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 .
Information Disclosure Statement
The information disclosure statement (IDS) submitted on August 20, 2024, was considered by the examiner. The submission is compliant with the provisions of 37 CFR 1.97.
Claim Objections
Claim 10 is objected to because of the following informalities: There are two periods within the claim, one at ‘10. The method of claim 1, wherein the machine learning model comprises a large language model.’ , and one at ‘and wherein the method is performed by one or more computing devices.’ Appropriate correction is required.
Claim 10 is a dependent claim and limitations of “training a machine learning model using training prompts querying a vector store to prevent unauthorized user disclosure of data derived from the vector store; receiving a prompt; generating, using the machine learning model, a response to the prompt based at least in part on the vector store; and wherein the method is performed by one or more computing devices” are same as those in claim 1, and claim 10 is objected to since it is not clear it these are meant to be new limitations or are there by error. Appropriate correction is required.
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.
Claim 1 recite the limitation "the hierarchical store" in “A method comprising: training a machine learning model using training prompts querying a vector store to prevent unauthorized user disclosure of data derived from the hierarchical store.” There is insufficient antecedent basis for this limitation in the claim. Appropriate correction is required.
Claim 10 is a dependent claim and limitations of “training a machine learning model using training prompts querying a vector store to prevent unauthorized user disclosure of data derived from the vector store; receiving a prompt; generating, using the machine learning model, a response to the prompt based at least in part on the vector store; and wherein the method is performed by one or more computing devices” are same as those in claim 1, and claim 10 is rejected under 35 U.S.C. 112(b) for being indefinite since it is not clear it these are meant to be new limitations or are there by error. Appropriate correction is required.
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 because the claimed invention is directed to an abstract idea (mental process) without significantly more.
Claim 1:
Regarding claim 1, in step 1 of the 101-analysis set forth in MPEP 2106, the claim recites “A method comprising: training a machine learning model using training prompts querying a vector store to prevent unauthorized user disclosure of data derived from the hierarchical store…”, and a method is one of the four statutory categories of invention.
In step 2A prong 1 of the 101-analysis set forth in the MPEP 2106, the examiner has determined that the following limitations recite a process that, under the broadest reasonable interpretation, covers a mental process but for recitation of generic computer components:
i. generating, … a response to the prompt based at least in part on the vector store; (mental process, a person can mentally evaluate and create a response to a prompt from viewing a vector store, see MPEP 2106.04(a)(2)(III)),
If claim limitations, under their broadest reasonable interpretation, cover performance of the limitations as a mental process but for the recitation of generic computer components, then it falls within the mental process grouping of abstract ideas. Accordingly, the claim “recites” an abstract idea.
In step 2A prong 2 of the 101-analysis set forth in MPEP 2106, the examiner has determined that the following additional elements do not integrate this judicial exception into a practical application:
ii. A method comprising: training a machine learning model using training prompts querying a vector store to prevent unauthorized user disclosure of data derived from the hierarchical store; (In step 2A, prong 2, this recites mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)),
iii. receiving a prompt (In step 2A, prong 2, this recites mere data gathering, which is considered an insignificant extra-solution activity – see MPEP 2106.05(g)),
iv. … using the machine learning model, (In step 2A, prong 2, this recites mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)),
v. … wherein the method is performed by one or more computing devices, (In step 2A, prong 2, this recites mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)),
Since the claim as a whole, looking at the additional elements individually and in combination, does not contain any other additional elements that are indicative of integration into a practical application, the claim is “directed” to an abstract idea.
In step 2B of the 101-analysis set forth in the 2019 PEG, the examiner has determined that the claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception.
As discussed above, additional elements ii, iv, and v recite mere instructions to apply the judicial exception using generic computer components, which are not indicative of significantly more. The additional element iii recites mere data gathering, and is considered an insignificant extra-solution activity. In step 2B, this insignificant extra-solution activity is a well understood routine and conventional activity, which includes receiving or transmitting data over a network from court case Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362 (utilizing an intermediary computer to forward information); TLI Communications LLC v. AV Auto. LLC, 823 F.3d 607, 610, 118 USPQ2d 1744, 1745 (Fed. Cir. 2016), – see MPEP 2106.05(d) (II)(i)),
Considering the additional elements individually and in combination, and the claim as a whole, the additional elements do not provide significantly more than the abstract idea. Therefore, the claim is not patent eligible.
Claim 2:
Regarding claim 2, it is dependent upon claim 1, and thereby incorporates the limitations of, and corresponding analysis applied to claim 1. Further, claim 2 recites the following additional elements:
i. 2. The method of claim 1, wherein training the machine learning model comprises:
penalizing the machine learning model based on a first response to a first training prompt comprising sensitive information corresponding to an individual one of the plurality of vector stores; (In step 2A, prong 2, this is considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)), (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)),
ii. and rewarding the machine learning model based on a second response to a second training prompt comprising aggregate information corresponding to multiple of the plurality of individual vector stores, (In step 2A, prong 2, this is considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)), (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)),
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Claim 3:
Regarding claim 3, it is dependent upon claim 1, and thereby incorporates the limitations of, and corresponding analysis applied to claim 1. Further, claim 3 recites the following additional elements:
i. 3. The method of claim 1, wherein the vector store stores one or more vector embeddings, (In step 2A, prong 2, this is considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)), (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)),
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Claim 4:
Regarding claim 4, it is dependent upon claim 3, and thereby incorporates the limitations of, and corresponding analysis applied to claim 3. Further, claim 4 recites the following abstract ideas:
i. 4. The method of claim 3, further comprising: generating a query vector based at least in part on the prompt; (This recites a mental process, since a person can mentally evaluate and create a query vector from viewing a prompt, see MPEP 2106.04(a)(2)(III)),
ii. comparing the query vector with the one or more vector embeddings; (This recites a mental process, since a person can mentally evaluate and compare a query vector and one or more vector embeddings, see MPEP 2106.04(a)(2)(III)),
iii. identify one or more vector similar vector embeddings from the vector store; wherein the one or more similar vector embeddings having a similarity to the query vector that meets or exceeds a predetermined threshold of similarity; (This recites a mental process, since a person can mentally evaluate and identify if one or more vector embeddings are similar, and if they are similar to the query vector that satisfies a preset threshold, see MPEP 2106.04(a)(2)(III)),
iv. and wherein the response is generated based at least in part on the one or more similar vector embeddings, (This recites a mental process, since a person can mentally evaluate and create one or more vector embeddings with pen and paper, see MPEP 2106.04(a)(2)(III)),
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Claim 5:
Regarding claim 5, it is dependent upon claim 4, and thereby incorporates the limitations of, and corresponding analysis applied to claim 4. Further, claim 5 recites the following abstract idea:
i. The method of claim 4, further comprising determining the similarity of the one or more similar vector embeddings to the query vector based at least in part on a cosine similarity of the one or more vector embeddings to the query vector. (This recites a mental process, since a person can mentally evaluate and determine a similarity of the vector embeddings using a cosine similarity, see MPEP 2106.04(a)(2)(III)),
If claim limitations, under their broadest reasonable interpretation, cover performance of the limitations as a mental process but for the recitation of generic computer components, then it falls within the mental process grouping of abstract ideas. Accordingly, the claim “recites” an abstract idea. Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Claim 6:
Regarding claim 6, it is dependent upon claim 4, and thereby incorporates the limitations of, and corresponding analysis applied to claim 4. Claim 6 recites the following abstract ideas:
i. 6. The method of claim 4, further comprising: decoding the one or more similar vector embeddings to obtain data related to the prompt; (This recites a mental process, since a person can mentally evaluate and identify if one or more vector embeddings is related to the prompt, see MPEP 2106.04(a)(2)(III)),
ii. and wherein the response is generated based at least in part on the data related to the prompt, (This recites a mental process, since a person can mentally evaluate and create a response for data related to the prompt, see MPEP 2106.04(a)(2)(III)),
If claim limitations, under their broadest reasonable interpretation, cover performance of the limitations as a mental process but for the recitation of generic computer components, then it falls within the mental process grouping of abstract ideas. Accordingly, the claim “recites” an abstract idea.
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Claim 7: Regarding claim 7, it is dependent upon claim 6, and thereby incorporates the limitations of, and corresponding analysis applied to claim 6. Further, claim 7 recites the following additional element:
i. 7. The method of claim 6, wherein the data related to the prompt comprises a context associated with the one or more similar vector embeddings. (In step 2A, prong 2, this recites an indication to a field of use or technological environment – see MPEP 2106.05(h)), (In step 2B, this also recites a field of use or technological environment – see MPEP 2106.05(h)),
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Claim 8:
Regarding claim 8, it is dependent upon claim 7, and thereby incorporates the limitations of, and corresponding analysis applied to claim 7. Further, claim 8 recites the following additional elements:
i. 8. The method of claim 7, wherein the machine learning model is a first machine learning model, the method further comprising determining, using a second machine learning model, whether the response comprises sensitive information corresponding to an individual one of the plurality of vector stores, (In step 2A, prong 2, this is considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)), (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)),
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Claim 9:
Regarding claim 9, it is dependent upon claim 1, and thereby incorporates the limitations of, and corresponding analysis applied to claim 1. Further, claim 9 recites the following abstract ideas:
i. The method of claim 1, further comprising: generating a plurality of individualized responses to the prompt, each of the plurality of individualized responses corresponding to one of the plurality of vector stores; (mental process, a person can mentally evaluate and create responses to a prompt, see MPEP 2106.04(a)(2)(III)),
ii. determining a similarity between the response and each of the plurality of individualized responses; (mental process, a person can mentally evaluate and determine a similarity, see MPEP 2106.04(a)(2)(III)),
iii. determining to mask the response based at least in part on the response having a similarity to fewer than a predetermined number of the plurality of individualized responses; (mental process, a person can mentally evaluate and mask or obscure with pen and paper a response, see MPEP 2106.04(a)(2)(III)),
iv. and regenerating, … the response to the prompt based at least in part on the vector store, (mental process, a person can mentally evaluate and re-create a response to a prompt from viewing a vector store, see MPEP 2106.04(a)(2)(III)),
Further, claim 9 recites the following additional element:
v. using the machine learning model, (In step 2A, prong 2, this is considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)). (In step 2B, this is also considered mere instructions to apply an exception using generic computer – see MPEP 2106.05(f)).
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Claim 10:
Regarding claim 10, it is dependent upon claim 1, and thereby incorporates the limitations of, and corresponding analysis applied to claim 1. Further, claim 10 recites the following additional elements:
i. The method of claim 1, wherein the machine learning model comprises a large language model, (In step 2A, prong 2, this recites an indication to a field of use or technological environment – see MPEP 2106.05(h)), (In step 2B, this also recites a field of use or technological environment – see MPEP 2106.05(h)),
Since claim 10 recites similar limitations as corresponding claim 1 listed above, it is rejected for similar reasons under 35 U.S.C. 101.
Claim 11:
Regarding claim 11, in step 1 of the 101-analysis set forth in MPEP 2106, the claim recites “One or more non-transitory storage media storing instructions which, when executed by one or more computing devices, cause:…”, and a non-transitory storage media or machine is one of the four statutory categories of invention.
Since the claim as a whole, looking at the additional elements individually and in combination, does not contain any other additional elements that are indicative of integration into a practical application, the claim is “directed” to an abstract idea.
In step 2B of the 101-analysis set forth in the 2019 PEG, the examiner has determined that the claim does not include additional elements that are sufficient to amount to significantly more than the judicial exception.
As discussed above, additional element i recites mere instructions to apply the judicial exception using generic computer components, which is not indicative of significantly more.
Regarding claim 11, it recites similar limitations as corresponding claim 1 listed above, and it is rejected for similar reasons under 35 U.S.C. 101.
Claims 12-19:
Since claims 12-19 recite similar limitations as corresponding claims 2-9 listed above, they are rejected for similar reasons under 35 U.S.C. 101.
Claim 20:
Regarding claim 20, it is dependent upon claim 11, and thereby incorporates the limitations of, and corresponding analysis applied to claim 11. Further, claim 20 recites the following additional element:
i. The one or more non-transitory storage media of claim 11, wherein the machine learning model comprises a large language model. (In step 2A, prong 2, this recites an indication to a field of use or technological environment – see MPEP 2106.05(h)), (In step 2B, this also recites a field of use or technological environment – see MPEP 2106.05(h)),
Since the claim does not recite additional elements that either integrate the judicial exception into a practical application, nor provide significantly more than the judicial exception, the claim is not patent eligible.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claims 1, 3, 10, 11 and 20 are rejected under 35 U.S.C. 103 as being unpatentable over Jeong, C. et al. in “A study on the implementation of generative AI services using an enterprise data-based LLM application architecture,” published on September 3, 2023, available at https://arxiv.org/pdf/2309.01105v1, (hereafter, Jeong), in view of Feyisetan, O., in “Leveraging hierarchical representations for preserving privacy and utility in text,” published on October 20, 2019, available at https://arxiv.org/pdf/1910.08917, (hereafter, Feyisetan).
Claim 1:
Regarding claim 1, Jeong teaches “A method comprising: training a machine learning model using training prompts querying a vector store … of data derived from the hierarchical store;”
See Jeong in page 9 note "4) Evaluation and Fine-tuning: After training, the model's performance is evaluated using a separate test dataset that wasn't used during training. Based on the evaluation results, fine-tuning might be necessary to improve the model's performance. This could involve adjusting hyperparameters, changing the architecture, or further training on additional data. The goal is to enhance the model's capabilities based on the insights gained from evaluation."
See Jeong in page 12, section 2.3.2 note "In recent times, it has become common to train language models like LLMs with neural network structures using large-scale
document collections to create "Learned Embeddings." During this training process, the model takes in various words as input and learns semantic relationships by making embedding vectors of contextually similar words closer together and those of dissimilar words farther apart. This approach allows the model to capture meaningful relationships within the context. " Here, Jeong shows during training, the model takes words as inputs.
Also, see Jeong in page 14, step 5) note "Input the completed prompt into the LLM and return the generated answer." Examiner construes training prompts to be input prompts or input information used during training of a model. Here, Jeong shows using words or prompts for training, and involves querying a vector database illustrated in figures 6 and 9.
PNG
media_image1.png
564
1120
media_image1.png
Greyscale
PNG
media_image2.png
562
1013
media_image2.png
Greyscale
Also, see Jeong in page 16 note in section 2.3.5 AI Chatbot, “Through AI chatbots like ChatGPT, generative AI services can be harnessed, allowing users to input prompts for desired questions and receive corresponding responses. Today, AI-driven chatbot systems encompass not only simple question-answer scenarios like FAQs, but have advanced to the level of analyzing human emotions and intentions to provide tailored responses (Jeong, C.S., 2023a).”
Also, see Jeong in page 18, in section 3.2.1 RAG based implementation procedure note "5) Integration of Prompt and Search Results: This step involves searching for information based on the prompted question and integrating relevant information. To search for contextually relevant information based on the prompt, appropriate chunks are retrieved from the vector database. These retrieved chunks are then sent to LLM to aid in the response generation process." Here, Jeong shows using the input prompts described from pages 12, 14 and 16 in retrieving information (i.e. query) from a vector database (i.e. vector store).
See Jeong in page 24, section 4.2.8 Summary of implementation cases and expected effects, mention "One crucial consideration is the choice of a vector database. While ChromaDB was used in the presented cases, there are numerous vector databases available, each with varying features, performance, and scalability. It's essential to carefully evaluate and select a database that suits the specific use case for optimal performance and efficiency". Here, Jeong shows using multiple vector databases or multiple individual vector databases. Hierarchical is described by the specification paragraph [0016] notes “The hierarchical vector store may be protected by fine-grained access control rules. The hierarchical vector store may include multiple individual vector stores, each of which is associated with a different entity”, to mean using multiple individual vector databases, which Jeong shows that a vector database is similar to a vector store, see page 11, section 2.3 for details.
Further, Jeong teaches “receiving a prompt,”
See Jeong in page 6 mention "In particular, open-source Large Language Models (LLMs) have gained traction since the release of Alpaca, which is built upon the LLaMA architecture. Following Alpaca's introduction, various LLMs like GPT4All have continued to be unveiled. GPT4All, in particular, is based on the LLaMA 7B model and was inspired by Alpaca. It collected 800,000 prompt-response pairs from the GPT-3.5-Turbo model, encompassing code, conversations, and narratives. Among these pairs, around 430,000 were designed in an assistant-style prompt-response format," Here, Jeong shows collecting prompt -response pairs from the model, which relate to receiving prompts.
See Jeong in page 3 describe " Subsequently, when user queries arise—such as inquiries regarding company dress codes through a chatbot—pertinent information can be retrieved and presented to the LLM through prompts, proving to be a more practical and efficient approach." Here, Jeong shows information retrieved through prompts, and also enhances the receiving information with prompts.
Further, Jeong teaches “generating, using the machine learning model, a response to the prompt based at least in part on the vector store;”
See Jeong in page 18, section 3.2.1 RAG based implementation procedure, mention "3) Embedding: The generated chunk-level text data is transformed into numerical vector representations. This step involves mapping words or sentences to vectors, ... 5) Integration of Prompt and Search Results: This step involves searching for information based on the prompted question and integrating relevant information. To search for contextually relevant information based on the prompt, appropriate chunks are retrieved from the vector database. These retrieved chunks are then sent to LLM to aid in the response generation process. Various search engines available within LangChain for vector store similarity search are utilized. 6) Answer Generation: Using the retrieved information as a basis, the response text is generated. At this stage, the type, length, and linguistic style of the generated text can be specified. LLM, such as OpenAI's GPT-3.5-turbo model or GPT4All, uses the similarity search module in LangChain to retrieve relevant documents and generate responses. " Here, Jeong mentions while searching for contextually relevant information based on the prompt (i.e. data related to the prompt ), appropriate vector representations or vector embeddings are retrieved from the vector database (i.e. vector store). Examiner construes store to mean a database, repository or a place that holds information related to vector and vector embeddings.
Further, Jeong teaches “… and wherein the method is performed by one or more computing devices,”
See Jeong on page 4, section 2.1 Background for Generative AI, note “encompassing unsupervised and semi-supervised algorithms that enable computers to respond to short prompts and create new content using previously generated content such as text, audio, video, images, and code”. Also, see Jeong on page 6 note “open-source Large Language Models (LLMs) have gained traction since the release of Alpaca, which is built upon the LLaMA architecture … Among these pairs, around 430,000 were designed in an assistant-style prompt-response format, making them approximately 16 times larger in scale compared to Alpaca's dataset. A notable advantage of this model is its ability to run on CPUs without the need for GPUs.” Here, Jeong mentions a computer or computing device that includes CPU memory tools to run models.
However, Jeong did not teach “to prevent unauthorized user disclosure of data derived from the hierarchical store;”
In an analogous art, Feyisetan teaches “to prevent unauthorized user disclosure of data derived from the hierarchical store;”
See Feyisetan in page 1, abstract mention "unlocking access to vast data stores for training machine learning models and supporting data driven decisions. We address this problem through the lens of dχ privacy, a generalization of Differential Privacy to non Hamming distance metrics. In this work, we explore word representations in Hyperbolic space as a means of preserving privacy in text. We provide a proof satisfying dχ-privacy, then we define a probability distribution in Hyperbolic space and describe a way to sample from it in high dimensions. Privacy is provided by perturbing vector representations of words in high dimensional Hyperbolic space to obtain a semantic generalization." Here, Feyisetan mentions getting access to data store, by exploring vector representations of words.
See Feyisetan in page 1, Introduction note "This is motivated by the ability to better encode hierarchical and semantic information in Hyperbolic space than in Euclidean space... At a high level, our algorithm preserves privacy by providing plausible deniability [20] over the contents of a user’s query. " Here, Feyisetan shows training a model to prevent privacy attacks (i.e. unauthorized user disclosure of data ) from hierarchical information.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the reference of Jeong of using a model in producing a response to the prompt from a vector store, with Feyisetan’s teaching of preventing unauthorized user disclosure of data.
One of ordinary skill in the art would be motivated to do so because by integrating Feyisetan’s framework into the methods of Jeong, one with ordinary skill in the art would achieve the goal of providing “the main advantage of this model is that it is conformal as a result, earlier research into Hyperbolic word embeddings have leveraged on this model, ” (see Feyisetan in page 4, in section A. Poincar´e ball model ).
Claim 3:
Regarding claim 3, Jeong in view of Feyisetan, teach the limitations of claim 1. Further, Jeong teaches “3. The method of claim 1, wherein the vector store stores one or more vector embeddings.”
See Jeong in page 21, section 4.2.4 Implementation of Embeddings and Vector Database Construction Steps mention “First, Chroma, the vector database used to store embeddings, is installed. The chunks are then sent to the OpenAI embedding engine to be transformed into numerical vectors. Each of these vectors, generated in this manner, is stored in Chroma for future retrieval, as illustrated in Figure 13.” Here, Jeong describes using a vector database (i.e. vector store) to store embeddings, which are vector embeddings. See figure 13 by Jeong for details.
PNG
media_image3.png
421
1226
media_image3.png
Greyscale
Claim 10:
Regarding claim 10, Jeong in view of Feyisetan, teach the limitations of claim 1. Further, Jeong teaches “…wherein the machine learning model comprises a large language model …data derived from the vector store;”
See Jeong in page 4 mention “The core aspects of the RAG model—information retrieval and utilization—are exhaustively investigated. The process of embedding document information using vectorized databases, its storage, and subsequent retrieval is elaborated. An innovative method is introduced to convert retrieval outcomes into prompts for LLM through an orchestration framework, facilitating the effective synthesis of user queries and search results to generate optimal responses.” Here, Jeong discusses using information from vectorized databases (i.e. which relate to data from the vector store).
Also, see Jeong in page 8, section 2.2.2 LLM (Large Language Model) note “Generative AI models, particularly Large Language Models (LLMs), are widely used in chatbot services like ChatGPT. These models are trained on language data, such as text, and provide results through generating text-based responses.” Here, Jeong also mentions using an LLM, which shows using a large language model.
Regarding claim 10, it comprises of similar additional limitations as independent claim 1, and is rejected under the same rationale under 35 U.S.C. 103.
Claim 11:
Regarding claim 11, it comprises of similar additional limitations as independent claim 1, and is rejected under the same rationale under 35 U.S.C. 103.
Further, Jeong teaches "11. One or more non-transitory storage media storing instructions which, when executed by one or more computing devices, cause:..."
See Jeong on page 4, section 2.1 Background for Generative AI, note “encompassing unsupervised and semi-supervised algorithms that enable computers to respond to short prompts and create new content using previously generated content such as text, audio, video, images, and code”. Also, see Jeong on page 6 note “open-source Large Language Models (LLMs) have gained traction since the release of Alpaca, which is built upon the LLaMA architecture … Among these pairs, around 430,000 were designed in an assistant-style prompt-response format, making them approximately 16 times larger in scale compared to Alpaca's dataset. A notable advantage of this model is its ability to run on CPUs without the need for GPUs.” Here, Jeong mentions a computer or computing device that includes CPU memory tools, which inherently has a non-transitory storage media, to run operations.
Claim 20:
Regarding claim 20, Jeong in view of Feyisetan, teach the limitations of claim 11. Further, Jeong teaches "The one or more non-transitory storage media of claim 11, wherein the machine learning model comprises a large language model."
See Jeong in page 4 note in section 2.1 Background for Generative AI "Generative AI is a form of artificial intelligence that utilizes extensive trained data models to generate new content such as text, images, audio, and videos. Prominent examples in this domain include ChatGPT, which is a language model service trained on vast amounts of data, and models like DALL-E and Midjourney, which focus on generating images. The categorization of generative AI models varies based on their output generation." Here, Jeong shows using a language model like ChatGPT.
Claims 2 and 12 are rejected under 35 U.S.C. 103 as being unpatentable over Jeong in view of Feyisetan, and further in view of Mosallanezhad A. et al., in US PG Pub No. US20220374606A1, published on November 24, 2022, (hereafter, Mosallanezhad), and further in view of Arora, S. et al., in "Reasoning over Public and Private Data in Retrieval-Based Systems," published on August 7, 2023, available at: https://direct.mit.edu/tacl/article/doi/10.1162/tacl_a_00580/117168%20 , (hereafter, Arora).
Claim 2:
Further, Jeong teaches "information corresponding to an individual one of the plurality of vector stores; ... to multiple of,… plurality of individual vector stores,"
See Jeong in page 21, section 4.2.4 Implementation of Embeddings and Vector Database Construction Steps mention “First, Chroma, the vector database used to store embeddings, is installed. The chunks are then sent to the OpenAI embedding engine to be transformed into numerical vectors. Each of these vectors, generated in this manner, is stored in Chroma for future retrieval, as illustrated in Figure 13.” Here, Jeong describes using a vector database (i.e. vector store) to store embeddings, which are vector embeddings. See figure 13 by Jeong for details. Jeong notes each of these vectors to show individual vectors from a vector database.
Also, see Jeong in page 3 mentioned in third paragraph “by exploring methods to implement LLM applications using the RAG architecture. It proposes procedures and methodologies to facilitate the facile implementation of such applications and presents illustrative implementation cases. To this end, the study initially examines approaches to overcome the information scarcity of LLM through fine-tuning or direct utilization of document information. It subsequently delves into the operational mechanics and key phases of the RAG model, discussing methods of information storage and retrieval using vectorized databases." Later, see Jeong in page 4 mention "The process of embedding document information using vectorized databases, its storage, and subsequent retrieval is elaborated.” Here, Jeong shows retrieving information related to vectorized databases. These vectorized databases contain embedding information from vector embeddings, and relates to information to vector stores. Here, Jeong also shows using vectorized databases, which is a plural term and relates to multiple vector stores since a store also means a database.
See Jeong in page 14 note “The Question Answering process using general embedding, which allows LLM to answer a given question based on information additionally injected, is shown in Figure 7. (Kim G. H., 2023). 1) Divide the entire information-rich text into smaller chunks and calculate embeddings for each chunk. 2) Compute the embedding for the query. 3) Calculate cosine similarity between query embedding and each chunk embedding, sorting the chunks based on similarity.” Also, see Jeong in page 18 in section 3.2.1 note “5) Integration of Prompt and Search Results: This step involves searching for information based on the prompted question and integrating relevant information. To search for contextually relevant information based on the prompt, appropriate chunks are retrieved from the vector database. These retrieved chunks are then sent to LLM to aid in the response generation process. Various search engines available within LangChain for vector store similarity search are utilized.” Here, Jeong shows individual vector embeddings for each chunk, and retrieving appropriate chunks (i.e. showing individual embedding data) from the vector database (i.e. vector store).
However, Jeong in view of Feyisetan, did not teach “The method of claim 1, wherein training the machine learning model comprises: penalizing the machine learning model based on a first response to a first training prompt comprising sensitive information … and rewarding the machine learning model based on a second response to a second training prompt comprising aggregate information corresponding to multiple of the plurality of individual … stores.”
In an analogous art, Mosallanezhad teaches “The method of claim 1, wherein training the machine learning model comprises: penalizing the machine learning model based on a first response to a first training prompt comprising sensitive information …”
See Mosallanezhad in [0045] describe “Where l indicates label. For each private-attribute attacker pk, the confidence score Cpk is within the range [−1,1]. A positive value demonstrates that the attacker has predicted private-attribute accurately, and a negative value indicates that the attacker was not able to infer user's private-attribute. According to this definition, the reward will be positive if action at has caused information hiding and will be negative if the action at was not able to hide sensitive information. Having confidence of private-attribute inference attackers, reward function at state st+1 is defined as:
PNG
media_image4.png
54
528
media_image4.png
Greyscale
”
Here, Mosallanezhad describes punishing the model from not being able to hide sensitive information,
Also, see Mosallanezhad mention in [0047] "Classification is one of the common tasks. In order to preserve the utility of data, it is necessary to ensure that preserving privacy of data does not destroy the semantic meaning of the text embedding representation 112 with respect to the given task. This challenge is approached by changing the agent's reward function with respect to the data utility. The utility checker component, i.e., utility classifier component DU, 144 is added to the RL environment 140 to assess the quality of resultant embedding representation. The confidence of the utility classifier component 144 is used for the given task to measure the utility of the text embedding representation 112 using the text embedding representation vector vu the its correct label i.
PNG
media_image5.png
60
515
media_image5.png
Greyscale
" . Here, Mosallanezhad mentions using text embedding vector to see if this information represents the same meaning of the original sensitive data.
Also, see Mosallanezhad in [0029] note “Blocks 220 and 230 are iteratively repeated until a resultant final text embedding representation 112 that maximizes the reward value is achieved.” Here, Mosallanezhad shows that the whole method is repeated including training steps, see algorithm 1 in paragraphs [0053-0054] show multiple training iterations (i.e. relate to having a second, third and subsequent training prompts or inputs).
PNG
media_image6.png
633
946
media_image6.png
Greyscale
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the references of Jeong and Feyisetan of using a model in producing a response to the prompt from a vector store, with Mosallanezhad’s teaching of reward if aggregate information and punishment if has sensitive data.
One of ordinary skill in the art would be motivated to do so because by integrating Mosallanezhad’s framework into the methods of Jeong and Feyisetan, one with ordinary skill in the art would achieve "RLTA-GEN performs better in terms of utility, in comparison to the unaltered present system 100 which incorporates both gender and location attackers. Moreover, results show that both RLTA-GEN and RLTA-LOC have better utility than other baselines," (See [0082] from Mosallanezhad).
However, Jeong in view of Feyisetan, further in view of Mosallanezhad, did not teach “rewarding the machine learning model based on a second response to a second training prompt comprising aggregate information corresponding to multiple of the plurality of individual … stores.”
In an analogous art, Arora teaches “rewarding the machine learning model based on a second response to a second training prompt comprising aggregate information corresponding to multiple of the plurality of individual … stores,”
See Arora in page 909, mention in section Baselines “We evaluate using four retrieval baselines: (1) CONCURRENTQA-MDR, a dense retriever trained on the CONCURRENTQA train set (15.2k examples), to understand the value of in-domain training data for the task.” Here, Arora shows using various training examples around 15.2k, and any one of these can be chosen to be a first, second or subsequent training prompt.
See Arora in page 907, section 4.2 ConcurrentQA overview, note “Example questions from ConcurrentQA are included in Table 1. The corpora contain 47k emails (DP) and 5.2M Wikipedia passages (DG), and the benchmark contains 18,439 examples (Table 2). Questions require three main reasoning patterns: (1) bridge questions require identifying an entity or fact in Hop1 on which the second retrieval is dependent”. Here, Arora shows using a second retrieval, or a second response to the model ConcurrentQA’s question or prompt. Each entity relates to individual data or store. See Table 1 on page 907, where Arora illustrates examples of responses and prompts:
PNG
media_image7.png
347
981
media_image7.png
Greyscale
See Arora mention using multiple models in page 903 , note “Privacy-Preserving Retrieval, We then propose a framework for reasoning about the privacy tradeoffs required for SoTA models to achieve as good performance on public-private QA as is achieved in public-QA. We evaluate performance when no private information is revealed, and models trained only on public data (e.g., Wikipedia) are utilized. Under this privacy standard, models sacrifice upwards of 19% performance under SPIRAL constraints to protect document privacy and 57% to protect query privacy.” Here, Arora shows models achieve good performance, which is viewed as similar as rewarded when private information is protected. This applies to a first, second or subsequent responses to a prompt.
See Arora mention in page 913, section 7. Conclusion, “quantitatively demonstrated the effects of applying models trained on one distribution (e.g., public) to the mixed-distribution (e.g., public and private) setting in Sections 5 and 6. Private iterative retrieval is underexplored and we hope the benchmark-resource and evaluations we provide inspire further research on this topic, for instance under alternate privacy models.” When Arora mentions using alternate models, Arora shows using more than one model, and this relates to using a second or subsequent model.
Also, see Arora in page 905, section 2.3 Privacy preserving retrieval, note “Partial query privacy via fake query augmentation for high-latency retrieval from public databases. Another class of privacy techniques for hiding the user’s intentions is query-obfuscation or k-anonymity. The user’s query is combined with fake queries or queries from other users to increase the difficulty of linking a particular query to the user’s true intentions (Gervais et al., 2014).” Here, Arora mentions combining the information with queries to increase difficulty of connecting a query to a user’s purpose relate to aggregating information to obscure sensitive or private data.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the references of Jeong, Feyisetan, and Mosallanezhad, of using a model in producing a response to the prompt from a vector store, with Arora’s teaching of reward if aggregate information and punishment if has sensitive data.
One of ordinary skill in the art would be motivated to do so because by integrating Arora’s framework into the methods of Jeong, Feyisetan, and Mosallanezhad, one with ordinary skill in the art would achieve " Further, we find the performance of the model trained on Wikipedia improves by 4.3% if we retrieve the top k/2 passages from each distribution vs. retrieving the over all top k passages," (see Arora in page 903).
Claim 12:
Regarding claim 12, it comprises of similar additional limitations as corresponding dependent claim 2, and is rejected under the same rationale under 35 U.S.C. 103.
Claims 4, 5, 6, 7 and 14, 15, 16, and 17 are rejected under 35 U.S.C. 103 as being unpatentable over Jeong in view of Feyisetan, and further in view of Steedman Henderson, M., in US PG Pub. No. US20210141798A1, published on May 13, 2021 , (hereafter, Henderson), and further in view of Sharifi, M. et al., in US PG Pub. No. US20220335932A1, published on October 20, 2022, (hereafter, Sharifi).
Claim 4:
Regarding claim 4, Jeong in view of Feyisetan, teach the limitations of claim 3. Further, Jeong teaches “comparing the query vector with the one or more vector embeddings,”
See page 14 of Jeong, describe in section 2.3.3 Vector Database note "A vector database is a novel type of database developed to address the problem of long-term memory deficiency in LLMs. This database is specialized in efficiently storing and managing high-dimensional real-valued vector indices. Unlike traditional databases, a vector database represents queries in the form of real-valued vectors (embeddings) and supports a method to extract data with similar vectors." Here, Jeong describes a vector database that stores vector embeddings, and each vector is indexed shows that each vector is embedded or encoded in a form that is easily retrieved with number representations. The indexed vectors here mean the same as the embeddings as Jeong indicated.
See page 14, where Jeong describes in step 2. "Querying: In the vector database, indexed query vectors are compared to indexed vectors in the dataset to find the nearest neighbors. The similarity metric used in the indexing is applied during this process." Here, Jeong mentions the query vectors are compared to indexed vectors, which from section 2.3.3 the index vectors represent vector embeddings. This comparison will find the nearest neighbors, which means searching for the most similar content in the data. See Jeong in page 18 , step 4, describe for details.
Further, Jeong teaches “identify one or more vector similar vector embeddings from the vector store;” and “wherein the one or more similar vector embeddings having a similarity to the query vector … of similarity;”
Vector store is construed to be any location where vector embeddings can be kept, such as a database or repository. See Jeong in page 12, section 2.3.2 Using embedding in LLM note " In the context of LLM (Language Model), embedding, or embedding vectors, refer to the representation of text as fixed-size arrays of floating-point numbers. These arrays form real-valued vector shapes. When a specific word, sentence, or document is input into an embedding generation model, it outputs a vector composed of floating-point values. Although these vectors might be hard for humans to directly comprehend, calculating the distances between embeddings of different words or documents can reveal semantic relationships... During this training process, the model takes in various words as input and learns semantic relationships by making embedding vectors of contextually similar words closer together and those of dissimilar words farther apart. This approach allows the model to capture meaningful relationships within the context." Here, Jeong shows identifying the embedded vectors (i.e. vector embeddings) of similar words closer together as being similar. Jeong shows identifying embeddings with similarities. See Jeong in page 13 and figure 9 (on page 17) for more information.
Also, see Jeong in page 21 section 4.2.5 Implementation of question(Prompt) and search result integration steps mention "In this step, we utilize similarity search to retrieve relevant chunks for each query. When a user submits a query, the first task is to convert it into an embedding to find the most relevant chunks within the Chroma DB. The code to achieve this using LangChain is presented in Figure 14. For instance, if a user asks, "What is the work dress code for male employees?" the vector database containing the 'dress code standards.pdf' document will be queried to find chunks with content similar to the query." Here, Jeong shows the method of using similarity search to find the most similar vector embeddings from a vector database for a specific subject the user asks from a query. Jeong applies this similarity to a query (which also contain information via chunks such as vectors, see Jeong in page 11 section 2.3.1 Architecture of RAG model note “Figure 5 illustrates the data processing pipeline for RAG usage, involving breaking down the original data into smaller chunks and converting text data into numerical vectors through embedding, which are then stored in a vector repository.”) See figure 9 for details.
PNG
media_image8.png
505
1017
media_image8.png
Greyscale
However, Jeong in view of Feyisetan did not teach “4. The method of claim 3, further comprising: generating a query vector based at least in part on the prompt; wherein the one or more similar vector embeddings having a similarity to the query vector that meets or exceeds a predetermined threshold of similarity; and wherein the response is generated based at least in part on the one or more similar vector embeddings”
In an analogous art, Henderson teaches “4. The method of claim 3, further comprising: generating a query vector based at least in part on the prompt;”
See Hendersen in paragraph [0079] mention "In this example, in step S401, the user enters the phrase “Is that place affordable?” In step S402 and S403, the phrase is encoded using a trained encoder 206 to produce a context vector hX." Here, Hendersen shows that the context vector represents the query vector, and relates to the prompt asked by the user here in step S401.
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to combine the references of Jeong and Feyisetan, and incorporate into the teachings of Henderson because all references teach using a model to produce a response to the prompt using vector data.
One of ordinary skill in the art would be motivated to do so because this would achieve “The model is therefore memory efficient and training efficient, while maintaining performance in a response selection task,” (see Henderson in [0030]).
However, Jeong in view of Feyisetan, further in view of Henderson, did not teach “wherein the one or more similar vector embeddings having a similarity to the query vector that meets or exceeds a predetermined threshold of similarity; and wherein the response is generated based at least in part on the one or more similar vector embeddings.”
In an analogous art, Sharifi teaches “wherein the one or more similar vector embeddings having a similarity to the query vector that meets or exceeds a predetermined threshold of similarity;”
See Sharifi in paragraph [0057] mention "When two or more responses are compared and determined to be within a threshold similarity, assistant coordination module 325 may cluster responses. Referring to FIG. 4B, a first cluster 425 includes Response 1 and Response 2, and second cluster 430 includes Response 3. For the query “What is the weather,” Response 1 may be “The current temperature is 55 degrees” and Response 2 may be “The temperature is 55 degrees,” both of which may have vector representations that are close in vector space (i.e., include similar information)." Here, Sharifi mentions using a threshold similarity to compare two or more responses for similarity in vector representations are within that threshold (i.e. meet a threshold of similarity).
Further, see Sharifi in [0024] note "In some implementations, the primary automated assistant can combine responses only when the confidence score of a cluster exceeds the threshold of one or more of the individual responses. For example, a first response can have a confidence score of 0.95 and a cluster of responses may have a determined confidence score of 0.94." Here, Sharifi shows examples of preset values for a threshold.
Further, Sharifi teaches “and wherein the response is generated based at least in part on the one or more similar vector embeddings.”
See Sharifi in paragraph [0056] mention " In various implementations, responses can be processed to identify a semantic property that is included in a given response. Further, an embedding (e.g., a word2vec representation) of the identified semantic property can be generated for a response and compared to embeddings of the semantic labels associated with the other responses (e.g., respective embedding representations). Further, it may be determined that the semantic property matches a given embedding, of the plurality of embeddings of the respective semantic labels of other responses, based on the comparison. For example, assume the embeddings are word2vec representations. In this example, a cosine distance between the word2vec representation of the semantic property and each of the word2vec representations of the respective semantic labels can be determined, and a given semantic label that is associated with a respective cosine distance that satisfies a distance threshold can be utilized to determine the semantic property of the spoken utterance matches the given semantic label that is, for example an exact match or soft match." Here, Sharifi shows the embeddings satisfying a distance threshold of a cosine similarity metric,
Also, see Sharifi in [0057] note "When two or more responses are compared and determined to be within a threshold similarity, assistant coordination module 325 may cluster responses. Referring to FIG. 4B, a first cluster 425 includes Response 1 and Response 2, and second cluster 430 includes Response 3. For the query “What is the weather,” Response 1 may be “The current temperature is 55 degrees” and Response 2 may be “The temperature is 55 degrees,” both of which may have vector representations that are close in vector space (i.e., include similar information)." Here, Sharifi shows the response that is produced based on similarity between a given embedding's semantic labels.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the references of Jeong, Feyisetan, and Henderson, of using a model in producing a response to the prompt from a vector store, with Sharifi’s teaching of using a similarity to the query vector that meets or exceeds a predetermined threshold of similarity.
One of ordinary skill in the art would be motivated to do so because by integrating Sharifi’s framework into the methods of Jeong, Feyisetan, and Henderson, one with ordinary skill in the art would achieve "utilization of techniques described herein mitigates the need of a user to explicitly invoke each of multiple automated assistants to determine, from each of the automated assistants, a response to a query that the user has interest in being provided. This can make the overall duration of the user/assistant(s) interaction more efficient." (see Sharifi in [0007]).
Claim 5:
Regarding claim 5, Jeong in view of Feyisetan, further in view of Henderson, and further in view of Sharifi, teach the limitations of claim 4.
Further, Henderson teaches “5. The method of claim 4, further comprising determining the similarity of the one or more similar vector embeddings to the query vector based at least in part on a cosine similarity of the one or more vector embeddings to the query vector,”
See Henderson mention in paragraphs [0081-0082] “ The encoder is trained to output a vector (a context vector hX) for an input query. The training has been performed using queries and corresponding responses. The encoder 206 has been trained so that the encoder 206 for a query produces a context vector hX which is very similar to a response vector hY for a suitable response. The similarity between the two vectors can be determined using a similarity measure such as the cosine similarity (which is obtained as (kX·hY)/(|hX∥hY|). The R possible response vectors hY have been generated off-line and reside in a database. Therefore, once the phrase is encoded using the trained encoder and the hX context vector is produced, the next stage S404 is to look for a similar response vector for this context vector in the response vector database 209.” Here, Henderson mentions that determining similarity occurs between the context vector (i.e. query vector) and possible response vectors (i.e. other vector embeddings see [0061]) by using a cosine similarity measure.
Also, see Henderson in [0061] describe "The response selection model 109 is stored in the storage 107. For example, the response vectors and their associated responses, and various parameters of the response selection model 109 such as the embeddings and weights and bias vectors in the encoder may be stored in the storage 107." Here, Henderson shows that the response vectors contain embeddings which are vector embeddings.
It would have been obvious for one of ordinary skill in the art before the effective filing date of the claimed invention to combine the references of Jeong and Feyisetan, and incorporate into the teachings of Henderson because the references teach using a model to produce a response to the prompt using vector data based at least in part on a cosine similarity of the one or more vector embeddings to the query vector.
One of ordinary skill in the art would be motivated to do so because this would achieve “The model is therefore memory efficient and training efficient, while maintaining performance in a response selection task,” (see Henderson in [0030]).
Claim 6:
Regarding claim 6, Jeong in view of Feyisetan, further in view of Henderson, and further in view of Sharifi, teach the limitations of claim 4.
Further, Jeong teaches “The method of claim 4, further comprising: decoding the one or more similar vector embeddings to obtain data related to the prompt;
See Jeong in page 13, section Question Answering, mention "Therefore, to obtain answers to questions related to private information, the text containing that information needs to be included in the prompt to LLM. However, current LLM services have limitations on the length of input text, which requires breaking down long texts into smaller chunks. In cases where longer texts need to be processed, relevant chunks are selected and included along with the question in the prompt." Here, Jeong shows that to retrieve responses to user questions (i.e. prompts) about private data, the prompt needs to include text that has the sensitive information within the prompt (i.e. obtain data related to the prompt) for the model to read. See Jeong illustrate in figure 7 for details.
PNG
media_image9.png
530
1336
media_image9.png
Greyscale
See Jeong in pages 11-12 mentioned in section 2.3.1 Architecture of the RAG Model "The RAG (Retrieval-Augmented Generation) model is designed for text generation tasks, performing a process that involves retrieving information from given source data and utilizing that information to generate desired text. Figure 5 illustrates the data processing pipeline for RAG usage, involving breaking down the original data into smaller chunks and converting text data into numerical vectors through embedding,
Chunking of Searchable Units: Source data is divided into smaller units known as chunks. Chunks are typically small text fragments, such as sentences or paragraphs, making it easier to search for and utilize information at this granular level.
• Embedding: Generated chunks undergo embedding, a process of converting text into meaningful vector representations. Pre-trained language models are often used to transform text into dense vectors, capturing the meaning and related information in vector form.
• Construction of Vector Database: A vector database is built based on the embedded chunks. This database represents the positions of each chunk within the vector space, enabling efficient retrieval and similarity calculations.
• Search and Information Integration: To retrieve information relevant to the context of the text to be generated, appropriate chunks are searched within the vector database. The retrieved chunks are decoded back into original text data to extract information," Here, Jeong mentions decoding the relevant vector embedding representations, where Jeong calls these chunks, back into original text data to extract information. The relevant information from those vector embeddings are viewed as similar vector embeddings related to the prompt.
Further, Jeong teaches "and wherein the response is generated based at least in part on the data related to the prompt."
See Jeong in page 18, section 3.2.1 RAG based implementation procedure mention "The RAG model, as discussed in the architecture of the RAG model in Chapter 2, is a search-augmented generative model used to retrieve and generate responses based on information relevant to given questions or topics... 5) Integration of Prompt and Search Results: This step involves searching for information based on the prompted question and integrating relevant information. To search for contextually relevant information based on the prompt, appropriate chunks are retrieved from the vector database. These retrieved chunks are then sent to LLM to aid in the response generation process. 6) Answer Generation: Using the retrieved information as a basis, the response text is generated." Here, Jeong explicitly shows producing a response based on 'searching for information based on the prompted question' (i.e. data related to the prompt).
Claim 7:
Regarding claim 7, Jeong in view of Feyisetan, further in view of Henderson, and further in view of Sharifi, teach the limitations of claim 6.
Further, Jeong teaches “7. The method of claim 6, wherein the data related to the prompt comprises a context associated with the one or more similar vector embeddings,”
See Jeong in page 18, section 3.2.1 RAG based implementation procedure, mention "3) Embedding: The generated chunk-level text data is transformed into numerical vector representations. This step involves mapping words or sentences to vectors, ... 5) Integration of Prompt and Search Results: This step involves searching for information based on the prompted question and integrating relevant information. To search for contextually relevant information based on the prompt, appropriate chunks are retrieved from the vector database. These retrieved chunks are then sent to LLM to aid in the response generation process. Various search engines available within LangChain for vector store similarity search are utilized. 6) Answer Generation: Using the retrieved information as a basis, the response text is generated. At this stage, the type, length, and linguistic style of the generated text can be specified. LLM, such as OpenAI's GPT-3.5-turbo model or GPT4All, uses the similarity search module in LangChain to retrieve relevant documents and generate responses. " Here, Jeong mentions while searching for contextually relevant information based on the prompt (i.e. data related to the prompt ), appropriate vector representations or vector embeddings are retrieved (here, appropriate vector embeddings also means finding similar vector embeddings).
Further, see Jeong in page 21 section 4.2.5 Implementation of question(Prompt) and search result integration steps mention "In this step, we utilize similarity search to retrieve relevant chunks for each query. When a user submits a query, the first task is to convert it into an embedding to find the most relevant chunks within the Chroma DB. The code to achieve this using LangChain is presented in Figure 14. For instance, if a user asks, "What is the work dress code for male employees?" the vector database containing the 'dress code standards.pdf' document will be queried to find chunks with content similar to the query." Here, Jeong describes data related to the user's question (i.e. prompt), where this information is used to find other content similar to the user's original question.
Further, see Jeong in pages 9-10, section 2.2.3 Prompt engineering mention "Prompt engineering is important because it seeks to improve the quality of answers without the need for extensive parameter updates through large-scale data or fine-tuning processes. One effective approach to enhance answer quality is by providing example answer instances when prompting questions. This method guides the model to generate responses that are similar to the provided examples, thereby improving response quality...Prompts serve as instructions and guidelines for LLMs, varying from simple to complex, and can specify the format, content, and style of the text LLMs are expected to produce." Here, Jeong mentions using a prompt to provide information and context.
Also, see page 14 of Jeong, describe in section 2.3.3 Vector Database note "A vector database is a novel type of database developed to address the problem of long-term memory deficiency in LLMs. This database is specialized in efficiently storing and managing high-dimensional real-valued vector indices. Unlike traditional databases, a vector database represents queries in the form of real-valued vectors (embeddings) and supports a method to extract data with similar vectors." Here, Jeong describes a vector database that stores vector embeddings, and each vector is indexed shows that each vector is embedded or encoded in a form that is easily retrieved. Here, Jeong also mentions the process of extracting data with similar vector embeddings for the LLM model that processes the prompts, from page 11.
Also, see Jeong in page 11, section 2.3 Overview of the RAG model, note "strategies have been proposed, such as fine-tuning LLMs with new data or directly injecting information into the prompt context. However, the former approach incurs significant costs, while the latter approach of embedding all information into prompts is not practical. As an alternative, the RAG (Retrieval-Augmented Generation) model has emerged. This model stores information in databases and retrieves the necessary information to provide it to the LLM when needed. By providing LLMs with relevant questions and associated reference materials beforehand, the model utilizes these references to generate more accurate and reliable answers. In this section, we delve into the details of the RAG model, which is employed as an implementation method in this paper. The RAG model stores information, leverages databases, and enhances the generation process through the inclusion of relevant reference materials". Here, Jeong shows data related to the prompt. Jeong also uses LLM models to extract data or context from the prompt that is connected to information from vector embeddings mentioned in page 14 in section 2.3.3. Vector Database.
Claim 14:
Regarding claim 14, it comprises of similar additional limitations as corresponding dependent claim 4, and is rejected under the same rationale under 35 U.S.C. 103.
Claim 15:
Regarding claim 15, it comprises of similar additional limitations as corresponding dependent claim 5, and is rejected under the same rationale under 35 U.S.C. 103.
Claim 16:
Regarding claim 16, it comprises of similar additional limitations as corresponding dependent claim 6, and is rejected under the same rationale under 35 U.S.C. 103.
Claim 17:
Regarding claim 17, it comprises of similar additional limitations as corresponding dependent claim 7, and is rejected under the same rationale under 35 U.S.C. 103.
Claims 8 and 18 are rejected under 35 U.S.C. 103 as being unpatentable over Jeong in view of Feyisetan, and further in view of Gupta S., et al., in US PG Pub. No. US20230080686-A1, published on March 16, 2023, (hereafter, Gupta).
Claim 8:
Regarding claim 8, Jeong in view of Feyisetan, further in view of Sharifi, teach the limitations of claim 7. Further, Jeong teaches “8. The method of claim 7, wherein the machine learning model is a first machine learning model, … whether the response comprises … information corresponding to an individual one of the plurality of vector stores.”
See Jeong in page 4 mention “The core aspects of the RAG model—information retrieval and utilization—are exhaustively investigated. The process of embedding document information using vectorized databases, its storage, and subsequent retrieval is elaborated. An innovative method is introduced to convert retrieval outcomes into prompts for LLM through an orchestration framework, facilitating the effective synthesis of user queries and search results to generate optimal responses.” Here, Jeong discusses using a model with information from vectorized databases (i.e. which relate to data from the vector store) to produce responses.
However, Jeong in view of Feyisetan, further in view of Sharifi, did not teach “… the method further comprising determining, using a second machine learning model, whether the response comprises sensitive information …”
In an analogous art, Gupta teaches “… the method further comprising determining, using a second machine learning model, whether the response comprises sensitive information …”
See Gupta in paragraph [0071] mention “Referring to FIG. 8 , a visual representation of the rules engine model (also referred to herein as the second model). The second model may use various predefined rules to determine whether a column included PII. The flow chart 800 illustrates how the second model can be utilized to generate/predict a second score indicative of a column including PII.” Here, Gupta shows using a second model specifically to identify if a column (construed to mean information or data) contains private or sensitive information (here abbreviated as PII or personally identifiable information).Also, see Gupta in [0038] note for more details.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the references of Jeong and Feyisetan of using a model in producing a response to the prompt from a vector store, and incorporate with Gupta’s teaching of using a second model to detect private information.
One of ordinary skill in the art would be motivated to do so because by integrating Gupta’s framework into the methods of Jeong and Feyisetan, one with ordinary skill in the art would achieve "This multi-layered process helps identify misclassified PII data column(s) or assign the PII/No-PII sensitivity to a new column," (see Gupta in [0005]).
Claim 18:
Regarding claim 18, it comprises of similar additional limitations as corresponding dependent claim 8, and is rejected under the same rationale under 35 U.S.C. 103.
Claims 9 and 19 are rejected under 35 U.S.C. 103 as being unpatentable over Jeong in view of Feyisetan, and further in view of Sharifi, and further in view of Ghorab, M. et al., in “Personalised Information Retrieval: survey and classification,” published on May 13, 2012, available at: https://link.springer.com/article/10.1007/s11257-012-9124-1#citeas , (hereafter, Ghorab).
Claim 9:
Regarding claim 9, Jeong in view of Feyisetan, teach the limitations of claim 1.
Further, Jeong teaches “9. The method of claim 1, further comprising: generating a plurality of individualized responses to the prompt, each of the plurality of individualized responses corresponding to one of the plurality of vector stores;”
See Jeong in page 16 note in section 2.3.5 AI Chatbot, “Through AI chatbots like ChatGPT, generative AI services can be harnessed, allowing users to input prompts for desired questions and receive corresponding responses. Today, AI-driven chatbot systems encompass not only simple question-answer scenarios like FAQs, but have advanced to the level of analyzing human emotions and intentions to provide tailored responses (Jeong, C.S., 2023a).” Here, Jeong mentions producing tailored or individualized responses to input prompts. Each response relate to a corresponding vector database or vector store.
See Jeong in page 18 section 3.2.1 RAG based implementation procedure “The RAG model, as discussed in the architecture of the RAG model in Chapter 2, is a search-augmented generative model used to retrieve and generate responses based on information relevant to given questions or topics. Each step follows the procedure outlined in Figure 9. 1) Source Data Collection and Extraction: During the data collection phase, both structured and unstructured data are gathered. Structured data is stored in standardized formats such as CSV, JSON, or XML, while unstructured data is stored in formats like PDF, TXT, HTML, images, and videos. Preparatory materials related to the task, such as regulations, user manuals, and terms and conditions, are loaded into LangChain using the LangChain module. 2) Chunk Generation: Source data is processed to split it into smaller units known as chunks. These chunks typically consist of sentences or paragraphs, serving as smaller text fragments that can be used to search and retrieve information from LLM. LangChain's module is utilized to split data into chunks that are suitable for retrieval.
3) Embedding: The generated chunk-level text data is transformed into numerical vector representations. This step involves mapping words or sentences to vectors, and libraries provided by OpenAI or GPT4All can be employed for this purpose.
4) Building the Vector Database: Based on the embedded chunks, the vector database is constructed. This database represents the positions of each chunk in the vector space, facilitating efficient search and similarity calculations. Typically, content from each document is included, embeddings and documents are stored in the vector repository, and documents are indexed using embeddings. Tools like Chroma or FAISS for vector indexing can be used.
5) Integration of Prompt and Search Results: This step involves searching for information based on the prompted question and integrating relevant information. To search for contextually relevant information based on the prompt, appropriate chunks are retrieved from the vector database. These retrieved chunks are then sent to LLM to aid in the response generation process. Various search engines available within LangChain for vector store similarity search are utilized. 6) Answer Generation: Using the retrieved information as a basis, the response text is generated. At this stage, the type, length, and linguistic style of the generated text can be specified. LLM, such as OpenAI's GPT-3.5-turbo model or GPT4All, uses the similarity search module in LangChain to retrieve relevant documents and generate responses.” Here, Jeong shows retrieving appropriate information from the vector database, and relates to information such as responses retrieved from using the relevant information from the corresponding vector database from page 16, section 2.3.5 AI Chatbot.
See Jeong in figures 6 and 9 for details.
PNG
media_image1.png
564
1120
media_image1.png
Greyscale
PNG
media_image2.png
562
1013
media_image2.png
Greyscale
Further, Jeong teaches “and regenerating, using the machine learning model, the response to the prompt based at least in part on the vector store,”
See Jeong mention in page 6 note in section 2.2.2 LLM (Large Language Model) “Generative AI models, particularly Large Language Models (LLMs), are widely used in chatbot services like ChatGPT. These models are trained on language data, such as text, and provide results through generating text-based responses.”
Also, see Jeong in page 9 mention in section 2.2.3 Prompt Engineering in ChatGPT, “the quality of the generated responses greatly depends on how detailed and specific the prompts (questions or requests) are conveyed. This is why the exploration of prompt engineering, which involves finding combinations of prompt input values from LLM, plays a crucial role. Prompt engineering is important because it seeks to improve the quality of answers without the need for extensive parameter updates through large-scale data or fine-tuning processes. One effective approach to enhance answer quality is by providing example answer instances when prompting questions. This method guides the model to generate responses that are similar to the provided examples,” and later see Jeong in page 16 section 2 .3.5 AI Chatbot Through AI chatbots like ChatGPT, note “generative AI services can be harnessed, allowing users to input prompts for desired questions and receive corresponding responses. Today, AI-driven chatbot systems encompass not only simple question-answer scenarios like FAQs, but have advanced to the level of analyzing human emotions and intentions to provide tailored responses.” Here, Jeong shows generating more than one responses and uses fine-tuning to update (construed to mean similar to the term ‘regenerate’ ) responses with updated relevant information.
However, Jeong in view of Feyisetan, did not teach “determining a similarity between the response and each of the plurality of individualized responses; determining to mask the response based at least in part on the response having a similarity to fewer than a predetermined number of the plurality of individualized responses;”
In an analogous art, Sharifi teaches “determining a similarity between the response and each of the plurality of individualized responses;
See Sharifi in paragraph [0020] note "Each response can be embedded in a vector space (e.g., using “word2vec” or other techniques) based on processing the information, that is included in the response, using one or more embedding machine learning models (e.g., a trained encoder neural network). Once the responses are represented as vectors, the primary automated assistant can identify similar responses based on similarity of the vectors in the vector space...For example, a first automated assistant can determine a response of “The current temperature is 55 degrees” and a second assistant may determine a response of “The weather right now is sunny and around 55 degrees.” Because the two responses include similar information (e.g., a current temperature), the vectors representing the two responses will likely be located near each other in vector space and therefore include similar information. "
Also, see Sharifi in paragraph [0057] mention "When two or more responses are compared and determined to be within a threshold similarity, assistant coordination module 325 may cluster responses. Referring to FIG. 4B, a first cluster 425 includes Response 1 and Response 2, and second cluster 430 includes Response 3. For the query “What is the weather,” Response 1 may be “The current temperature is 55 degrees” and Response 2 may be “The temperature is 55 degrees,” both of which may have vector representations that are close in vector space (i.e., include similar information)." Here, Sharifi shows determining a similarity between two different responses of response 1 and response 2. Since each response is different, each response is also considered individualized. Therefore, Sharifi produces a similarity measure between a response and for each of the individualized responses. Sharifi also shows this similarity is applied to vector embeddings. Here, Sharifi also mentions using a threshold similarity to compare two or more responses for similarity in vector representations are within that threshold.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the references of Jeong and Feyisetan of using a model in producing a response to the prompt from a vector store, with Sharifi’s teaching of using a similarity between responses.
One of ordinary skill in the art would be motivated to do so because by integrating Sharifi’s framework into the methods of Jeong and Feyisetan, one with ordinary skill in the art would achieve "Thus, utilization of techniques described herein mitigates the need of a user to explicitly invoke each of multiple automated assistants to determine, from each of the automated assistants, a response to a query that the user has interest in being provided. This can make the overall duration of the user/assistant(s) interaction more efficient." (see Sharifi in [0007]).
However, Jeong in view of Feyisetan, and further in view of Sharifi did not teach “determining to mask the response based at least in part on the response having a similarity to fewer than a predetermined number of the plurality of individualized responses;”
In an analogous art, Ghorab teaches “determining to mask the response based at least in part on the response having a similarity to fewer than a predetermined number of the plurality of individualized responses;”
See Ghorab in page 415, section 4.2.3.2 Result adaptation mention “The other common approach to search personalisation is result adaptation. Adaptation of result lists can be performed by result scoring, result re-ranking, or result filtering. Result re-ranking takes place after an initial set of documents have been retrieved by the system, where an additional ranking round is performed to re-order documents based on certain adaptation aspects (e.g. displaying certain documents at higher ranks in the result list based on the user’s interests). Result filtering can be considered as a special case (or a step further) of result re-ranking, where after the result list is sorted in descending order of relevance scores, results that fall below a certain threshold are not displayed to the user. Result scoring involves incorporating adaptation features directly in the primary scoring function of the retrieval component of the system … After the concepts of the snippets have been deduced, they are compared to the concepts in the user model using cosine similarity. The results are then re-ranked in descending order of the conceptual similarity score.” Here, Ghorab shows if results are below a certain threshold (inherently preset), then they are not displayed (similar to masked) from the user.
See page 396 where Ghorab notes “Furthermore, the system also has a threshold for the number of terms stored in a vector (i.e. a maximum of M interest terms per cluster). If the threshold of N vectors was reached, and a new vector comes in, then all the vectors in the user model, in addition to the incoming vector, are textually compared to each other using cosine similarity (Baeza-Yates and Ribeiro-Neto 2011). The two most similar vectors are then combined together in one vector. This is done by grouping together the terms from the two vectors, sorted in descending order of weights, and then keeping only the top M terms.”
See also page 399, where Ghorab notes “This challenge was overcome by textually comparing such queries with all other queries in the logs (using cosine similarity) to find ones that can be deemed similar.” Ghorab uses cosine similarity to compare the query results.
Further, see page 420, where Ghorab mentions “After the categories are identified, the query is actually submitted multiple times for retrieval; once without specifying any categories, and one time for each of the identified categories. This leads to the retrieval of multiple result lists. The system then performs result adaptation by ranking and merging the results into a single list. A weighted voting based algorithm was used, where results that appeared on more than one list were favoured.” Here, Ghorab shows that query is submitted to produce results (i.e. viewed as responses to the user’s query), and from page 399 shows if the results that are below a threshold for similarity (mentioned from page 396 as cosine similarity), then those results are not displayed (i.e. masked) to the user. For more details, see page 419, where Ghorab notes in section 4.2.3.3. Query adaptation and result adaptation.
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to combine the references of Jeong, Feyisetan, and Sharifi, of using a model in producing a response to the prompt from a vector store, with Ghorab’s teaching of masking sensitive data.
One of ordinary skill in the art would be motivated to do so because by integrating Ghorab’s framework into the methods of Jeong, Feyisetan, and Sharifi, one with ordinary skill in the art would achieve “Personalised Information Retrieval (PIR) systems take a step further to better satisfy the user’s specific information needs by providing search results that are not only of relevance to the query but are also of particular relevance to the user who submitted the query. PIR has thereby attracted increasing research and commercial attention as information portals aim at achieving user loyalty by improving their performance in terms of effectiveness and user satisfaction,” (See Ghorab, in page 381 in abstract).
Claim 19:
Regarding claim 19, it comprises of similar additional limitations as corresponding dependent claim 9, and is rejected under the same rationale under 35 U.S.C. 103.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to WENWEI ZENG whose telephone number is (571)272-7111. The examiner can normally be reached Monday-Friday, 8am-5pm.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Usmaan Saeed can be reached at (571) 272-4046. 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.
/WenWei Zeng/Examiner, Art Unit 2146
/USMAAN SAEED/Supervisory Patent Examiner, Art Unit 2146