Computer-implemented method for searching in a data inventory for a data set corresponding to a search data set

CN122935768APending Publication Date: 2026-10-16ABP SERVICES LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202580021224.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2024-05-17
Filing Date
2025-03-11
Publication Date
2026-10-16

AI Technical Summary

Technical Problem

[0002]为了将如长文本、图像、视频或设计图纸之类的复杂数据集与搜索数据集或初始文档进行比较,需要非常大量时间

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122935768A_ABST
    Figure CN122935768A_ABST
Patent Text Reader

Abstract

The invention relates to a computer-implemented method for searching a data stock for data sets corresponding to a search data set, comprising the following method steps: conducting a search from a search data set, which comprises search data set portions, wherein the search data set portions are each semantically, graphically or technically associated in themselves; assigning vector embeddings to the search data set and / or the search data set portions; assigning vector embeddings in the data stock to the data sets and / or data set portions stored therein; ascertaining and evaluating by means of two correlation values and sorting and data-technically processing the individual data set portions according to the second correlation value as a result.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to a computer-implemented method for searching a dataset corresponding to a search dataset in a data store. Background Technology

[0002] Comparing complex datasets such as long texts, images, videos, or design drawings with search datasets or initial documents requires a significant amount of time. However, the problem lies in the fact that, overall, too many documents are being searched or identified as relevant by AI-powered searches, resulting in an excessive number of hits. This leads to an excessive load on computational time and infrastructure. Consequently, the relevant dataset is often hidden within the sheer volume of the dataset.

[0003] Existing technologies have extensively attempted to make long and complex texts comparable by using computer-aided methods such as classification and keyword tagging. A key challenge lies in the fact that public content within a dataset can be contained in different locations. These various public locations within the dataset can have varying degrees of relevance to the search. Summary of the Invention

[0004] The objective of this invention is to overcome the shortcomings of the prior art and provide a method by which a dataset can be compared with a search dataset in a low-computational-power, simple-infrastructure, and resource-efficient manner.

[0005] This task is solved by the method claimed in the claims.

[0006] A computer-implemented method according to the present invention is used to search for a dataset corresponding to a search dataset in a data store. The method includes the following steps: performing a search starting from a search dataset, the search dataset including search dataset portions, wherein the search dataset portions are semantically, graphically, or technically related to each other; embedding vectors in the search dataset and / or search dataset portions; embedding vectors in the data store and / or dataset portions stored therein; determining a first relevance based on comparisons with the search dataset and evaluating it using a first relevance value of the dataset; evaluating the dataset using the first relevance value based on comparisons with the search dataset; comparing each dataset portion of the dataset having a first relevance value exceeding a first relevance threshold, wherein the comparisons are performed individually with the search dataset portions of the search dataset that are semantically or technically related to each other; determining a second relevance for each dataset portion of a dataset based on comparisons with the search dataset portions of the search dataset and evaluating it using a second relevance value; and sorting and performing data processing on each dataset portion according to the second relevance value as the result.

[0007] The advantage of this approach in the first embodiment is that the method is divided into multiple steps, saving computation time by performing more computationally intensive steps or comparisons only after the first relevance identification and filtering of datasets that no longer need to be processed. Additionally, it is advantageous that the accuracy of search results in terms of dataset details can be significantly improved in order to sort according to the second relevance value.

[0008] A dataset with a first relevance value exceeding the first relevance threshold can also be referred to as a hit dataset. A dataset or hit dataset can be, for example, a single document such as a patent application or patent document or instruction manual or CAD drawing or CAD model or scientific article or book or online book, or it can be a photograph, photo series or video of a physical object, or it can be the object itself or other digital descriptions or illustrations of the object, or it can be other documents.

[0009] The dataset section or the hit dataset section can be a part of the dataset or the hit dataset, especially a part of the aforementioned document.

[0010] Furthermore, the present invention relates to a processor configured to implement the method. This ensures efficient processing of the method steps.

[0011] Additionally, the present invention relates to a computer program comprising commands that, when executed by a computer, cause the computer to perform the method. This allows all method steps to be controlled or emulated from the program code.

[0012] Furthermore, the present invention relates to a computer-readable storage medium having instructions that, when executed by a computer, cause the computer to perform the method. This allows for advantageous sharing of the instructions.

[0013] Similarly, the present invention relates to an interface configured to access the computer program and to pass results, search datasets, search dataset portions, datasets, and dataset portions to a second computer program product for further processing.

[0014] In the method step "perform a search starting from a search dataset, which includes portions of the search dataset, wherein each portion of the search dataset is semantically, graphically, or technically related," the search can be performed across the entire data store. The entire data store may, for example, include all available, previously filed patent applications. The entire data store may also include all other previously published documents.

[0015] Furthermore, it is conceivable that the data stock includes legal cases. In particular, it can be stipulated that the data stock includes judgments of adjudicated legal cases. This can also be referred to as precedent. Specifically, it can be stipulated that the data stock is limited to judgments of adjudicated legal cases in selected countries.

[0016] Determining the first relevance based on comparisons with the search dataset and evaluating the dataset using the first relevance value serves the purpose of limiting the entire stock of all previously published documents to further method steps. Therefore, in further method steps, datasets with first relevance values ​​exceeding the first relevance threshold can be considered. Within this limited dataset, further searches and sorting or sequential listing of the datasets can then be performed.

[0017] In the method step “comparing the individual dataset portions of the dataset having a first relevance value exceeding a first relevance threshold, wherein the comparison is performed individually with semantically or technically related portions of the search dataset,” as described in the preceding paragraph and also herein, only those datasets having a first relevance value exceeding the first relevance threshold can be considered. Therefore, this method step can be performed on a dataset that has already been narrowed down. The individual dataset portions of the dataset having a first relevance value exceeding the first relevance threshold can thus be compared with semantically or technically related portions of the search dataset.

[0018] A second relevance can be identified by comparing each dataset portion with a semantically or technically related portion of the search dataset, and an evaluation using the second relevance value can be performed for each dataset portion. Here, a second relevance value relating to consistency with each search dataset portion in all pre-selected datasets can be assigned to each dataset portion of all pre-selected datasets.

[0019] Here, sorting and data processing of each dataset part based on the second relevance value can be performed separately for each search dataset part.

[0020] In another alternative variation, the method further includes the following steps: identifying a third relevance for the dataset and evaluating it using a third relevance value, the third relevance being a weighted calculation based on the second relevance values ​​of the various dataset portions of the dataset; and storing the search and evaluation results of the dataset and the dataset portions to which it belongs, based on the third relevance value, and performing data processing based on the third relevance value, or displaying it in a sorted manner in a user interface based on the third relevance value.

[0021] Another advantage of this alternative approach is that the search results can be further improved significantly, and other computationally expensive steps for data processing techniques such as storage or display can be limited to datasets that are particularly relevant after being reordered using a third relevance value.

[0022] Furthermore, it is advantageous that the vector embedding is a numerical representation of text, words, phrases, images, or documents. This allows for searches not only based on exact consistency but also on similar datasets.

[0023] Another advantageous approach is to use vector embeddings for searching dataset portions and / or dataset portions, generated through machine learning processes or by utilizing large language models. This allows the search to find results that are consistent only with indirect language, graphics, or technology.

[0024] Another alternative approach is characterized in that the vector embeddings for the dataset portion are performed independently of the search during preparation. This allows computationally intensive steps to be pre-processed, independent of the search query. Consequently, the search becomes faster and more efficient in terms of infrastructure and processor load.

[0025] Also advantageous is the storage of further information about the source and association of the dataset portion in the dataset or adjacent dataset portions, in relation to the vector embedding used for the dataset portion, so that the associations are readable by the software or result in better search hits.

[0026] For example, it can be envisioned that the dataset or hit dataset has a predetermined structure for its construction. This could be the case, for example, in the case of a patent application or patent. Such documents may, for example, have an abstract, a specification introduction, a brief description of the figures, a list of figure labels, patent claims, and figures. For example, information such as in which part of the document, for example, in which part of a patent application, the dataset portion can be given in the vector embedding used for the dataset portion. Furthermore, it can be envisioned that precise source information, such as paragraph or page and line numbers or patent claim numbers, is stored in the vector embedding used for the dataset portion. In particular, it can be specified that the search considers when two dataset portions are very close in source, for example, in the same paragraph or the same patent claim.

[0027] Another advantageous improvement to the method is that the vector embedding is a sentence embedding or a word embedding. Here, sentence embedding can be understood as the embedding of a grammatically complete sentence, containing all the necessary sentence components from a grammatical perspective. However, sentence embedding can also be understood as the embedding of phrases that do not constitute a complete sentence from a strictly grammatical perspective. Here, the phrases can also be words that do not form sentence fragments from a grammatical perspective, but are arbitrarily related from a grammatical perspective, yet efficiently reproduce information without filler words, such as adjectives, adverbs, or particles. This particularly improves the processing of related information existing in textual form.

[0028] What is particularly efficient is that the vector embedding includes context so that the dataset in which the search is performed is smaller and thus can be processed more efficiently.

[0029] Another advantageous approach is that the search dataset and / or a portion thereof is a formula, structure, text, video, diagram, CAD file, image, or audio file, or a portion thereof. Thus, in the case of text, formulas can be individually recognized and processed by the software. In the case of text, the text can be processed directly or in parts to make processing faster and more efficient for the software. The same applies to diagrams, CAD files, images, or audio files. Similarly, the relevance of the invested computational and / or storage capacity can be optimized.

[0030] A particular advantage is that the search is a hybrid search utilizing both keyword and vector searches, allowing different search paths to be used simultaneously, or the more load-efficient path to be executed first. Here, the keyword search can be a Boolean search, in which other operators for the search can be entered besides the keywords themselves.

[0031] The method can be further improved by allowing selection of a portion of the document pool for searching the dataset via prompt-based search on the user interface. This keeps the number of documents to be processed low while maintaining a consistent result hit rate.

[0032] An efficient improvement to the method could advantageously be that the data inventory and the vector embeddings are stored in one or more distributed databases. This approach ensures that the software can process the data easily and systematically, even when the dataset is distributed and located in a distributed manner.

[0033] Equally advantageous is the creation of terms from the search dataset portion and the dataset portion or the hit dataset portion, and the indexing of the search dataset portion and the dataset portion using an inference engine. Thus, software-based search can be implemented energy-efficiently even with very large data stores.

[0034] A particularly advantageous approach is to asynchronously reload the dataset or a portion thereof for additional search and evaluation results after the initial hit results have been displayed on the user interface for an extended period. This allows the processor's computational power to be distributed over a longer timeframe.

[0035] Furthermore, it is conceivable that additional search and evaluation results of the dataset or a portion thereof could be asynchronously reloaded upon certain user actions, such as scrolling with the mouse wheel to a predetermined point or moving the scroll bar to a predetermined point.

[0036] An advantageous variation involves displaying the multidimensional matrix in two dimensions on the user interface, such that each portion of the search dataset is displayed in a single row of the matrix alongside multiple different portions of the dataset. This allows complex relationships to be displayed with fewer display switches, thereby utilizing computational power efficiently. For the user, this also results in a significantly improved detection of multidimensional relationships. This leads to energy savings.

[0037] This can also be improved by selectively displaying multiple dataset portions in the corresponding dataset matrix columns on the user interface, either individually, side-by-side, or overlapping, but in the same matrix rows as the first search dataset portion to be compared. This also allows for efficient processor load management for calculating the information to be displayed.

[0038] Of particular advantage is that changes to the search dataset trigger the determination of new relevance values ​​for the second and / or third relevance values, and therefore the new ranking can be calculated more easily without having to start a completely new computation run.

[0039] Also advantageous is that manual changes to the second relevance value via user interaction on the user interface trigger the determination of a new relevance value for the third relevance value. This also avoids entirely new searches, and already performed calculation steps can be retained or only partially modified.

[0040] To make the search more computationally efficient and faster, a portion of the search dataset or changes to that portion can be stored and used for further machine learning or fine-tuning of the search model.

[0041] Advantageously, datasets and / or portions thereof are stored and / or displayed on a user interface in multiple different languages ​​for a single dataset and a vector embedding, so that a single computational run can process more related information.

[0042] It is also advantageous that the search dataset is a patent claim, because efficient and repeatable accurate computation of search results can be achieved due to the standardized language construction.

[0043] It is also highly advantageous that the search dataset portion and the dataset portion can each be a single feature from a patent claim, patent specification, or patent drawings. Therefore, finer-grained and computationally intensive searches can be limited to sub-regions of a larger dataset.

[0044] Furthermore, it can be specified that the vector embedding is performed using a graphics processor. This offers the advantage that, due to its architecture, a graphics processor is particularly well-suited for computing the vector embedding.

[0045] Furthermore, it can be specified that the search dataset is a description of the relationships between legal cases. This description of relationships can recreate the picture (Bild) of the legal cases. When searching for legal cases from the entire stock of adjudicated legal cases in a selected country, individual legal cases can be found in the stock of cases whose relationship descriptions are similar to those in the search dataset. This can support legal work, particularly in countries with a so-called "common law" legal order.

[0046] In particular, it is conceivable to divide the various features of the legal cases into different parts of the search dataset.

[0047] Furthermore, it can be stipulated that, in the case of utilizing a large language model, argumentative text is written based on at least a few datasets from the datasets found in the search. This offers the advantage that the datasets found in the search can be automatically processed into the argumentative text.

[0048] In particular, it can be specified that only datasets that are selected from search results by user input in a preliminary selection step are included in the argument text and cited there.

[0049] In a first embodiment, the descriptive text may be a notification from the patent office, wherein the search dataset is an independent patent claim of a patent application, and the dataset consists of relevant prior art documents cited in the notification. Specifically, the notification may discuss a lack of inventiveness or novelty relative to the prior art, wherein various sources of the dataset may be cited in connection with the search dataset. In particular, it may be specified that the search dataset is copied verbatim into the descriptive text, with corresponding location descriptions of the sources provided.

[0050] In another embodiment, the argumentative text may be specified as a document or forming the basis for an opposition to a patent, either as grounds for opposition or in opposition proceedings, wherein the search dataset is an independent patent claim of the patent application, and the dataset is relevant prior art documents cited in the opposition.

[0051] In another embodiment, the argumentative text may be specified as a written statement by one party or the basis for such a statement, wherein the search dataset is a description of the relationship between legal cases, and the dataset consists of relevant precedents cited in the written statement. In particular, it may be specified that the case numbers of the relevant precedents are cited in accordance with the usual citation methods in the corresponding legal order.

[0052] In another embodiment, the argumentative text may be specified as being the basis for a court judgment or formation thereof, wherein the search dataset is a description of the relationships between legal cases, and the dataset is relevant precedents cited in the judgment.

[0053] Furthermore, it can be stipulated that a legal database is accessed when the argumentative text is created, wherein, in addition to the dataset, laws are cited in the argumentative text.

[0054] Regarding patents, the legal database may include, for example, patent laws, provisions and regulations of patent conventions, examination guidelines of the patent office, compilations of decisions of the Patent Office Board, compilations of decisions of national courts, etc.

[0055] In the case of legal matters, the legal database may be a national legal database and annotations that interpret national laws.

[0056] Furthermore, it can be specified that the dataset is divided into groups, wherein datasets are assigned to the corresponding groups such that, for each individual search dataset portion within the search dataset, none of the second relevance values ​​exceeds a predetermined second relevance threshold. This provides the advantage that a clear structure of search results can be achieved through this measure. For example, this measure allows datasets to be listed in a first group where the dataset has a dataset portion with a second relevance value exceeding the predetermined second relevance threshold for each search dataset portion; and datasets to be listed in a second group where the dataset does not have a dataset portion with a second relevance value exceeding the predetermined second relevance threshold for any search dataset portion within the search dataset.

[0057] Within the group, the dataset can be displayed in sorted order according to the third relevance value, wherein, within the corresponding group, the following dataset portions can be disregarded in the weighted calculation of the second relevance values ​​of the various dataset portions of the dataset: in the dataset portions, for individual search dataset portions within each search dataset portion of the search dataset, none of the second relevance values ​​exceeds a determined second relevance threshold.

[0058] Furthermore, it can be stipulated that, in order to rank the groups, a portion of the search dataset is selected with the aid of user input: for this portion of the search dataset, none of the second relevance values ​​are allowed to exceed a predetermined second relevance threshold. This offers the advantage that, with the aid of user input, it is possible to determine which features need not be found. This may be relevant, for example, when, for the purpose of evaluating inventiveness, a document about the closest prior art is known, which has certain portions of the search dataset, and these already known portions of the search dataset are no longer relevant for further retrieving a second document for combination with the closest prior art.

[0059] Furthermore, it can be stipulated that, if none of the datasets in the search dataset has a second relevance value exceeding a determined second relevance threshold for all portions of the search dataset, the dataset with the most second relevance values ​​or the highest third relevance value is assigned to the first group, and the dataset containing the missing portions of the search dataset from the first group is assigned to the second group. This offers the advantage that, in the absence of documents that violate novelty by possessing all the features, for example, the closest possible prior art can be listed in the first group, and possible complementary documents can be listed in the second group.

[0060] Furthermore, it can be stipulated that the data stock used for the search has been limited from the globally available data stock. This offers the following advantages: it saves computation time and computing power. For example, it can be envisioned that this pre-limitation is performed using Boolean operators. For instance, it can be limited to certain categories of patent classifications, such as the IPC classification. Additionally, it can be limited to other bibliographical data in patent documents, such as specific application or publication periods, specific countries, specific applicants, and similar items.

[0061] Large Language Models (LLMs) are understood as extensive language models based on machine learning, capable of understanding, generating, and responding to natural language. Models such as GPT-3 (Generative Pretrained Transformer) or LLaMA are examples of LLMs.

[0062] In the context of artificial intelligence (AI)—as is also the case in this application—the term "token" often refers to the basic unit of text processed by an AI model. There are two common AI-related meanings here: A lexical is understood as a unit such as a word or a subunit of a word. For example, the sentence "This cat runs fast (Die Katze läuft schnell)" is divided into lexicals, where each lexical is a single word: "this (Die)," "cat (Katze)," "run (läuft)," and "fast (schnell)." This lexicalization allows AI models to understand and analyze text. Different large language models can only operate with a limited number of lexicals. Language models such as Llama 2 have up to 4096 lexicals, or language models such as CodeLlama have up to 16384 lexicals.

[0063] In the field of artificial intelligence (AI)—as is the case here—the term "cue word" describes a short instruction or text snippet given to an AI model to obtain a desired response or output. A cue word is like a task setting that requires the model to generate or understand specific information.

[0064] A feature is a technically and / or functionally interrelated part of a larger set of features, such as in a patent claim, invention, technical specification, or product. Such part may also consist of a single word, sentence portion, conjunction, or linking word, but such part cannot be further subdivided without losing its technical relevance or semantic meaning—when viewed in isolation.

[0065] A graphics processing unit (GPU) stands for "Graphics Processing Unit" and is a dedicated hardware component optimized for graphics processing and parallel computing—particularly for applications in the fields of machine learning and artificial intelligence.

[0066] Vector embeddings are numerical representations of words, phrases, or documents that help LLMs understand the meaning of words and their relationships. These embeddings improve the output of AI models in terms of speed and scope. Vector embeddings are a way of converting words, sentences, and other data into numbers that capture their meaning and relationships. They represent different data types as points in a multidimensional space, where data points that are similar to each other are closer to each other. These numerical representations help machines understand and process this data more effectively.

[0067] Word embeddings represent individual words as vectors. The model learns word embeddings by capturing semantic relationships and contextual information from large texts.

[0068] Sentence embeddings represent an entire sentence or phrase as a vector. Models such as Universal Sentence Encoder (USE) and SkipThought generate embeddings that capture the meaning and context of the sentence as a whole.

[0069] Vector search is fundamental to the semantic or similarity search experience. Because meaning and context are captured in the embeddings, vector search understands the user's intent even without exact keyword matching. This approach works for text data, documents, images, videos, graphics, and audio data.

[0070] Furthermore, a computer-implemented method may be specified for sorting hit datasets of a search, wherein the search, as a result, has multiple hit datasets corresponding to the search dataset in a data store. The method includes the following steps: - Starting from the search dataset, perform the search within the data inventory using artificial intelligence algorithms to find the matching dataset; - Calculate and configure a value consistent with the search dataset for each of the hit datasets; - Sort the hit dataset based on the consistency value; - On the output device The search dataset is displayed, wherein the search dataset is divided into at least one first search dataset portion and one second search dataset portion, wherein the first search dataset portion and the second search dataset portion are displayed separately and are visually distinguishable from each other; Display a top-ranked hit dataset, the hit dataset having a first hit dataset portion and a second hit dataset portion, wherein the first hit dataset portion is graphically assigned to the first search dataset portion when displayed, and wherein the second hit dataset portion is graphically assigned to the second search dataset portion when displayed, wherein a first consistency value between the top-ranked hit dataset and the search dataset is displayed. Display a second-ranked hit dataset, the hit dataset having a first hit dataset portion and a second hit dataset portion, wherein the first hit dataset portion is graphically assigned to the first search dataset portion when displayed, and wherein the second hit dataset portion is graphically assigned to the second search dataset portion when displayed, wherein a second consistency value between the second-ranked hit dataset and the search dataset is displayed; - Provide an input element for user input to adapt to the search dataset, or for user input to adapt to at least one of the consistent values; - The hit dataset is reordered taking into account user input entered in the input element.

[0071] The method offers the following advantages: by reordering the hit dataset based on user input, a simple adaptation of search results can be achieved without forcing computationally intensive and costly new searches. Therefore, even in cases of minimal adaptation in the evaluation of the search dataset or hit dataset, multiple computationally intensive search runs can be avoided. Furthermore, the method according to the invention reduces the time spent on searches. Thus, computational power can be saved, thereby reducing the energy consumption of the computer on which the method is executed.

[0072] Furthermore, it can be specified that each part of the search dataset be identified. This identification can be, for example, a number, such as 1, 2, 3 or I, II, III. The identification can also be a letter, such as A, B, C or a, b, c. Of course, any other conceivable identification can be used, such as symbols like + and -, or identifiers from different languages. Additionally, it can be specified that the location of the corresponding hit dataset part is displayed within the corresponding hit dataset.

[0073] Furthermore, it can be specified that the input element for user input includes an evaluation element, wherein the evaluation element is configured for user input to at least one of the hit datasets, and, taking into account the user input input to the evaluation element, the consistency value is recalculated for at least the hit dataset to which the evaluation element is configured, and the hit datasets are reordered based on the recalculated consistency value. This brings the following advantages: by reordering the hit datasets based on the recalculation of the consistency value, which is performed by user input input to the evaluation element, the search results can be adapted without performing computationally intensive and costly new searches. Therefore, this measure can save computing power and thereby reduce the energy consumption of the computer on which the method is executed. In addition, this measure can reduce the interaction time between the user and the computer executing the method, because the method according to the invention does not require possible user actions, such as manually moving or deleting the hit datasets. It is also advantageous for other data processing techniques, such as storing the hit dataset, or subsequently requiring less computational power, if the consistent value is recalculated at least for the hit dataset to which the evaluation element is configured when user input is given to the evaluation element, since all relevant data can be taken into account here, and thus repeatability can be achieved without overusing computational power.

[0074] The evaluation element used to influence the consistency value can be, for example, a forward sorting element, by which the hit dataset to which the evaluation element is assigned is ranked before the hit dataset with a higher consistency value. Here, the consistency value of the hit dataset to which the evaluation element is assigned can be automatically adjusted.

[0075] The evaluation element used to influence the consistency value can be, for example, a shift-sorting element, by which the hit dataset to which the evaluation element is assigned is arranged after the hit dataset with a lower consistency value. Here, the consistency value of the hit dataset to which the evaluation element is assigned can be automatically adjusted.

[0076] The evaluation element used to influence the consistency value can be, for example, a deletion element, by which the hit dataset to which the evaluation element is assigned can be deleted from the displayed hit dataset list.

[0077] Furthermore, it can be specified that the input element for user input includes an input field, which allows the search dataset to supplement another part of the search dataset. This offers the following advantages: the search dataset can be extended with additional features. Instead of completely re-executing the search within the existing data using artificial intelligence algorithms, only the values ​​calculated and configured for each hit dataset are updated to match the search dataset, and the hit datasets are reordered considering the supplemented search dataset portion. This significantly saves computational resources.

[0078] Furthermore, it can be specified that the input element for user input includes a separating element, by means of which one of the search dataset portions can be divided into another search dataset portion. This brings the following advantages: this measure can improve the results of the hit dataset ranking, and it is not necessary to completely re-execute the search, thereby saving computational power.

[0079] Furthermore, it can be specified that the input element for user input includes an adaptation field, by which content from one of the search dataset portions can be adapted. This offers the advantage that the results of ranking the hit dataset can be improved, and the search does not need to be completely re-executed, thereby saving computational power. In a further improved embodiment, however, it can be specified that when a change in one of the search dataset portions exceeds a certain extent, all subsequent method steps already performed using the artificial intelligence algorithm in the data inventory and derived from the changed search dataset are re-executed.

[0080] Furthermore, it can be specified that the input element for user input includes a deletion element, by which a portion of the search dataset can be deleted. This measure allows the search to be easily adapted. In particular, it can be specified that all subsequent method steps already performed using the artificial intelligence algorithm in the data stock and derived from the modified search dataset are re-executed.

[0081] Furthermore, it is suitable that the artificial intelligence algorithm be automatically adapted based on the user input for future searches within the data store. This offers the advantage that future searches can provide better results and therefore require less computational power.

[0082] Furthermore, it can be stipulated that a consistency value is calculated for each hit dataset, corresponding to the search dataset, so that future searches within the data inventory are automatically adapted based on the user input. This offers the advantage that future searches can provide better results and therefore require less computational power.

[0083] Furthermore, it can be specified that the reordered hit dataset is fed to data processing. This data processing may, for example, store the hit dataset in a storage medium for later output on an output device. Alternatively, it may transfer the hit dataset to another computer for display on an output device. Or, it may transfer the hit dataset to a printing device to print the search results. These features offer the advantages that the data processing can handle the correctly ordered hit dataset and that no further adaptation of the hit dataset is required during processing, resulting in energy savings. Furthermore, these measures improve the quality of the data structure used for data processing, which in turn improves the reproducibility of the results.

[0084] Also advantageous is the following implementation scheme, which specifies that a consistency value with the search dataset is calculated for each hit dataset using a consistency value calculation algorithm, based on at least one first consistency value portion of the first hit dataset portion and the first search dataset portion, and a second consistency value portion of the second hit dataset portion and the second search dataset portion. This offers the advantage that the respective consistency value portions can be considered in the calculation of the consistency value of the corresponding hit dataset. Therefore, improved search results can be achieved, thereby saving overall computation time.

[0085] Furthermore, it can be specified that the consistency value calculation algorithm forms an average of the individual consistency value components, wherein only the following hit datasets are listed in the display of the hit datasets: in the hit datasets, the individual consistency value components exceed a consistency value component threshold. This avoids the unnecessary display of irrelevant hit datasets, thereby saving computational resources. In addition, this measure can shorten the total time that must be spent on searching.

[0086] In an alternative implementation variant, it may be specified that the consistency value calculation algorithm forms the consistency value calculation based on the respective consistency value portions, wherein weights are assigned to the respective consistency value portions. It may also be specified that only the following hit datasets are listed in the display of the hit datasets: in the hit datasets, the respective consistency value portions exceed a consistency value portion threshold.

[0087] Specifically, it can be specified that the weighting of the respective consistent value portions is performed taking into account the search dataset hierarchy of the corresponding search dataset portion. For example, search dataset portions at lower search dataset hierarchies may have smaller weights.

[0088] Also advantageous is the following implementation scheme, which specifies that at least the first hit dataset portion of the second-ranked hit dataset is equipped with a first evaluation element for user input, and the second hit dataset portion of the second-ranked hit dataset is equipped with a second evaluation element for user input, wherein the consistency value is recalculated at least for hit datasets that are equipped with either a first evaluation element or a second evaluation element addressed by user input. This provides the advantage that the respective consistency value portions can be considered in the calculation of the consistency value of the corresponding hit dataset. Therefore, improved search results can be achieved, thereby saving overall computation time.

[0089] According to a further extension, at least one of the hit dataset portions may have at least one first hit dataset position and one second hit dataset position, wherein the first hit dataset position and the second hit dataset position are visually distinguishable and displayed separately, wherein each hit dataset position is assigned a hit dataset position consistency value, and wherein the hit dataset positions are displayed in order according to their respective hit dataset position consistency values. This offers the advantage that not only can a search dataset portion be assigned one hit dataset portion, but multiple hit dataset positions within a corresponding hit dataset portion can also be assigned to a corresponding search dataset portion. This measure can significantly improve search results and thus save overall search time. The search results can therefore correspond to a multidimensional matrix.

[0090] Alternatively, the consistency value portion can be assigned a consistency value for the hit dataset location based on the highest consistency value among the hit dataset locations. This offers the advantage that, within the corresponding hit dataset portion, the source is used for evaluation based on the highest consistency with its corresponding search dataset portion. This measure can significantly improve search results and thus save overall search time.

[0091] Furthermore, it can be specified that the input element for user input includes an evaluation element, wherein the location evaluation element is assigned to at least one of the hit dataset locations for user input, and considering the user input input to the location evaluation element, the consistency value is recalculated at least for the hit dataset location to which the evaluation element is assigned. This brings the following advantages: by reordering the hit dataset based on the recalculation of the consistency value, which is performed by the user input input to the evaluation element, the search results can be adapted without performing a computationally intensive and costly new search. Therefore, this measure can save computing power and thereby reduce the energy consumption of the computer on which the method is executed. In addition, this measure can reduce the interaction time between the user and the computer executing the method, because the method according to the invention does not require possible user actions, such as manually moving or deleting hit dataset locations.

[0092] Furthermore, it is suitable that the various search dataset portions can be assigned to different search dataset portion levels, wherein the affiliation of the search dataset portions to their respective search dataset portion levels is visually distinguishable and displayed separately. This brings the following advantages: this measure can improve search accuracy, thereby saving computation time. For example, it can be stipulated that a feature A forms a search dataset portion, which belongs to a first search dataset portion level. Features A1 and A2 can be, for example, functions of feature A, which each form a search dataset portion, which belongs to a second search dataset portion level.

[0093] The identifiers for each level can be, for example, numbers such as 1, 1.1, 1.2, 1.1.1, 1.1.2, 2, 3, or 1, 1-I, 1-II, 1-II, 1-I-II, 2, 3, or I, II, I-II, II, III, or I, I-1, I-2, II, III. The identifiers for each level can also be letters, such as A, Aa, Ab, B, C, or a, aA, aB, b, c. The identifiers for each level can also be a combination of letters and numbers, such as A, A1, A2, B, or 1, 1A, 1B, 2. The identifiers for each level can also be any other combination of numbers and letters. Of course, any other conceivable identifier can be used, such as symbols like + and -, or identifiers from different languages.

[0094] Furthermore, it can be specified that when one of the hit datasets is selected, the display is updated, wherein the following content is displayed on the output device: The search dataset is displayed, wherein the search dataset is divided into at least one first search dataset portion and one second search dataset portion, wherein the first search dataset portion and the second search dataset portion are visually distinguishable and displayed separately. The selected hit dataset is displayed, the hit dataset having a first hit dataset portion and a second hit dataset portion, wherein the first hit dataset portion is graphically assigned to the first search dataset portion during display, and wherein the second hit dataset portion is graphically assigned to the second search dataset portion during display. The selected hit dataset is displayed in full text, as a text extract, or as a graph, with portions of the hit dataset highlighted within the text, extract, or graph. This offers the advantage that particularly interesting hit datasets can be analyzed in detail. This can significantly reduce the overall time required to perform a retrieval.

[0095] In a first embodiment, the full text may be specified as the complete written portion of the patent application. Furthermore, the text extract may be specified as the specification of the patent application. Additionally, the text extract may be specified as a patent claim of the patent application. Furthermore, it may be specified that a text extract is only a part of the specification or patent claim of the patent application.

[0096] Furthermore, it can be stipulated that multiple text excerpts be displayed. For example, it can be envisioned that the patent claims and the specification of a patent application are displayed separately from each other.

[0097] Alternatively, it can be envisioned that the graphics are displayed separately.

[0098] Furthermore, it can be specified that a text portion can be selected from the full text or the text extract, and this text portion can be assigned to one of the search dataset portions, thus becoming a hit dataset portion, wherein a consistency value portion can also be assigned to the hit dataset portion. This brings the following advantages: search results can be improved through this measure.

[0099] In particular, it can be specified that the selected text portion and the configured consistent value portion are used to train the artificial intelligence algorithm.

[0100] The consensus value or a portion thereof may be a percentage of consensus in numerical terms, or consensus expressed numerically in other ways. The consensus value may also be a vector value or matrix of consensus, or it may be a matrix-vector product.

[0101] The consistency value can be displayed by giving a numerical value, especially a percentage.

[0102] The display of the consistent values ​​can be done in a hierarchical manner.

[0103] The consistency value can be displayed using color grading, such as green, yellow, and red, where a consistency value range can be assigned to each color. For example, a consistency value range of 100% to 80% can be assigned to the green color.

[0104] The consistency value can be displayed graphically, such as using symbols "↑", "→", and "↓", where a consistency value range can be assigned to each symbol. For example, the symbol "↑" can be assigned a consistency value range of 100% to 80%.

[0105] Furthermore, it can be envisioned that the display of the consistency value is achieved through a color gradient, such as a transition from green to red.

[0106] Alternatively, it can be envisioned that the display of the consistency value is achieved through a graphical gradient, such as a bar that has a blackening or marking up to the area corresponding to the consistency value.

[0107] Also advantageous is the following implementation scheme, which specifies a two-dimensional display of a multidimensional matrix on the output device, wherein each portion of the search dataset is displayed in a row of the matrix along with multiple portions of different hit datasets. This approach enables a clear display of search results, thereby reducing the time required for searching. Complex relationships can be displayed with fewer display switches, thus utilizing computing power efficiently. For the user, this also results in a significantly improved detection of multidimensional relationships. This leads to energy savings.

[0108] According to a further extension, the various search dataset portions of the search dataset are displayed in a row of a matrix along with multiple different hit dataset portions and the hit dataset positions of the multiple different hit datasets belonging to the corresponding hit dataset portions. This measure enables a clear display of search results, thereby reducing the time required for searching. It also allows for efficient processor load calculation of the information to be displayed, resulting in energy savings.

[0109] Here, the hit dataset positions can be displayed side-by-side in a row distinct from each other, within the same search dataset section, where switching between the various hit dataset positions is possible using a scroll bar or by scrolling. Furthermore, it is conceivable that the various hit dataset positions can be displayed vertically in a row distinct from each other, within the same search dataset section, where switching between the various hit dataset positions is possible using a scroll bar or by scrolling. In particular, it can be specified that the various hit dataset positions are arranged in separate windows, where scrolling or scrolling bars first operate within the window of the hit dataset position and jump to the next hit dataset position at the end of the window.

[0110] Furthermore, it is suitable that one of the hit datasets can be selected as the closest prior art, wherein the consistency value portion with each search dataset portion may be below a threshold or may be non-existent, wherein additional hit datasets with high consistency values ​​between the hit dataset portion and these selected closest prior art portions, which have fewer or no consistency values, are preferentially displayed in the ranking. This provides the advantage that the closest prior art can be defined, and related documents can be searched. This can be particularly advantageous when only hit datasets with low consistency are found for the search dataset. Attached Figure Description

[0111] To better understand the present invention, it will be explained in more detail with the aid of the following drawings.

[0112] The following are illustrated in highly simplified diagrams: Figure 1 A schematic diagram of the infrastructure used to perform the method according to embodiment 1 is shown; Figure 2 A schematic illustration showing a first embodiment of a multidimensional matrix in a two-dimensional display; Figure 3 A schematic illustration showing a second embodiment of a multidimensional matrix in a two-dimensional display; Figure 4 A schematic illustration showing a third embodiment of a multidimensional matrix in a two-dimensional display; Figure 5 A schematic flowchart illustrating a first embodiment of a method for searching a dataset corresponding to a search dataset in a data inventory. Detailed Implementation

[0113] First, it should be determined that in embodiments described differently, the same components are given the same reference numerals or the same component names, wherein the disclosure contained throughout the specification can be semantically transferred to the same components having the same reference numerals or the same component names. Furthermore, the positional descriptions selected in the specification, such as upper, lower, sideways, etc., relate to the directly described and shown figures, and these positional descriptions are semantically transferred to the new position when the position changes.

[0114] exist Figure 1 The diagram illustrates a first embodiment of the infrastructure used to perform the method, and explains the method steps with the aid of the diagram.

[0115] To perform the method, system 1 can be set up in hardware and software infrastructure 9. The system may have one or more processors 2, which may include at least one central processing unit 3 and / or a graphics processing unit 4.

[0116] To operate the system 1, a management program 6 can be used to access information in a data store 14, which is stored in a database 7. This information can be displayed via a user interface 5 using the management program 6. Specifically, the user interface 5 can be specified to be displayed on an output device 101, such as a screen. The information in the data store 14 can be, for example, a vector embedding 10, a search dataset 12, a search dataset portion 13, a dataset 15, a dataset portion 16, a hit dataset 113, or a hit dataset portion 114, 115, 116. The search result 21 can be displayed on the user interface 5. A search dataset 12, a search dataset portion 13, a dataset 15, a dataset portion 16, a hit dataset 113, or a hit dataset portion 114, 115, 116 can be text 22, video 23, drawings 24, CAD files 25, images 26, audio files 27, a structure 32, or a portion thereof.

[0117] The computer-implemented method for searching a dataset 15 in data storage 14 that corresponds to a search dataset 12 includes the following steps: performing a search starting from the search dataset 12, wherein the search dataset preferably includes at least two or more search dataset portions 13, wherein the search dataset portions 13 are semantically, graphically, or technically related; configuring vector embeddings 10 in the search dataset 12 and / or search dataset portions 13; configuring vector embeddings 10 in the data storage 14 in the datasets 15 and / or dataset portions 16 stored therein; determining a first relevance based on comparisons with the search dataset 12, and... The dataset 15 is evaluated using a first relevance value 17; each dataset portion 16 of the dataset 15 having a first relevance value exceeding a first relevance threshold 18 is compared, wherein the comparison is performed individually with each semantically or technically related search dataset portion 13 of the search dataset 12; based on the comparison with each search dataset portion 13 of the search dataset 12, a second relevance is determined for each dataset portion 16 of a dataset 15 and evaluated using a second relevance value 19; and each dataset portion 16 is sorted and data processed according to the second relevance value 19 to obtain the result 21.

[0118] In addition, the method may also include the following steps: identifying a third relevance for the dataset 15 and evaluating it using a third relevance value 20, the third relevance being a weighted calculation based on the second relevance values ​​19 of each dataset portion 16 of the dataset 15; storing the search and evaluation results of the dataset 15 and its dataset portion 16 according to the third relevance value 20; and performing data processing according to the third relevance value 20, or displaying it in sorted form in the user interface 5 according to the third relevance value 20.

[0119] The vector embedding 10 can be a numerical representation of text 22, words, phrases, images 26, or documents.

[0120] The vector embeddings 10 used for searching dataset portion 13 and / or dataset portion 16 can be generated through a machine learning process or by utilizing a large language model 11.

[0121] The vector embedding 10 for the dataset portion 16 can be performed independently of the search during the preparation process. This process can run asynchronously and is designed to efficiently utilize the infrastructure 9 and processor 2, particularly the graphics processor 4.

[0122] In terms of the vector embedding 10 used for dataset portion 16, further information about the source and association of said dataset portion 16 in dataset 15 or adjacent dataset portions 16 can be stored.

[0123] Among them, vector embedding 10 can use sentence embedding or word embedding.

[0124] The vector embedding 10 may include context. This may include, for example, a defined technical field, an existing associated dataset 15 in the data stock 15, which may be automatically selected by the user or AI, or defined by search or prompt words.

[0125] The search dataset 12 and / or the search dataset portion 16 and / or the dataset 15 and / or the dataset portion 16 may be a formula, text 22, video 23, drawing 24, CAD file 25, image 26, or audio file 27 or a portion thereof.

[0126] The search can be a hybrid search that utilizes both keyword and vector search. This involves combining classic search queries such as text or numbers with AI-based search functionality.

[0127] A portion of the document pool in the data store 14 can be selected via prompt-based search on the user interface 5 for searching the dataset 15.

[0128] The data storage 14 and the vector embedding 10 can be stored in one or more distributed databases 7.

[0129] Lexical units can be created from search dataset portion 13 and dataset portion 16, and indexed using inference engine 29.

[0130] After the initial hit results are displayed on user interface 5 for an extended period of time, additional search and evaluation results for dataset 15 or dataset section 16 can be loaded asynchronously.

[0131] The vector embedding 10 can be performed using a graphics processor.

[0132] The system 1 may have a processor configured to implement the method.

[0133] The computer program 28 may include commands that cause the computer or processor to perform the method when the program is executed by the computer or by the processor 2 of the system 1 in the infrastructure 9.

[0134] The computer-readable storage medium may have instructions that, when executed by a computer or by a processor 2 of system 1 in infrastructure 9, cause the computer or processor to perform.

[0135] The system 1 may also have an interface 31, which may be configured to access the computer program 28 and to pass the result 21, the search dataset 12, the search dataset portion 13, the dataset 15, and the dataset portion 16 to a second computer program product for further processing.

[0136] Figure 2 A schematic illustration of a first embodiment of the multidimensional matrix 8 in a two-dimensional display is shown, for example, on the user interface 5 of the management program 6. The user interface 5 can be output and visually displayed on the output device 101.

[0137] The multidimensional matrix 8 can be displayed in two dimensions on the user interface 5 such that the various search dataset portions 13 of the search dataset 12 are displayed together in a row of the matrix 8 with the multiple different dataset portions 16 of the dataset 15.

[0138] Multiple dataset portions 16 in the dataset 15 can also be selectively displayed individually, side-by-side, or overlappingly in the columns of the matrix 8 of the corresponding dataset 15 on the user interface 5, but in the same rows of the matrix 8 as the search dataset portion 13 to be compared.

[0139] Changes to the search dataset portion 13 can trigger the determination of new relevance values ​​for the second relevance value 19 and / or the third relevance value 20.

[0140] Manually changing the second correlation value 19 through user interaction on user interface 5 can trigger the determination of a new correlation value for the third correlation value 20.

[0141] Changes to search dataset section 13 or dataset section 16 can be stored and used for fine-tuning of other machine learning or search models, such as in large language model 11.

[0142] Dataset 15 and / or dataset section 16 may be stored and / or displayed on user interface 5 in multiple different languages ​​for a dataset 15 and a vector embedding 10.

[0143] The search dataset 12 may be a patent claim.

[0144] The search dataset portion 13 and the dataset portion 16 or the hit dataset portion may be individual features in the patent claims, patent specification or patent drawings.

[0145] Figure 3 A schematic illustration of a second embodiment of the multidimensional matrix 8 in a two-dimensional display is shown, as can be visually perceived, wherein the same components are used again as before. Figures 1 to 2The same reference numerals or part names are used in the accompanying drawings. To avoid unnecessary repetition, refer to or cite previous drawings. Figures 1 to 2 The detailed description is in the text.

[0146] Here, the display can also be performed on the user interface 5 of the management program 6.

[0147] As from Figure 3 Therefore, it can be specified that the search datasets 12 and 102, when optically displayed, include a first search dataset portion 103, a second search dataset portion 104, and a third search dataset portion 105, or generally described as at least two, but optionally also including multiple search dataset portions 13. The various search dataset portions 13, 103, 104, and 105 can be semantically related. The search datasets 12 and 102 can form the basis of a search. In particular, it can be specified that the search datasets 12 and 102 are extracted from documents or created from information in documents.

[0148] For example, it can be envisioned that the respective search dataset portions 13, 103, 104, and 105 constitute independent claims of a patent application. In other words, the respective search dataset portions 13, 103, 104, and 105 can be derived from or reproduce the content of independent claims of a patent application.

[0149] The first search dataset portion 103 may be, for example, feature combination A. The second search dataset portion 104 may be, for example, feature combination B. The third search dataset portion 105 may be, for example, feature combination C. Of course, it is also conceivable that the search dataset 102 may only include the first search dataset portion 103 with feature combination A and the second search dataset portion 104 with feature combination B+C, or other partitions.

[0150] Furthermore, it is conceivable that other search dataset portions exist besides the third search dataset portion 105. The aforementioned combination of features can therefore be further subdivided into individual search dataset portions 13.

[0151] The division of features or feature combinations into the search dataset section 13 can be done by the user or automatically.

[0152] As from Figure 3 Furthermore, it can be stipulated that the respective search dataset portions 13, 103, 104, and 105 are assigned to the first search dataset portion level 106. Therefore, the respective search dataset portions 13, 103, 104, and 105 can, in principle, have the same value.

[0153] Specifically, it can be specified that the various search dataset portions 13, 103, 104, and 105 are distinguishably identified from each other by means of a search dataset portion identifier 107. Such a search dataset portion identifier 107 may be, for example, the number of the various search dataset portions 13, 103, 104, and 105.

[0154] In one first embodiment, the partitioning of the search dataset 102 may be performed by an artificial intelligence algorithm. For example, the search dataset 102 may be based on an independent patent claim of a selected patent application, which is partitioned into search dataset portions 13, 103, 104, and 105 using the artificial intelligence algorithm. Furthermore, it is conceivable that other search dataset portions may be formed, for example, by dependent patent claims of the selected patent application.

[0155] Furthermore, it can be specified that an input element 108 is provided for user input. The input element 108 may, for example, include a separator element 109, by which one of the search datasets 12, 102, or the search dataset portions 13, 103, 104, 105 can be divided into other search dataset portions 13. The separator element 109 can be implemented, for example, by positioning the cursor within a text portion of one of the search dataset portions 13.

[0156] Here, the separation itself can be made, for example, by clicking with a mouse. Alternatively, it can be envisioned that the separation is made, for example, by pressing a keyboard key, such as the Enter key.

[0157] Furthermore, it is conceivable that the input element 108 includes an adaptation field 110, by means of which one of the search dataset portions 13, 103, 104, 105 can be adapted in terms of its content. This can be achieved, for example, by making editing possible when one of the search dataset portions 13, 103, 104, 105 is clicked or otherwise selected.

[0158] Furthermore, it is conceivable that the input element 108 includes a deletion element 111, which is used to delete one of the search dataset portions 13, 103, 104, and 105.

[0159] Furthermore, it can be specified that the input element 108 includes an input field 112, which is used to add another search dataset portion 13.

[0160] Therefore, there are various possibilities to adapt the search datasets 12, 102 or their search dataset portions 13, 103, 104, 105.

[0161] Furthermore, it is conceivable that the search dataset 102 or the various search dataset portions 12, 103, 104, 105 are not taken from existing documents, but are directly entered by the user.

[0162] The various search datasets 13, 103, 104, and 105 can form the basis of the search.

[0163] As from Figure 3 It can be further seen that the various search dataset parts 13, 103, 104, and 105 can be displayed stacked vertically in a column.

[0164] Furthermore, it is conceivable that different hit datasets 113 found in the search may be displayed in other columns on the output device 101. The hit datasets 113 may be a selection of those datasets 15 that are also displayed.

[0165] In particular, it can be specified that the various hit datasets 113 are displayed side-by-side in each column according to a specific sorting.

[0166] Here, for example, it can be envisioned that a hit dataset 113a, ranked first, is displayed directly next to the search datasets 12 and 102. Next to it, a hit dataset 113b, ranked second, and next to it, a hit dataset 113c, ranked third, can be displayed. Furthermore, it is of course envisioned that other hit datasets 113 can be displayed. In particular, it can be specified that the displayed hit datasets 113 can be selected using a scroll bar, or that the display can be moved, so that multiple hit datasets 113 can be displayed.

[0167] As from Figure 3 It can be further seen that at least some of the hit datasets 113 have a first hit dataset portion 114, which is assigned to the first search dataset portion 103.

[0168] Furthermore, at least some of the hit datasets 113 may have a second hit dataset portion 115, which is configured to the second search dataset portion 104.

[0169] Furthermore, at least some of the hit datasets 113 may have a third hit dataset portion 116, which is configured to the third search dataset 105.

[0170] Each first hit dataset portion 114 can be arranged in the same row as the first search dataset portion 103. Each second hit dataset portion 115 can be arranged in the same row as the second search dataset portion 104. Each third hit dataset portion 116 can be arranged in the same row as the third search dataset portion 105.

[0171] Here, a row is understood as a row in a matrix or table rather than a single line of text. Therefore, it is conceivable to display longer text with multiple lines of text in a single row of a matrix or table.

[0172] In particular, it is conceivable that scrollbars or scrolling possibilities exist in the various search dataset sections 13, 103, 104, 105 or in the various hit dataset sections 114, 115, 116, so that longer text can be displayed in a clear manner.

[0173] As from Figure 3 Furthermore, it can be specified that a consistency value 117 is displayed for each of the hit datasets 113a, b, c, etc. For example, the first hit dataset 113a may have a first consistency value 117a that is 98% consistent with the search dataset 102. Additionally, for example, the second-ranked hit dataset 113b may have a second consistency value 117b that is 80% consistent with the search dataset 102. In particular, it can be specified that the ranking of the hit datasets 103 is based on the consistency value 117. The consistency value 117 may correspond to a value also referred to in this document as the third relevance value 20, and this value can be determined using the described computer-implemented method.

[0174] As from Figure 3 It can be further seen that the first hit dataset portion 114a of the first hit dataset 113a ranked first has a first consistent value portion 118a with the first search dataset portion 103.

[0175] Furthermore, it can be specified that the second hit dataset portion 115a of the first hit dataset 113a has a second consistent value portion 119a with the second search dataset portion 104.

[0176] Furthermore, it can be specified that the third hit dataset portion 116a of the first hit dataset 113a has a third consistent value portion 120a with the third search dataset portion 105.

[0177] Furthermore, it can be specified that the first hit dataset portion 114b of the second-ranked hit dataset 113b has a first consistent value portion 118b with the first search dataset 103. This can be similarly continued for all other hit dataset portions 114, 115, 116.

[0178] In the case of the various consistency value portions 118, 119, and 120, it can also be specified that the consistency value portion is displayed as a numerical value, such as a percentage. Furthermore, it can be specified that the consistency value 117a of the first-ranked hit dataset 113a is calculated using a consistency value calculation algorithm based on the first consistency value portion 118a, the second consistency value portion 119a, and the third consistency value portion 120a.

[0179] The respective hit dataset portions 114, 115, and 116 can be, for example, text fragments of the hit dataset 113. The respective consistency value portions 118, 119, and 120 can display the consistency values ​​between the corresponding hit dataset or its text fragments and the corresponding search dataset portions 103, 104, and 105 or their text fragments.

[0180] As is particularly evident from the second hit dataset portion 115a of the first-ranked hit dataset 113a, it can be specified that this hit dataset portion 115a has a first hit dataset position 121a and a second hit dataset position 122a. Of course, other hit dataset positions may also exist. For example, this can occur when multiple text fragments exist in the first-ranked hit dataset 113a, and these text fragments have sufficient consistency with the second search dataset portion 104. Here, the first hit dataset position 121a may have a consistency value 123a with the first hit dataset position of the second search dataset portion 104. The second hit dataset position 122a may have a consistency value 124a with the second hit dataset position of the second search dataset portion 104. These first hit dataset position consistency values ​​123 or second hit dataset position consistency values ​​124, etc., may correspond to the second relevance value 19, as already described or derived therefrom.

[0181] Here, the second consistency value portion 119a of the first hit dataset 113a can be equal to the highest hit dataset position consistency value 123a of the hit dataset positions 121 and 122a. This scheme can be similarly applied to the remaining hit datasets 113 or the remaining hit dataset portions 114, 115, and 116.

[0182] In other words, in each hit dataset portion 114, 115, 116 of the corresponding hit dataset 113, the following hit dataset positions 121, 122 can be displayed: the hit dataset positions have the highest consistency with the corresponding involved search dataset portions 13, 103, 104, 105, wherein the consistency is measured at hit dataset position consistency values ​​123, 124.

[0183] As from Figure 3 Furthermore, it can be specified that the input element 108 for user input includes an evaluation element 126. Specifically, it can be specified that each of the hit datasets 113 is assigned an evaluation element 126. The evaluation element 126 can influence the consistency value 117 of the corresponding hit dataset 113.

[0184] Similarly, it can be specified that a first evaluation element 127b is provided for user input in the first hit dataset portion 114b of the second-ranked hit dataset 113b. Furthermore, it can be specified that a second evaluation element 128b is provided for user input in the second hit dataset portion 115b of the second-ranked hit dataset 113b. This can also be continued similarly for the remaining hit datasets 113.

[0185] Furthermore, it is conceivable that evaluation elements are also set for each of the hit dataset locations 121 and 122.

[0186] Specifically, it can be stipulated that when a user input is made to one of the input elements 108, the consistency value of the corresponding hit dataset 113 involved is recalculated, and the hit dataset 113 is reordered in accordance with the corresponding result, taking into account the user input made in the input element 108.

[0187] Figure 4 A schematic illustration of a third embodiment of the multidimensional matrix 8 in a two-dimensional display is shown, wherein the same components are used again as before. Figures 1 to 3 The same reference numerals or part names are used in the accompanying drawings. To avoid unnecessary repetition, refer to or cite previous drawings. Figures 1 to 3 The detailed description is in the text.

[0188] As from Figure 4 It can be specified that the search dataset 102 has a first search dataset portion 103, which is arranged in the first search dataset portion hierarchy 106.

[0189] The second search dataset portion 104 can be arranged in the second search dataset portion hierarchy 129. The third search dataset portion 105 can also be arranged in the second search dataset portion hierarchy 129. A fourth search dataset portion can again be arranged in the first search dataset portion hierarchy 106. This hierarchical arrangement can be meaningful, for example, for components and their constituent parts, or for components and their sub-components, or for features and their functions.

[0190] The configuration of the various search dataset parts 13, 103, 104, and 105 to the search dataset part levels 106 and 129 can, of course, be performed in other orders. Alternatively, they can be configured to additional search dataset part levels.

[0191] Figure 5 Another embodiment of a possible method flow for searching dataset 15 corresponding to the search datasets 12, 102 or displayed as hit dataset 113 is shown.

[0192] In a search dataset creation step 130, it may be specified that a search dataset 12, 102 or its search dataset portions 13, 103, 104, 105 are defined.

[0193] In a first configuration step 131, the vector embedding 110 may then be configured to search datasets 12, 102 or to search dataset portions 13, 103, 105.

[0194] Furthermore, it can be specified that in a second configuration step 132, the vector embedding in the data store 114 is configured to the dataset 15 or dataset portion 16 stored therein.

[0195] In a first evaluation step 133, a first relevance can then be determined and evaluated using the first relevance value 17. For this purpose, the vector embeddings in the data store 114 for the dataset 15 or dataset portion 16 stored therein, and the configuration of the vector embeddings 110 to search datasets 12, 102 or search dataset portions 13, 103, 105 can be used as a reference or comparison.

[0196] In a first selection step 134, a dataset 15 can then be selected for further data processing if it has a first correlation value 17 that exceeds a first correlation threshold 18. This results in the selection of the dataset 15, thus eliminating the need to continue analyzing all datasets 15 in the data stock 114.

[0197] In a first comparison step 135, each dataset portion 16 of the dataset 15 selected in the first selection step 134 may then be individually compared with semantically or technically related search dataset portions 13, 103, 105 of the search datasets 12, 102.

[0198] In a second evaluation step 136, a second correlation can then be identified for each dataset portion 16 of a dataset 15 based on the comparison in the first comparison step 135, and assigned as a second correlation value 19 to the corresponding dataset portion 16.

[0199] Here, each of the dataset portions 16 may have a second relevance value 19 relating to each of the search dataset portions 13, 103, 105. To understand this as an analogy to a database, an m:n relationship may be involved.

[0200] In a first sorting step 137, the respective dataset portions 16 can then be sorted as a result based on the second correlation value 19.

[0201] Subsequently, in a third evaluation step 138, a third relevance can be determined for each dataset 15 selected in the first selection step 134, wherein a third relevance value 20 can be assigned to each of these datasets 15. In particular, it can be specified that the third relevance value 20 is based on a weighted calculation of a second relevance value 19 for each dataset portion 16 of the corresponding dataset 15.

[0202] In a second sorting step 139, the individual datasets 15 can then be sorted according to a third relevance value 20 assigned to the datasets.

[0203] In a display step 140, each sorted dataset 15 can then be displayed as the first-ranked hit dataset 113a, the second-ranked hit dataset 113b, the third-ranked hit dataset 113c, etc., as already combined. Figure 3 or Figure 4As described in detail. The consistency value 117 shown here for the corresponding displayed hit dataset 113 can be a mapping or corresponding display of the third relevance value 20, or a direct numerical representation of the third relevance value 20. The displayed hit dataset position consistency values ​​123, 124 can be a mapping or corresponding display of the second relevance value 19 for the corresponding search dataset portions 13, 103, or a direct numerical representation of the second relevance value 19 for the corresponding search dataset portions 13, 103. The consistency values ​​118, 119, 120 can be reproduced according to the above embodiment, showing that the dataset portion 16, which is ranked as the most relevant in the first sorting step 137—displayed as the first hit dataset position 121—has the highest second relevance value 19.

[0204] In a minor change step 141, a small change in the user input can be considered. This can cause the described process steps to be recalculated or executed from the first comparison step 135.

[0205] In a major change step 142, a major change in the user input may be considered. This may cause the described process steps to be recalculated or executed from the first configuration step 131.

[0206] For example, adding a new feature to the search dataset 12 could be a minor change step 141. Slight modifications to the search dataset 12 could also be a minor change step 141. Manually inputting data into one of the evaluation elements 126, 127, and 128 could also be a minor change step 141.

[0207] For example, removing a feature from the search dataset 12 could be a major change step 142. A major modification to the search dataset 12 could also be a major change step 141.

[0208] The embodiments illustrate possible implementation variations, wherein it should be noted that the invention is not limited to the particularly shown implementation variations, but rather various combinations of the various implementation variations are possible, and such variations are based on the ability of those skilled in the art to be taught through the technical actions of the invention.

[0209] The scope of protection is determined by the claims. However, the specification and drawings should be used to interpret the claims. A single feature or combination of features from the different embodiments shown and described can constitute an independent inventive solution on its own. The task on which such an independent inventive solution is based can be derived from the specification.

[0210] All descriptions of numerical ranges in this specification shall be understood to include any and all subranges derived therefrom. For example, the description of 1 to 10 shall be understood to include all subranges starting from the lower limit of 1 and the upper limit of 10, that is, all subranges starting from the lower limit of 1 or greater and ending at the upper limit of 10 or less, such as 1 to 1.7, or 3.2 to 8.1, or 5.5 to 10.

[0211] For clarity, it should be noted that, in order to better understand the structure, some components are not shown to scale and / or enlarged and / or reduced.

[0212] List of reference numerals

[0213] 1 System

[0214] 2 processors

[0215] 3. Central Processing Unit

[0216] 4. Graphics Processor

[0217] 5. User Interface

[0218] 6. Management Procedures

[0219] 7. Database

[0220] 8-matrix

[0221] 9. Infrastructure

[0222] 10 Vector Embedding

[0223] 11 Large-scale language models

[0224] 12 Search Datasets

[0225] 13 Search Dataset Section

[0226] 14. Data Stock

[0227] 15 Datasets

[0228] 16. Dataset Section

[0229] 17 First correlation value

[0230] 18. Correlation threshold

[0231] 19 Second correlation value

[0232] 20 Third correlation value

[0233] 21 Results

[0234] 22 texts

[0235] 23 Videos

[0236] 24. Attached Figure

[0237] 25 CAD files

[0238] 26 images

[0239] 27 audio files

[0240] 28 Computer Programs

[0241] 29. Inference Engine

[0242] 30 Storage media

[0243] 31 Interface

[0244] 32 Structural Formula

[0245] 101 Output Device

[0246] 102 Search Dataset

[0247] 103 First Search Dataset Part

[0248] 104 Second Search Dataset Part

[0249] 105 Third Search Dataset Part

[0250] 106 Partial Hierarchy of the First Search Dataset

[0251] 107 Search dataset partial identifiers

[0252] 108 input elements

[0253] 109 Separating element

[0254] 110 Adaptation Fields

[0255] 111 Delete component

[0256] 112 Input Field

[0257] 113 hit dataset

[0258] 114 First Hit Dataset Part

[0259] 115 Second Hit Dataset Part

[0260] 116 Third Hit Dataset Part

[0261] 117 Consistent Value

[0262] 118 First Consistent Value Part

[0263] 119 Second Consistent Value Part

[0264] 120 Third Consistent Value Part

[0265] 121 First hit dataset location

[0266] 122 Second hit dataset location

[0267] 123 First Hit Dataset Location - Consistency Value

[0268] 124 Second Hit Dataset Location - Consistency Value

[0269] 125 Position Evaluation Component

[0270] 126 Evaluation Components

[0271] 127 First Evaluation Component

[0272] 128 Second Evaluation Component

[0273] 129 Second Search Dataset Partial Hierarchy

[0274] 130 Search Dataset Creation Steps

[0275] 131 First configuration step

[0276] 132 Second configuration step

[0277] 133 First evaluation step

[0278] 134 First Choice Step

[0279] 135 First comparison step

[0280] 136 Second evaluation step

[0281] 137 First sorting step

[0282] 138 Third evaluation step

[0283] 139 Second sorting step

[0284] 140 Show steps

[0285] 141 Small Changes Steps

[0286] 142 Major Changes

Claims

1. A computer-implemented method for searching a dataset corresponding to a search dataset in a data store, the method comprising the following steps; The search is performed starting from a search dataset, which includes a search dataset portion, wherein... The search dataset components are semantically, graphically, or technically related. Configure vector embeddings in the search dataset and / or a portion of the search dataset; Vectors are embedded in the data store and configured in the dataset and / or dataset portion stored therein; The first relevance is determined based on comparison with the search dataset, and the first relevance value of the dataset is used for evaluation. The individual dataset portions of the dataset having a first relevance value exceeding a first relevance threshold are compared, wherein the comparison is performed separately with the semantically or technically related portions of the search dataset. Based on comparisons with the search dataset portions of the search dataset respectively, a second relevance is identified for each dataset portion of a dataset and evaluated using the second relevance value; and The results are obtained by sorting and processing the data in each dataset based on the second correlation value.

2. The method according to claim 1, further comprising the following method steps: A third correlation is identified for the dataset and evaluated using a third correlation value, which is a weighted calculation based on the second correlation values ​​of the various dataset portions of the dataset; and Based on the third relevance value, the search and evaluation results of the dataset and the dataset portion to which it belongs are stored, and data processing is performed based on the third relevance value, or the results are displayed in a sorted manner in the user interface based on the third relevance value.

3. The method according to claim 1 or 2, wherein, The vector embedding is a numerical representation of text, words, phrases, images, or documents.

4. The method according to any one of claims 1 to 3, wherein, The vector embeddings used to search for parts of the dataset and / or parts of the dataset are generated through machine learning processes or by utilizing large language models.

5. The method according to any one of claims 1 to 4, wherein, The vector embeddings used for the dataset portion are performed independently of the search during the preparation process.

6. The method according to any one of claims 1 to 5, wherein, In terms of vector embedding for a portion of the dataset, further information about the origin and association of that portion of the dataset within the dataset or adjacent portions of the dataset is stored.

7. The method according to any one of claims 1 to 6, wherein, The vector embedding is either sentence embedding or word embedding.

8. The method according to any one of claims 1 to 7, wherein, The vector embedding includes context.

9. The method according to any one of claims 1 to 8, wherein, The search dataset and / or a portion of the search dataset and / or the dataset and / or a portion of the dataset. Formulas, structural expressions, text, videos, diagrams, CAD files, images, or audio files Or a part thereof.

10. The method according to any one of claims 1 to 9, wherein, The search is a hybrid search using both keyword and vector search.

11. The method according to any one of claims 1 to 10, wherein, On the user interface, select a portion of the document pool to search the dataset using prompt-based search.

12. The method according to any one of claims 1 to 11, wherein, The data storage and the vector embedding are stored in one or more distributed databases.

13. The method according to any one of claims 1 to 12, wherein, Lexicons are created from the search dataset portion and the dataset portion, and the search dataset portion and the dataset portion are indexed using an inference engine.

14. The method according to any one of claims 1 to 13, wherein, After the initial hit results are displayed on the user interface for an extended period of time, additional search and evaluation results for the dataset or a portion of the dataset are asynchronously reloaded.

15. The method according to any one of claims 1 to 14, wherein, The multidimensional matrix is ​​displayed in two dimensions on the user interface, such that each part of the search dataset is displayed in one row of the matrix along with multiple different parts of the dataset.

16. The method according to any one of claims 1 to 15, wherein, Multiple dataset portions are selectively displayed individually, side-by-side, or overlapping in the matrix columns of the respective datasets on the user interface, but in the same matrix rows as the first search dataset portion to be compared.

17. The method according to any one of claims 1 to 16, wherein, Changes to the search dataset trigger the determination of new relevance values ​​for the second and / or third relevance values.

18. The method according to any one of claims 1 to 17, wherein, Manually changing the second correlation value through user interaction on the user interface triggers the determination of a new correlation value for the third correlation value.

19. The method according to any one of claims 1 to 18, wherein, Store portions of the search dataset or changes to portions of the dataset, and use them for fine-tuning of other machine learning or search models.

20. The method according to any one of claims 1 to 19, wherein, Data sets and / or dataset segments are stored and / or displayed on a user interface in multiple different languages ​​for a dataset and a vector embedding.

21. The method according to any one of claims 1 to 20, wherein, The search dataset is a patent claim.

22. The method according to any one of claims 1 to 21, wherein, The search dataset portion and the dataset portion are individual features in the patent claims, patent specification, or patent drawings.

23. The method according to any one of claims 1 to 22, wherein, The vector embedding is performed using a graphics processor.

24. The method according to any one of claims 1 to 20, wherein, The search dataset is a description of the relationships between legal cases.

25. The method according to any one of claims 1 to 24, characterized in that, Argumentative texts are written using large language models based on at least one of the datasets found in the search.

26. The method according to any one of claims 1 to 25, characterized in that, The dataset is divided into groups, wherein the following datasets are assigned to the corresponding groups, wherein for each individual search dataset portion of the search dataset, none of the second relevance values ​​exceed a determined second relevance threshold.

27. The method according to claim 26, characterized in that, To sort the groups, the following search dataset portions are selected with the help of user input, and for these search dataset portions, none of the second relevance values ​​are allowed to exceed a defined second relevance threshold.

28. The method according to any one of claims 26 to 27, characterized in that, If, for all portions of the search dataset, there is no second relevance value in the dataset that exceeds a determined second relevance threshold, the dataset with the most second relevance values ​​or the highest third relevance value is assigned to the first group, and the dataset in which the missing portions of the first group's dataset are located is assigned to the second group.

29. The method according to any one of claims 1 to 28, characterized in that, The amount of data available for the search has been limited from the globally available data pool.

30. A system comprising a processor configured to implement the method according to any one of claims 1 to 29.

31. A computer program comprising commands that, when executed by a computer, cause the computer to perform the method according to any one of claims 1 to 29.

32. A computer-readable storage medium having instructions that, when executed by a computer, cause the computer to perform the method according to any one of claims 1 to 29.

33. An interface configured to access the computer program of claim 31 and to pass results, a search dataset, a portion of the search dataset, the dataset, and the portion of the dataset to a second computer program product for further processing.