vectors stored in a non-volatile memory of the system.)
and provide a response to the input based on the one or more vectors. See Pg. 15, Paragraph 2, (Vectors included in at least one retrieval partition may be output as a query result or fed back to the user, i.e. provide a response to the input based on the one or more vectors.)
KUANG does not disclose the step wherein the vector database comprises (i) an index portion stored in a volatile memory of the system,
and wherein querying the vector database comprises: accessing the index portion from the volatile memory of the system;
identifying, from the plurality of second level vector partitions a particular second level vector partition based on the input vector,
SONG discloses the step wherein the vector database comprises (i) an index portion stored in a volatile memory of the system, See FIG. 1 & Paragraph [0057], (Disclosing a method for processing vector data. FIG. 1 illustrates the method comprising step S110 of accessing real-time vector data and writing said vector data into a cache. A vector data retrieval request utilizes a feature classification index established for newly added vector data, i.e. wherein the vector database comprises (i) an index portion stored in a volatile memory of the system (e.g. vector data is maintained in a cache).)
and wherein querying the vector database comprises: accessing the index portion from the volatile memory of the system; See Paragraph [0057], (A vector retrieval request may be executed to facilitate retrieval of vector data wherein a feature classification index is used to obtain a target vector from a related vector data cluster, i.e. wherein querying the vector database comprises: accessing the index portion from the volatile memory of the system;)
identifying, from the plurality of second level vector partitions a particular second level vector partition based on the input vector, See Paragraph [0087], (A master node may receive a vector retrieval request and distributes said request to multiple computing nodes to determine a number of similar representative vector data based on a similarity metric and determines a vector data cluster associated with the similar representative vector data according to the feature classification index. The master node may then obtain a final retrieval result by traversing multiple associated vector data clusters. Note [0085] wherein vector data retrieval is performs based on a target classification index constructed from a first-level index and a second-level index wherein the second-level index is formed through spatio-temporal classification used to perform vector retrieval, i.e. identifying, from the plurality of second level vector partitions a particular second level vector partition based on the input vector (e.g. vector data retrieval comprises searching a multi-level vector index structure including a second-level index).)
KUANG and SONG are analogous art because they are in the same field of endeavor, vector-based search systems. It would have been obvious to anyone having ordinary skill in the art before the effective filing date to modify the system of KUANG to include the method of scoring newly-added vector data into ac ache for use during vector retrieval operations as disclosed by SONG. Paragraph [0087] of SONG discloses that the use of the feature classification index may facilitate spatio-temporal classification of vector data which results in an improvement in the efficiency of vector retrieval and accelerates the process of vector retrieval.
KUANG-SONG does not disclose the step wherein the index portion comprises a plurality of first level vector partitions and a plurality of second level vector partitions,
wherein the particular second level vector partition comprises a subset of the plurality of first level vector partitions;
identifying, from the subset of the plurality of first level vector partitions, one or more first level vector partitions based on the input vector;
and retrieving one or more vectors corresponding to the one or more first level vector partitions from the non-volatile memory of the system;
Deep discloses the step wherein the index portion comprises a plurality of first level vector partitions and a plurality of second level vector partitions, See FIG. 7 & Paragraphs [0072]-[0073], (Disclosing a system for retrieving query target objects from partitions of a spatial index comprising vector data. FIG. 7 illustrates a plurality of partitioning approaches for processing vector objects including partitioning approach 753 of constructing a multi-level structure using hierarchical navigable small world (HNSW) graphs. An HNSW graph is constructed bottom-up for each of the partitions corresponding to one or more leaf nodes of a cover tree, i.e. wherein the index portion comprises a plurality of first level vector partitions and a plurality of second level vector partitions (e.g. an HNSW graph structure comprises a plurality of layers where links are separated across different layers with overlapping members).)
wherein the particular second level vector partition comprises a subset of the plurality of first level vector partitions; See FIG. 7 & Paragraphs [0072]-[0073], (FIG. 7 illustrates partitioning approach 753 of constructing a multi-level structure using hierarchical navigable small world (HNSW) graphs. An HNSW graph is constructed bottom-up for each of the partitions corresponding to one or more leaf nodes of a cover tree, i.e. wherein the particular second level vector partition comprises a subset of the plurality of first level vector partitions (e.g. an HNSW graph structure comprises a plurality of layers where links are separated across different layers with overlapping members);)
identifying, from the subset of the plurality of first level vector partitions, one or more first level vector partitions based on the input vector; See FIG. 7 & Paragraphs [0072]-[0073], (FIG. 7 illustrates partitioning approach 753 of constructing a multi-level structure using hierarchical navigable small world (HNSW) graphs. An HNSW graph is constructed bottom-up for each of the partitions corresponding to one or more leaf nodes of a cover tree.)
The examiner notes that searching an HNSW graph comprises entering a top layer of the HNSW graph and traversing edges across each layer to reach a nearest neighbor to the query vector representing a query, i.e. identifying, from the subset of the plurality of first level vector partitions, one or more first level vector partitions based on the input vector
and retrieving one or more vectors corresponding to the one or more first level vector partitions from the non-volatile memory of the system; See Paragraph [0031], (A query or partition request may be sent to one or more query targets and the system may determine whether the retrieved partition includes an embedding vector that satisfies the similarity criterion with respect to a first embedding vector.) See Paragraph [0024], (If a sufficiently similar vector is not present in the local cache, a partition map created for the spatial index may be consulted to identify a relevant record, i.e. retrieving one or more vectors corresponding to the one or more first level vector partitions (e.g. vectors may be retrieved form a spatial index such as the HNSW graphs which search a layered structure for relevant vectors and may therefore retrieve vectors from a topmost layer) from the non-volatile memory of the system (e.g. the system may consult non-cache memory to determine a relevant vector);)
KUANG, SONG and Deep are analogous art because they are in the same field of endeavor, vector-based search systems. It would have been obvious to anyone having ordinary skill in the art before the effective filing date to modify the system of KUANG-SONG to include the method of retrieving vector data based on a spatial index as disclosed by Deep. Paragraph [0025] of Deep discloses that the system may provide at least the following advantages with regard to object recognition tasks: (a) reducing the average time taken between the capture of an image and the recognition or non-recognition of objects or entities represented in the image, thus enabling response actions to be initiated rapidly regardless of whether objects were recognized or not, (b) reducing the overall amount of networking bandwidth that has to be utilized for object recognition applications, and/or (c) enhancing data security for object recognition applications by reducing the number of times potentially sensitive data is transferred over the public Internet or other types of less secure network pathways.
KUANG, SONG and Deep are analogous art because they are in the same field of endeavor, vector-based search systems. It would have been obvious to anyone having ordinary skill in the art before the effective filing date to modify the system of KUANG-SONG to include the method of retrieving vector data based on a spatial index as disclosed by Deep. Paragraph [0025] of Deep discloses that the system may provide at least the following advantages with regard to object recognition tasks: (a) reducing the average time taken between the capture of an image and the recognition or non-recognition of objects or entities represented in the image, thus enabling response actions to be initiated rapidly regardless of whether objects were recognized or not, (b) reducing the overall amount of networking bandwidth that has to be utilized for object recognition applications, and/or (c) enhancing data security for object recognition applications by reducing the number of times potentially sensitive data is transferred over the public Internet or other types of less secure network pathways.
Regarding dependent claim 2,
As discussed above with claim 1, KUANG-SONG-Deep discloses all of the limitations.
Deep further discloses the step wherein the index portion comprises an index that represents the plurality of first level vector partitions and the plurality of second level vector partitions, and wherein each second level vector partition in the plurality of second level vector partitions corresponds to a different subset of the plurality of first level vector partitions. See FIG. 7 & Paragraphs [0072]-[0073], (FIG. 7 illustrates a plurality of partitioning approaches for processing vector objects including partitioning approach 753 of constructing a multi-level structure using hierarchical navigable small world (HNSW) graphs. An HNSW graph is constructed bottom-up for each of the partitions corresponding to one or more leaf nodes of a cover tree, i.e. wherein the index portion comprises an index that represents the plurality of first level vector partitions (e.g. a topmost level) and the plurality of second level vector partitions, and wherein each second level vector partition in the plurality of second level vector partitions corresponds to a different subset of the plurality of first level vector partitions (e.g. layers of the HNSW graph contain different subsets of a complete graph. The bottom layer contains all nodes while upper layers contain fewer nodes).)
Regarding dependent claim 6,
As discussed above with claim 1, KUANG-SONG-Deep discloses all of the limitations.
KUANG further discloses the step wherein the non-volatile memory is a solid-state drive memory. See Pg. 12, Paragraph 5, (Storage resources may be provided by a hard disk which may be embodied as a solid-state drive (SSD), i.e. wherein the non-volatile memory is a solid-state drive memory.)
Regarding dependent claim 7,
As discussed above with claim 1, KUANG-SONG-Deep discloses all of the limitations.
KUANG further discloses the step wherein the vector database further comprises a non-indexed vector portion stored in the volatile memory, wherein the non-indexed vector portion comprises a second plurality of vectors that is not indexed. See Pg. 6, Paragraph 6, (The process of vector retrieval comprises constructing a vector base library containing vectors obtained by feature extraction of a large amount of data, i.e. wherein the vector database further comprises a non-indexed vector portion stored in the volatile memory, wherein the non-indexed vector portion comprises a second plurality of vectors that is not indexed (e.g. the method of KUANG does not disclose the use of an index, therefore vectors of the vector base library are stored in a vector format and retrieved in a vector format).)
Regarding independent claim 8,
KUANG discloses a method comprising: accessing, by a computer system, an input vector generated by an artificial intelligence (AI) model based on an input; See Pg. 12, Paragraph 10, (Disclosing a vector retrieval method for improving retrieval speed of vector-based data. The system may utilize a feature extraction model wherein a set of sample images may be input into a convolutional neural network (CNN) for feature extraction such that the CNN model outputs a vector corresponding to each sample image to be stored in a vector base library, i.e. accessing, by a computer system, an input vector generated by an artificial intelligence (AI) model based on an input (e.g. the system generates vectors using a CNN, i.e. an AI model).)
querying, by the computer system, a vector database using the input vector, See Pg. 10, Paragraph 9, (Vectors may be formed and stored in a vector base library to be queried after feature extraction by searching the vector base library, i.e. query a vector database using the input vector.)
wherein the vector database comprises (ii) a vector portion including a plurality of vectors stored in a non- volatile memory of the computer system, See Pg. 6, Paragraph 6, (The process of vector retrieval comprises constructing a vector base library containing vectors obtained by feature extraction of a large amount of data.) See Pg. 12, Paragraph 10, (Vector base library is located in memory 202 of computing device 20 or in storage device 30. Storage device 30 may be embodied as an independent storage medium or memory, i.e. (ii) a vector portion including a plurality of vectors stored in a non-volatile memory of the system.)
and causing, by the computer system, the AI model to generate a response to the input based on the one or more vectors. See Pg. 15, Paragraph 2, (Vectors included in at least one retrieval partition may be output as a query result or fed back to the user, i.e. causing, by the computer system, the AI model to generate a response to the input based on the one or more vectors (e.g. the CNN is used to generate vectors which are selected as output as part of the retrieval process).)
KUANG does not disclose the step wherein the vector database comprises (i) an index portion stored in a volatile memory of the computer system,
and wherein the querying the vector database comprises: accessing the index portion from the volatile memory of the computer system;
identifying, from the plurality of second level vector partitions, a particular second level vector partition based on the input vector,
SONG discloses the step wherein the vector database comprises (i) an index portion stored in a volatile memory of the computer system, See FIG. 1 & Paragraph [0057], (Disclosing a method for processing vector data. FIG. 1 illustrates the method comprising step S110 of accessing real-time vector data and writing said vector data into a cache. A vector data retrieval request utilizes a feature classification index established for newly added vector data, i.e. wherein the vector database comprises (i) an index portion stored in a volatile memory of the system (e.g. vector data is maintained in a cache).)
and wherein the querying the vector database comprises: accessing the index portion from the volatile memory of the computer system; See Paragraph [0057], (A vector retrieval request may be executed to facilitate retrieval of vector data wherein a feature classification index is used to obtain a target vector from a related vector data cluster, i.e. wherein querying the vector database comprises: accessing the index portion from the volatile memory of the system;)
identifying, from the plurality of second level vector partitions, a particular second level vector partition based on the input vector, See Paragraph [0087], (A master node may receive a vector retrieval request and distributes said request to multiple computing nodes to determine a number of similar representative vector data based on a similarity metric and determines a vector data cluster associated with the similar representative vector data according to the feature classification index. The master node may then obtain a final retrieval result by traversing multiple associated vector data clusters. Note [0085] wherein vector data retrieval is performs based on a target classification index constructed from a first-level index and a second-level index wherein the second-level index is formed through spatio-temporal classification used to perform vector retrieval, i.e. identifying, from the plurality of second level vector partitions a particular second level vector partition based on the input vector (e.g. vector data retrieval comprises searching a multi-level vector index structure including a second-level index).)
KUANG and SONG are analogous art because they are in the same field of endeavor, vector-based search systems. It would have been obvious to anyone having ordinary skill in the art before the effective filing date to modify the system of KUANG to include the method of scoring newly-added vector data into ac ache for use during vector retrieval operations as disclosed by SONG. Paragraph [0087] of SONG discloses that the use of the feature classification index may facilitate spatio-temporal classification of vector data which results in an improvement in the efficiency of vector retrieval and accelerates the process of vector retrieval.
KUANG-SONG does not disclose the step wherein the index portion comprises a plurality of first level vector partitions and a plurality of second level vector partitions,
wherein the particular second level vector partition comprises a subset of the plurality of first level vector partitions;
identifying, from the subset of the plurality of first level vector partitions, one or more first level vector partitions based on the input vector;
and retrieving one or more vectors corresponding to the one or more first level vector partitions from the non-volatile memory of the computer system;
Deep discloses the step wherein the index portion comprises a plurality of first level vector partitions and a plurality of second level vector partitions, See FIG. 7 & Paragraphs [0072]-[0073], (Disclosing a system for retrieving query target objects from partitions of a spatial index comprising vector data. FIG. 7 illustrates a plurality of partitioning approaches for processing vector objects including partitioning approach 753 of constructing a multi-level structure using hierarchical navigable small world (HNSW) graphs. An HNSW graph is constructed bottom-up for each of the partitions corresponding to one or more leaf nodes of a cover tree, i.e. wherein the index portion comprises a plurality of first level vector partitions and a plurality of second level vector partitions (e.g. an HNSW graph structure comprises a plurality of layers where links are separated across different layers with overlapping members).)
wherein the particular second level vector partition comprises a subset of the plurality of first level vector partitions; See FIG. 7 & Paragraphs [0072]-[0073], (FIG. 7 illustrates partitioning approach 753 of constructing a multi-level structure using hierarchical navigable small world (HNSW) graphs. An HNSW graph is constructed bottom-up for each of the partitions corresponding to one or more leaf nodes of a cover tree, i.e. wherein the particular second level vector partition comprises a subset of the plurality of first level vector partitions (e.g. an HNSW graph structure comprises a plurality of layers where links are separated across different layers with overlapping members);)
identifying, from the subset of the plurality of first level vector partitions, one or more first level vector partitions based on the input vector; See FIG. 7 & Paragraphs [0072]-[0073], (FIG. 7 illustrates partitioning approach 753 of constructing a multi-level structure using hierarchical navigable small world (HNSW) graphs. An HNSW graph is constructed bottom-up for each of the partitions corresponding to one or more leaf nodes of a cover tree.)
The examiner notes that searching an HNSW graph comprises entering a top layer of the HNSW graph and traversing edges across each layer to reach a nearest neighbor to the query vector representing a query.
and retrieving one or more vectors corresponding to the one or more first level vector partitions from the non-volatile memory of the computer system; See Paragraph [0031], (A query or partition request may be sent to one or more query targets and the system may determine whether the retrieved partition includes an embedding vector that satisfies the similarity criterion with respect to a first embedding vector.) See Paragraph [0024], (If a sufficiently similar vector is not present in the local cache, a partition map created for the spatial index may be consulted to identify a relevant record, i.e. retrieving one or more vectors corresponding to the one or more first level vector partitions (e.g. vectors may be retrieved form a spatial index such as the HNSW graphs which search a layered structure for relevant vectors and may therefore retrieve vectors from a topmost layer) from the non-volatile memory of the system (e.g. the system may consult non-cache memory to determine a relevant vector);
KUANG, SONG and Deep are analogous art because they are in the same field of endeavor, vector-based search systems. It would have been obvious to anyone having ordinary skill in the art before the effective filing date to modify the system of KUANG-SONG to include the method of retrieving vector data based on a spatial index as disclosed by Deep. Paragraph [0025] of Deep discloses that the system may provide at least the following advantages with regard to object recognition tasks: (a) reducing the average time taken between the capture of an image and the recognition or non-recognition of objects or entities represented in the image, thus enabling response actions to be initiated rapidly regardless of whether objects were recognized or not, (b) reducing the overall amount of networking bandwidth that has to be utilized for object recognition applications, and/or (c) enhancing data security for object recognition applications by reducing the number of times potentially sensitive data is transferred over the public Internet or other types of less secure network pathways.
Regarding dependent claim 9,
As discussed above with claim 8, KUANG-SONG-Deep discloses all of the limitations.
KUANG further discloses the step wherein the vector database further comprises a non-indexed vector portion stored in the volatile memory, and wherein the non-indexed vector portion comprises a second plurality of vectors that is not indexed. See Pg. 6, Paragraph 6, (The process of vector retrieval comprises constructing a vector base library containing vectors obtained by feature extraction of a large amount of data, i.e. wherein the vector database further comprises a non-indexed vector portion stored in the volatile memory, wherein the non-indexed vector portion comprises a second plurality of vectors that is not indexed (e.g. the method of KUANG does not disclose the use of an index, therefore vectors of the vector base library are stored in a vector format and retrieved in a vector format).)
Regarding dependent claim 10,
As discussed above with claim 9, KUANG-SONG discloses all of the limitations.
KUANG further discloses the step wherein the querying the vector database further comprises comparing the input vector with each of the second plurality of vectors in the non- indexed vector portion. See Pg. 13, Paragraph 6, (The vector retrieval stage includes selecting an unselected retrieval partition from a number "K" of retrieval partitions as a target retrieval partition and calculating a similarity between a query vector and each vector contained in the target retrieval partition, i.e. wherein the querying the vector database further comprises comparing the input vector with each of the second plurality of vectors in the non-indexed vector portion (e.g. the query vector is compared with the plurality of vector partitions until no unselected partitions of K partitions remain).)
Regarding dependent claim 11,
As discussed above with claim 9, KUANG-SONG-Deep discloses all of the limitations.
KUANG further discloses the step of obtaining an additional vector generated by the AI model; See Pg. 12, Paragraph 10, (The system may utilize a feature extraction model wherein a set of sample images may be input into a convolutional neural network (CNN) for feature extraction such that the CNN model outputs a vector corresponding to each sample image to be stored in a vector base library, i.e. obtaining an additional vector generated by the AI model (e.g. for example, a new batch of sample images may be submitted by computing device 20).)
and storing the additional vector in the non-indexed vector portion of the vector database. See Pg. 6, Paragraph 6, (The process of vector retrieval comprises constructing a vector base library containing vectors obtained by feature extraction of a large amount of data, i.e. storing the additional vector in the non-indexed vector portion of the vector database (e.g. vector data is stored in the vector base library embodied as an independent storage medium).)
Regarding dependent claim 13,
As discussed above with claim 9, KUANG-SONG-Deep discloses all of the limitations.
Deep further discloses the step wherein the index portion comprises an index that represents the plurality of first level vector partitions and the plurality of second level vector partitions, and wherein each second level vector partition in the plurality of second level vector partitions corresponds to a different subset of the plurality of first vector partitions. See FIG. 7 & Paragraphs [0072]-[0073], (FIG. 7 illustrates a plurality of partitioning approaches for processing vector objects including partitioning approach 753 of constructing a multi-level structure using hierarchical navigable small world (HNSW) graphs. An HNSW graph is constructed bottom-up for each of the partitions corresponding to one or more leaf nodes of a cover tree, i.e. wherein the index portion comprises an index that represents the plurality of first level vector partitions (e.g. a topmost level) and the plurality of second level vector partitions, and wherein each second level vector partition in the plurality of second level vector partitions corresponds to a different subset of the plurality of first level vector partitions (e.g. layers of the HNSW graph contain different subsets of a complete graph. The bottom layer contains all nodes while upper layers contain fewer nodes).)
Regarding dependent claim 14,
As discussed above with claim 8, KUANG-SONG-Deep discloses all of the limitations.
Deep further discloses the step wherein the plurality of vectors is divided into the plurality of first level vector partitions based on a first clustering process performed on the plurality of vectors, and wherein the plurality of first level vector partitions is divided into the plurality of second level vector partitions based on a second clustering process performed on the plurality of first level vector partitions. See FIG. 7 & Paragraphs [0072]-[0073], (FIG. 7 illustrates a plurality of partitioning processes including approach 752 wherein an algorithm A1 may be used to generate a cover tree and algorithm A2 of constructing partitions P1, P2, P3 using HNSW, wherein the plurality of vectors is divided into the plurality of first level vector partitions based on a first clustering process performed on the plurality of vectors and wherein the plurality of first level vector partitions is divided into the plurality of second level vector partitions based on a second clustering process performed on the plurality of first level vector partitions (e.g. an algorithm A1 may generate a cover tree that is associated with the partitions generated using HNSW comprising a plurality of partitions having layers).)
Regarding independent claim 15,
KUANG discloses a non-transitory machine-readable medium having stored thereon machine-readable instructions executable to cause a machine to perform operations comprising: obtaining an input vector generated by an artificial intelligence (AI) model; See Pg. 12, Paragraph 10, (Disclosing a vector retrieval method for improving retrieval speed of vector-based data. The system may utilize a feature extraction model wherein a set of sample images may be input into a convolutional neural network (CNN) for feature extraction such that the CNN model outputs a vector corresponding to each sample image to be stored in a vector base library, i.e. accessing, by a computer system, an input vector generated by an artificial intelligence (AI) model based on an input (e.g. the system generates vectors using a CNN, i.e. an AI model).)
querying a vector database using the input vector, See Pg. 10, Paragraph 9, (Vectors may be formed and stored in a vector base library to be queried after feature extraction by searching the vector base library, i.e. querying a vector database using the input vector.)
wherein the vector database comprises (ii) a vector portion including a plurality of vectors stored in a non-volatile memory, See Pg. 6, Paragraph 6, (The process of vector retrieval comprises constructing a vector base library containing vectors obtained by feature extraction of a large amount of data.) See Pg. 12, Paragraph 10, (Vector base library is located in memory 202 of computing device 20 or in storage device 30. Storage device 30 may be embodied as an independent storage medium or memory, i.e. (ii) a vector portion including a plurality of vectors stored in a non-volatile memory of the system.)
and causing the AI model to generate data based on the one or more vectors. See Pg. 15, Paragraph 2, (Vectors included in at least one retrieval partition may be output as a query result or fed back to the user, i.e. causing the AI model to generate a response to the input based on the one or more vectors (e.g. the CNN is used to generate vectors which are selected as output as part of the retrieval process).)
KUANG does not disclose the step wherein the vector database comprises (i) an index portion stored in a volatile memory,
wherein the querying the vector database comprises: accessing the index portion from the volatile memory;
SONG discloses the step wherein the vector database comprises (i) an index portion stored in a volatile memory, See FIG. 1 & Paragraph [0057], (Disclosing a method for processing vector data. FIG. 1 illustrates the method comprising step S110 of accessing real-time vector data and writing said vector data into a cache. A vector data retrieval request utilizes a feature classification index established for newly added vector data, i.e. wherein the vector database comprises (i) an index portion stored in a volatile memory of the system (e.g. vector data is maintained in a cache).)
wherein the querying the vector database comprises: accessing the index portion from the volatile memory; See Paragraph [0057], (A vector retrieval request may be executed to facilitate retrieval of vector data wherein a feature classification index is used to obtain a target vector from a related vector data cluster, i.e. wherein querying the vector database comprises: accessing the index portion from the volatile memory of the system;)
KUANG and SONG are analogous art because they are in the same field of endeavor, vector-based search systems. It would have been obvious to anyone having ordinary skill in the art before the effective filing date to modify the system of KUANG to include the method of scoring newly-added vector data into ac ache for use during vector retrieval operations as disclosed by SONG. Paragraph [0087] of SONG discloses that the use of the feature classification index may facilitate spatio-temporal classification of vector data which results in an improvement in the efficiency of vector retrieval and accelerates the process of vector retrieval.
KUANG-Song does not disclose the step wherein the index portion comprises a plurality of first level vector partitions and a plurality of second level vector partitions,
identifying, from the plurality of second level vector partitions, a particular second level vector partition based on the input vector, wherein the particular second level vector partition is linked to a subset of the plurality of first level vector partitions;
identifying, form the subset of the plurality of first level vector partitions, one or more first level vector partitions based on the input vector;
and retrieving one or more vectors corresponding to the one or more first level vector partitions from the non-volatile memory;
Deep discloses the step wherein the index portion comprises a plurality of first level vector partitions and a plurality of second level vector partitions, See FIG. 7 & Paragraphs [0072]-[0073], (Disclosing a system for retrieving query target objects from partitions of a spatial index comprising vector data. FIG. 7 illustrates a plurality of partitioning approaches for processing vector objects including partitioning approach 753 of constructing a multi-level structure using hierarchical navigable small world (HNSW) graphs. An HNSW graph is constructed bottom-up for each of the partitions corresponding to one or more leaf nodes of a cover tree, i.e. wherein the index portion comprises a plurality of first level vector partitions and a plurality of second level vector partitions (e.g. an HNSW graph structure comprises a plurality of layers where links are separated across different layers with overlapping members).)
identifying, from the plurality of second level vector partitions, a particular second level vector partition based on the input vector, wherein the particular second level vector partition is linked to a subset of the plurality of first level vector partitions; See Paragraph [0091]-[0092], (The one or more spatial queries may obtain a partition from multiple query targets based on the received one or more spatial queries. A search of the received partition may be conducted to determine a vector satisfying a similarity criterion, i.e. wherein the querying the vector database further comprises: identifying, from the plurality of second level vector partitions, a second second level vector partition based on the input vector (e.g. the vectors responsive to a query found at a layer of any of the HNSW partitions P1, P2, P3) , wherein the second second level vector partition is linked to a second subset of the plurality of first level vector partitions (e.g. an HNSW graph comprises a plurality of connected layers that may be traversed to retrieve a closest vector responsive to a query. Traversal includes accessing subsequent layers if no vectors of a current layer are responsive);)
identifying, form the subset of the plurality of first level vector partitions, one or more first level vector partitions based on the input vector; See FIG. 7 & Paragraphs [0072]-[0073], (FIG. 7 illustrates partitioning approach 753 of constructing a multi-level structure using hierarchical navigable small world (HNSW) graphs. An HNSW graph is constructed bottom-up for each of the partitions corresponding to one or more leaf nodes of a cover tree.)
The examiner notes that searching an HNSW graph comprises entering a top layer of the HNSW graph and traversing edges across each layer to reach a nearest neighbor to the query vector representing a query.
and retrieving one or more vectors corresponding to the one or more first level vector partitions from the non-volatile memory; See Paragraph [0031], (A query or partition request may be sent to one or more query targets and the system may determine whether the retrieved partition includes an embedding vector that satisfies the similarity criterion with respect to a first embedding vector.) See Paragraph [0024], (If a sufficiently similar vector is not present in the local cache, a partition map created for the spatial index may be consulted to identify a relevant record, i.e. retrieving one or more vectors corresponding to the one or more first level vector partitions (e.g. vectors may be retrieved form a spatial index such as the HNSW graphs which search a layered structure for relevant vectors and may therefore retrieve vectors from a topmost layer) from the non-volatile memory of the system (e.g. the system may consult non-cache memory to determine a relevant vector);
KUANG, SONG and Deep are analogous art because they are in the same field of endeavor, vector-based search systems. It would have been obvious to anyone having ordinary skill in the art before the effective filing date to modify the system of KUANG-SONG to include the method of retrieving vector data based on a spatial index as disclosed by Deep. Paragraph [0025] of Deep discloses that the system may provide at least the following advantages with regard to object recognition tasks: (a) reducing the average time taken between the capture of an image and the recognition or non-recognition of objects or entities represented in the image, thus enabling response actions to be initiated rapidly regardless of whether objects were recognized or not, (b) reducing the overall amount of networking bandwidth that has to be utilized for object recognition applications, and/or (c) enhancing data security for object recognition applications by reducing the number of times potentially sensitive data is transferred over the public Internet or other types of less secure network pathways.
Regarding dependent claim 16,
As discussed above with claim 15, KUANG-SONG-Deep discloses all of the limitations.
KUANG further discloses the step wherein the operations further comprise: receiving, via an interface, an input from a device, See Pg. 12, Paragraph 10, (Computing device 20 may input a plurality of sample images into a feature extraction model such as the CNN model, which outputs vectors corresponding to each sample image, i.e. wherein the operations further comprise: receiving, via an interface, an input from a device)
wherein the input vector is generated based on the input; See Pg. 12, Paragraph 10, (The input image(s) from computing device 20 are provided as input to the CNN which outputs a vector corresponding to each sample image, i.e. wherein the input vector is generated based on the input;)
and transmitting the data to the device via the interface as a response to the input. See Pg. 15, Paragraph 2, (Vectors included in at least one retrieval partition may be output as a query result or fed back to the user, i.e. and transmitting the data to the device via the interface as a response to the input.)
Regarding dependent claim 17,
As discussed above with claim 15, KUANG-SONG-Deep discloses all of the limitations.
KUANG further discloses the step wherein the vector database further comprises a non-indexed vector portion stored in the volatile memory, and wherein the non-indexed vector portion comprises a second plurality of vectors that is not indexed. See Pg. 6, Paragraph 6, (The process of vector retrieval comprises constructing a vector base library containing vectors obtained by feature extraction of a large amount of data, i.e. wherein the vector database further comprises a non-indexed vector portion stored in the volatile memory, wherein the non-indexed vector portion comprises a second plurality of vectors that is not indexed (e.g. the method of KUANG does not disclose the use of an index, therefore vectors of the vector base library are stored in a vector format and retrieved in a vector format).)
Regarding dependent claim 18,
As discussed above with claim 17, KUANG-SONG-Deep discloses all of the limitations.
KUANG further discloses wherein the querying the vector database further comprises comparing the input vector with each of the second plurality of vectors in the non-indexed vector portion. See Pg. 13, Paragraph 6, (The vector retrieval stage includes selecting an unselected retrieval partition from a number "K" of retrieval partitions as a target retrieval partition and calculating a similarity between a query vector and each vector contained in the target retrieval partition, i.e. wherein the querying the vector database further comprises comparing the input vector with each of the second plurality of vectors in the non-indexed vector portion (e.g. the query vector is compared with the plurality of vector partitions until no unselected partitions of K partitions remain).)
Regarding dependent claim 19,
As discussed above with claim 15, KUANG-SONG-Deep discloses all of the limitations.
Deep further discloses the step wherein the index portion comprises an index that represents the plurality of first level vector partitions and the plurality of second level vector partitions, and wherein each second level vector partition in the plurality of second level vector partitions corresponds to a different subset of the plurality of first level vector partitions. See FIG. 7 & Paragraphs [0072]-[0073], (FIG. 7 illustrates a plurality of partitioning approaches for processing vector objects including partitioning approach 753 of constructing a multi-level structure using hierarchical navigable small world (HNSW) graphs. An HNSW graph is constructed bottom-up for each of the partitions corresponding to one or more leaf nodes of a cover tree, i.e. wherein the index portion comprises an index that represents the plurality of first level vector partitions (e.g. a topmost level) and the plurality of second level vector partitions, and wherein each second level vector partition in the plurality of second level vector partitions corresponds to a different subset of the plurality of first level vector partitions (e.g. layers of the HNSW graph contain different subsets of a complete graph. The bottom layer contains all nodes while upper layers contain fewer nodes).)
Regarding dependent claim 20,
As discussed above with claim 15, KUANG-SONG-Deep discloses all of the limitations.
Deep further discloses the step wherein the plurality of vectors is divided into the plurality of first level vector partitions based on a first clustering process performed on the plurality of vectors, and wherein the plurality of first level vector partitions is divided into the plurality of second level vector partitions based on a second clustering process performed on the plurality of first level vector partitions. See FIG. 7 & Paragraphs [0072]-[0073], (FIG. 7 illustrates a plurality of partitioning processes including approach 752 wherein an algorithm A1 may be used to generate a cover tree and algorithm A2 of constructing partitions P1, P2, P3 using HNSW, wherein the plurality of vectors is divided into the plurality of first level vector partitions based on a first clustering process performed on the plurality of vectors and wherein the plurality of first level vector partitions is divided into the plurality of second level vector partitions based on a second clustering process performed on the plurality of first level vector partitions (e.g. an algorithm A1 may generate a cover tree that is associated with the partitions generated using HNSW comprising a plurality of partitions having layers).)
Claim(s) 3-5 is/are rejected under 35 U.S.C. 103 as being unpatentable over KUANG in view of SONG and Deep as applied to claim 2 above, and further in view of Mirman (US Patent No.: 11,995,109; Date of Patent: May 28, 2024).
Regarding dependent claim 3,
As discussed above with claim 2, KUANG-SONG-Deep discloses all of the limitations.
KUANG-SONG-Deep does not disclose the step wherein the plurality of vectors is divided into the plurality of first level vector partitions based on a first clustering process performed on the plurality of vectors, and wherein the plurality of first level vector partitions is divided into the plurality of second level vector partitions based on a second clustering process performed on the plurality of first level vector partitions.
Mirman discloses the step wherein the plurality of vectors is divided into the plurality of first level vector partitions based on a first clustering process performed on the plurality of vectors, and wherein the plurality of first level vector partitions is divided into the plurality of second level vector partitions based on a second clustering process performed on the plurality of first level vector partitions. See Col. 8, line 39 - Col. 9, line 17, (Disclosing a process for obtaining vectors in an embedding space. The process includes creating an HNSW graph and includes selecting a base layer for the graph that serves as the foundation of the HNSW structure. Each node in the base layer may be connected to its closest neighbors based on a distance metric. Additional layers may be constructed in a hierarchical manner wherein each successive layer may be smaller in the number of nodes than the one below it by containing a subset of elements from the previous layer, i.e. wherein the plurality of vectors is divided into the first level of vector partitions based on a first clustering process performed on the plurality of vectors, wherein vectors within each vector partition of the first level of vector partitions are further divided into a subset of vector partitions in the second level of vector partitions based on a second clustering process performed on the vectors (e.g. FIG. 7 illustrates a plurality of partitioning processes including approach 752 wherein an algorithm A1 may be used to generate a cover tree and algorithm A2 of constructing partitions P1, P2, P3 using HNSW).)
KUANG, SONG, Deep and Mirman are analogous art because they are in the same field of endeavor, vector-based search systems. It would have been obvious to anyone having ordinary skill in the art before the effective filing date to modify the system of KUANG-SONG-Deep to include the method of searching an HNSW structure to retrieve vector data as disclosed by Mirman. Col. 9, lines 14-17 of Mirman disclose that the process may be adapted to handle dynamic datasets where elements are added or removed over time, which allows the HNSW structure to adapt to the system's needs.
Regarding dependent claim 4,
As discussed above with claim 3, KUANG-SONG-Deep-Mirman discloses all of the limitations.
Mirman further discloses the step wherein the second clustering process is performed using a corresponding parameter determined based on a number of vectors associated with each of the plurality of first level vector partitions, and wherein the corresponding parameter specifies a number of first level vector partitions to be assigned to each of the second level vector partitions. See Col. 8, line 39 - Col. 9, line 17, (Disclosing a process for obtaining vectors in an embedding space. The process includes creating an HNSW graph and includes performing optimization techniques including tuning the parameters of the graph such as the number of layers, number of connections per element or criteria for selecting elements for higher layers. This allows the system to adapt to handle dynamic datasets where elements are added or removed over time, i.e. wherein the second clustering process is performed on each corresponding vector partition in the first level of vector partitions using a corresponding parameter determined based on a number of vectors associated with the corresponding vector partition, and wherein the corresponding parameter specifies a number of vector partitions into which the corresponding vector partition is divided (e.g. the parameters of the HNSW ).)
Regarding dependent claim 5,
As discussed above with claim 2, KUANG-SONG-Deep discloses all of the limitations.
KUANG-SONG-Deep does not disclose the step the particular second level vector partition is a first second level vector partition, wherein the subset of the plurality of first level vector partitions is a first subset of the plurality of first level vector partitions, See FIG. 7 & Paragraphs [0072]-[0073], (FIG. 7 illustrates partitioning approach 753 of constructing a multi-level structure using hierarchical navigable small world (HNSW) graphs. An HNSW graph is constructed bottom-up for each of the partitions corresponding to one or more leaf nodes of a cover tree.) See Paragraph [0091], (Spatial queries may be directed to the plurality of partitions to identify and receive a target partition of the plurality of partitions, i.e. wherein the particular second level vector partition is a first second level vector partition (e.g. FIG. 7 illustrates a plurality of partitions P1, P2, P3 comprising a plurality of layers each. A spatial search may identify any of the relevant partitions as a query target), wherein the subset of the plurality of first level vector partitions is a first subset of the plurality of first level vector partitions, (e.g. a query target may be discovered at a subsequent layer of an HNSW graph partition, for example a middle layer of P1).)
and wherein the querying the vector database further comprises: identifying, from the plurality of second level vector partitions, a second second level vector partition based on the input vector, wherein the second second level vector partition is linked to a second subset of the plurality of first level vector partitions; See Paragraph [0091]-[0092], (The one or more spatial queries may obtain a partition from multiple query targets based on the received one or more spatial queries. A search of the received partition may be conducted to determine a vector satisfying a similarity criterion, i.e. wherein the querying the vector database further comprises: identifying, from the plurality of second level vector partitions, a second second level vector partition based on the input vector), wherein the second second level vector partition is linked to a second subset of the plurality of first level vector partitions (e.g. an HNSW graph comprises a plurality of connected layers that may be traversed to retrieve a closest vector responsive to a query. Traversal includes accessing subsequent layers if no vectors of a current layer are responsive);)
and retrieving one or more additional vectors corresponding to the second subset of the plurality of first level vector partitions from the non-volatile memory of the system, See Paragraph [0031], (A query or partition request may be sent to one or more query targets and the system may determine whether the retrieved partition includes an embedding vector that satisfies the similarity criterion with respect to a first embedding vector.) See Paragraph [0024], (If a sufficiently similar vector is not present in the local cache, a partition map created for the spatial index may be consulted to identify a relevant record, i.e. retrieving one or more additional vectors corresponding to the second subset of the plurality of first level vector partitions from the non-volatile memory of the system (e.g. a responsive vector may be found at any layer of an HNSW graph. For example, of a first-traversed layer of a partition P1 does not contain the a similar enough vector, a next layer may be traversed to determine a responsive vector).)
KUANG-SONG-Deep does not disclose the step wherein providing the response is further based on the one or more additional vectors.
Mirman discloses the step wherein providing the response is further based on the one or more additional vectors. See FIG. 2 & Col. 11, lines 29-41, (Disclosing a process for obtaining vectors in an embedding space. FIG. 2 illustrates method 200 of retrieving vector data from an HNSW graph. Method 200 comprises step 218 of determining a current-closest pair of nodes in first and second subsets of unstructured records wherein the pair of nodes correspond to a pair of vectors in an embedding space. At step 214, the current-closest pair and minimum distance may be output upon detection of a stopping condition at step 212, i.e. wherein providing the response is further based on the one or more additional vectors (e.g. the response comprises a pair of nodes from each of a plurality of HNSW graphs and one from each first and second sets of unstructured records).)
KUANG, SONG, Deep and Mirman are analogous art because they are in the same field of endeavor, vector-based search systems. It would have been obvious to anyone having ordinary skill in the art before the effective filing date to modify the system of KUANG-SONG-Deep to include the method of searching an HNSW structure to retrieve vector data as disclosed by Mirman. Col. 9, lines 14-17 of Mirman disclose that the process may be adapted to handle dynamic datasets where elements are added or removed over time, which allows the HNSW structure to adapt to the system's needs.
Claim(s) 12 is/are rejected under 35 U.S.C. 103 as being unpatentable over KUANG in view of SONG and Deep as applied to claim 9 above, and further in view of Mortazavi et al. (US PGPUB No. 2016/0306836; Pub. Date: Oct. 20, 2016).
Regarding dependent claim 12,
As discussed above with claim 9, KUANG-SONG-Deep discloses all of the limitations.
KUANG-SONG-Deep does not disclose the step of determining that a size of the second plurality of vectors in the non-indexed vector portion exceeds a size threshold; and generating a second index for the vector database based on indexing the second plurality of vectors.
Mortazavi discloses the step of determining that a size of the second plurality of vectors in the non-indexed vector portion exceeds a size threshold; and generating a second index for the vector database based on indexing the second plurality of vectors. See Paragraph [0022], (Disclosing a method for storing data records as data segments in a storage element. The system may determine whether a storage element is nearly full, i.e. determining that a size of the second plurality of vectors in the non-indexed vector portion exceeds a size threshold, and consequently compresses a lightweight index that summarizes aspects of each of the data segments and appends this compressed lightweight index to the storage element. The lightweight index is generated by resolving a user-defined function on the attributes to generate maximum descriptors and minimum descriptors for each data segment, i.e. generating a second index for the vector database based on indexing the second plurality of vectors (e.g. the lightweight index is generated, compressed and appended to the storage element).)
KUANG, SONG, Deep and Mortazavi are analogous art because they are in the same field of endeavor, search systems. It would have been obvious to anyone having ordinary skill in the art before the effective filing date to modify the system of KUANG-SONG-Deep to include the method of appending an index to a storage element based on determining a fill level of said storage element as disclosed by Mortazavi. Paragraph [0029] of Mortazavi discloses that the data storage scheme 300 allows for optimal data band 310 storage space efficiency, which is beneficial for large amounts of data without requiring a significant operational cost as cold data is not frequently modified.
Response to Arguments
Applicant’s arguments with respect to claim(s) 1, 8 and 15 have been considered but are moot because the new ground of rejection does not rely on any reference applied in the prior rejection of record for any teaching or matter specifically challenged in the argument.
Regarding at least independent claim 1,
Applicant argues that the combination of KUANG and SONG does not disclose at least the following amended limitations:
query a vector database using the input vector, wherein the vector database comprises (i) an index portion stored in a volatile memory of the system and (ii) a vector portion including a plurality of vectors stored in a non-volatile memory of the system, wherein the index portion comprises a plurality of first level vector partitions and a plurality of second level vector partitions, and wherein querying the vector database comprises: accessing the index portion from the volatile memory of the system;
identifying, from the plurality of second level vector partitions a particular second level vector partition based on the input vector, wherein the particular second level vector partition comprises a subset of the plurality of first level vector partitions;
identifying, from the subset of the plurality of first level vector partitions, one or more first level vector partitions based on the input vector;
and retrieving one or more vectors corresponding to the one or more first level vector partitions from the non-volatile memory of the system;
The examiner agrees that KUANG and SONG do not disclose the entirety of the amended claim as currently presented. Deep et al. (US PGPUB No. 2024/0320951; Pub. Date: Sep. 26, 2024) is relied upon to disclose the limitations at issue in combination with KUANG and SONG as described in the rejection(s) above.
Applicant’s amendments modify the scope of the claimed invention and therefore necessitated the new grounds of rejection presented in this Office Action.
Conclusion
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any nonprovisional extension fee (37 CFR 1.17(a)) pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the mailing date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to Fernando M Mari whose telephone number is (571)272-2498. The examiner can normally be reached Monday-Friday 7am-4pm.
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, Ann J. Lo can be reached at (571) 272-9767. 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.
/FMMV/Examiner, Art Unit 2159 /ALBERT M PHILLIPS, III/Primary Examiner, Art Unit 2159
Prosecution Insights