OLLAMA-based data center clothing management method

Abstract

The invention relates to the technical field of data processing, in particular to a data center clothing management method based on OLLAMA, and the method comprises the following steps: collecting and calculating resource state data; and executing predictive semantic preprocessing: when the computing resource state data indicates that the graphics processor is in a preset low-load state, selecting to-be-pre-computed unstructured clothing data according to query mode information obtained by user query log analysis, and pre-computing the to-be-pre-computed unstructured clothing data according to the selected to-be-pre-computed unstructured clothing data. Driving an OLLAMA model to perform semantic analysis on the unstructured clothing data to be pre-calculated to obtain a pre-calculated semantic analysis result, and storing the pre-calculated semantic analysis result in a semantic cache; processing a user query request; through predictive semantic preprocessing, idle time of computing resources can be fully utilized, and a time-consuming semantic analysis task can be completed in advance, so that the delay of real-time query is remarkably reduced, and the response speed of a system is improved.

Description

[0001] The present application is a divisional application of the application with application number 2025113514355 and titled "Data center clothing management method and system based on OLLAMA", filed on September 22, 2025. TECHNICAL FIELD

[0002] The present application relates to the technical field of data processing, and particularly relates to a data center clothing management method based on OLLAMA. BACKGROUND

[0003] In the fast-changing fashion industry, enterprise data centers need to process massive and diversified clothing data, including structured information such as sales and inventory, and unstructured text such as design manuscripts, market reports, and social media comments. To extract business insights from them, enterprises introduce solutions based on OLLAMA models, but this solution faces many challenges: rapid fashion vocabulary updates, data cleaning lags, frequent manual intervention, unstable model reasoning performance, and inconsistent output, which seriously affect the real-time and accuracy of data analysis and restrict the enterprise's ability to respond quickly to the market.

[0004] Specifically, the rapid evolution of fashion terminology and the differences in expression of consumers in different regions pose significant challenges to traditional cleaning methods based on preset dictionaries and grammar patterns, making it difficult to accurately identify new terminology and multilingual mixed text. The lag in data cleaning makes the quality of data input into the OLLAMA model unstable, which in turn affects the depth of semantic understanding and the accuracy of analysis results.

[0005] To compensate for the shortcomings of the data cleaning link, data analysts have to adopt a compensatory strategy: by adjusting the wording strategy of the OLLAMA model questions more frequently and more finely, in order to obtain valuable information from incomplete or noisy raw data. This behavior of manually repeating trial and error and adjusting prompt words not only consumes a large amount of human resources, but also significantly lengthens the single response time of the OLLAMA model interface call and result processing flow due to the complex reasoning and context understanding of the OLLAMA model in each interaction, which cannot meet the real-time analysis demand.

[0006] In addition, as the market analysis team relies more and more on the OLLAMA model, and the parallel query strategy is adopted to deal with the above-mentioned response delay, the server graphics processors responsible for the inference of the OLLAMA model in the data center are in a long-term high-load operation state. This persistent high-intensity computing requirement causes the temperature of part of the server graphics processor cores to continuously rise and frequently touch its heat dissipation limit. When the temperature of the server graphics processor is too high, the hardware will automatically trigger the thermal throttling mechanism, that is, reduce its working frequency to prevent overheating damage. This mechanism causes the computing performance to periodically decline and recover, showing an unstable fluctuation, rather than a simple linear decline in performance. This performance fluctuation further exacerbates the delay of data processing and increases the uncertainty of the output results of the OLLAMA model.

[0007] In view of the above problems, the prior art needs to be improved. SUMMARY

[0008] The purpose of the present application is to solve the problems existing in the prior art and provide an OLLAMA-based data center clothing management method.

[0009] In a first aspect, the present application provides an OLLAMA-based data center clothing management method, which comprises the following steps: Collecting computing resource state data; Performing predictive semantic preprocessing, including: When the computing resource state data indicates that the graphics processor is in a preset low-load state, selecting unstructured clothing data to be precomputed according to the query pattern information obtained by analyzing the user query log, and driving the OLLAMA model to perform semantic analysis on the unstructured clothing data to be precomputed to obtain a pre-computed semantic analysis result, and storing the pre-computed semantic analysis result in a semantic cache; Processing a user query request, including: Receiving a user query request and searching in the semantic cache according to the user query request; Generating a query response according to the search result.

[0010] In a second aspect, an OLLAMA-based data center clothing management system is provided, which comprises: A collection module for collecting computing resource state data; A preprocessing module for performing predictive semantic preprocessing, including: When the computing resource state data indicates that the graphic processor is in a preset low load state, according to query mode information obtained by analyzing a user query log, non-structured clothing data to be pre-computed is selected, and an OLLAMA model is driven to perform semantic analysis on the non-structured clothing data to be pre-computed, so as to obtain a pre-computed semantic analysis result, and the pre-computed semantic analysis result is stored in a semantic cache; The query processing module is configured to process a user query request, and includes: The query processing module is configured to process a user query request, and includes: The query processing module is configured to process a user query request, and includes:

[0011] Compared with the prior art, the present application has the following beneficial effects: Through predictive semantic preprocessing, the present application can make full use of the idle time of computing resources, complete time-consuming semantic analysis tasks in advance, thereby significantly reducing the delay of real-time queries and improving the response speed of the system. At the same time, the introduction of the semantic cache reduces repeated calls to the OLLAMA model, reduces the real-time computing load of the graphic processor, avoids heat throttling and performance fluctuations caused by high load operation, and improves the operation stability of the model. In addition, through the automatic preprocessing and caching mechanism, the dependence on manual adjustment of the model by data analysts is reduced, the labor cost is reduced, and the real-time performance and accuracy of data analysis are improved, so that enterprises can respond to market changes more quickly. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 The present application is a method flowchart.

[0013] Figure 2 The present application is a system structure schematic diagram.

[0014] In the figure: 201, acquisition module; 202, preprocessing module; 203, query processing module. DETAILED DESCRIPTION

[0015] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the accompanying drawings are exemplary and are only used to explain the present application, and cannot be understood as a limitation of the present application.

[0016] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0017] To better understand the technical solution proposed in this application, the key terms involved are explained first. The OLLAMA model refers to a large-scale language model deployed and run in a data center environment, primarily used for semantic understanding, analysis, and generation of unstructured text data. The Graphics Processing Unit (GPU) is the core computing hardware upon which the OLLAMA model relies, and its workload directly affects the model's inference performance. The semantic cache is an area storing pre-computed semantic analysis results, designed to accelerate the response to subsequent user queries by trading space for time. Unstructured clothing data refers to clothing-related information that is not stored in a fixed format, such as design sketches, market reports, and social media comments. Query pattern information is extracted from historical user query logs, revealing patterns such as user query habits, high-frequency query terms, and query themes.

[0018] like Figure 1 The illustrated method for managing clothing in a data center based on OLLAMA includes the following steps: S101. Collect computing resource status data; It's important to note that the purpose of this step is to obtain real-time information about the OLLAMA model's operating environment within the data center. For example, a monitoring agent deployed on the server can periodically collect data such as the graphics processor's workload, temperature, and the response time of the OLLAMA model during inference tasks. This data can be stored in log files or transmitted in real-time to a central monitoring system via an API interface. Alternatively, a resource monitoring module can be integrated into the OLLAMA model scheduler to automatically record and summarize relevant resource usage before and after each model inference task.

[0019] S102. Perform predictive semantic preprocessing, including: When the computing resource status data indicates that the graphics processor is in a preset low load state, the unstructured clothing data to be pre-computed is selected based on the query pattern information obtained from the user query log analysis, and the OLLAMA model is driven to perform semantic analysis on the unstructured clothing data to be pre-computed, so as to obtain the pre-computed semantic analysis results and store the pre-computed semantic analysis results in the semantic cache. It should be noted that when the collected computing resource state data indicates that the graphics processor is in a preset low load state, the system intelligently selects the unstructured clothing data to be precomputed according to the query mode information obtained by analyzing the user query log. For example, if the historical query log shows that the user often queries information about “summer new dress”, and the current graphics processor load is low, the system will preferentially select the design manuscripts, market reports and other unstructured data related to “summer new dress” for preprocessing. Subsequently, the system drives the OLLAMA model to perform semantic analysis on the selected unstructured clothing data to be precomputed, and obtains pre-computed semantic analysis results. These results are then stored in the semantic cache for subsequent fast retrieval. As an alternative, the query mode information can also be used to predict user behavior through a machine learning model, so as to more accurately identify potential high-frequency query content.

[0020] The unstructured clothing data includes, for example, fabric 3D effect data in the digital sample room, clothing 3D effect data, pattern / design data, real person fitting map data generated by clothing modeling, fitting dynamic video data, and fabric basic information text and inventory quantity associated description text in the fabric management module. For the embodiment of “unstructured clothing data type limitation”: Assuming that a certain clothing data center accesses multiple data sources. In the digital sample room module, the designer generates 3D effect data of a summer dress through 3D modeling software, including the three-dimensional version of the clothing, the dynamic wrinkle effect of the material, etc.; At the same time, multiple pattern data are created for the design of the dress. In terms of fabric management, the system inputs the basic information of the fabric used to make the dress, such as fabric composition (60% cotton, 40% polyester fiber), grammage, width, etc. Text description and real-time updated fabric inventory data are recorded in the inventory table in the form of text. In addition, during the clothing modeling process, the AI generates fitting map data and dynamic show video data of different body models of the dress. When the data center performs predictive semantic preprocessing, the 3D effect data in the digital sample room, the pattern data, the basic information text of the fabric management, the inventory data, and the AI-generated fitting map and video data are all unstructured clothing data, which are included in the range to be precomputed, and drive the OLLAMA model to perform semantic analysis.

[0021] S103, processing a user query request, comprising: receiving a user query request and searching in the semantic cache according to the user query request; generating a query response according to the search result.

[0022] It should be noted that this step first receives the user's submitted query request. For example, the user may input "the latest fashion trend" or "a certain brand's autumn and winter series" and the like through the front-end interface. After the system receives these requests, it will first search in the semantic cache. If the semantic cache contains a pre-computed semantic analysis result that highly matches the user's query request, the result is directly extracted as the query response, thereby avoiding real-time calling of the OLLAMA model for reasoning, significantly shortening the response time. If no pre-computed result meeting the condition is found in the semantic cache, or the real-time reasoning result has a confidence lower than a preset threshold, the system will start the real-time reasoning process and generate a query response according to the reasoning result. For example, if the user queries "environmentally friendly fabric clothing", the system first searches in the cache, and if no match is found, the OLLAMA model is driven in real time to analyze the relevant data, and the analysis result is returned.

[0023] The overall working principle of the present application is to predictively pre-process potential user query content by prospectively utilizing the idle period of data center computing resources, thereby constructing an efficient semantic cache. When a user submits a query request, the system first attempts to quickly obtain a pre-computed result from the semantic cache to meet the query with higher real-time requirements. This mechanism effectively alleviates the computing pressure faced by the OLLAMA model during peak periods and reduces the response time of model reasoning. Specifically, by continuously collecting computing resource state data, the system can intelligently identify the low-load period of the graphics processor. In these periods, the system combines deep analysis of user query patterns to accurately select unstructured clothing data related to high-frequency or potentially high-frequency queries. Subsequently, the OLLAMA model is driven to perform semantic analysis on these data, and the obtained pre-computed semantic analysis result is stored in the semantic cache. When a user query request arrives, the system first searches in the semantic cache. If a hit is found, the pre-computed result is directly returned, greatly accelerating the response process. If a hit is not found or the real-time reasoning result has insufficient confidence, the system will start real-time reasoning and may combine a multi-model arbitration mechanism to ensure the accuracy and consistency of the query response. This collaborative working method enables each technical feature to work closely together to solve the problems of data processing lag, long response time, unstable model performance, and inconsistent results in the prior art, significantly improving the efficiency and accuracy of the data center clothing management method.

[0024] Compared with the prior art, the core innovation of the present application is to introduce a predictive semantic preprocessing mechanism based on the state of computing resources. Traditional methods often start the inference process of the OLLAMA model only after receiving a user query request, which leads to a long response time when processing massive unstructured data, especially when the graphics processor is heavily loaded, the performance bottleneck is more prominent. The present application effectively avoids this problem by performing semantic analysis on the unstructured clothing data that the user may query in advance and caching the results when the graphics processor is in a preset low load state. For example, during the night or during the low business peak period, the system can utilize the idle computing resources to preprocess the latest fashion trend reports, social media comments, etc. The semantic analysis results of high-frequency queries such as "2024 spring and summer popular colors" and "sustainable fashion brands" are pre-stored. When users query these contents during the daytime peak period, the system can directly extract the results from the semantic cache without real-time calling the OLLAMA model for time-consuming inference, thereby shortening the query response time from several seconds or even tens of seconds to milliseconds. This forward-looking processing method not only significantly improves the user experience, but also optimizes the utilization efficiency of data center computing resources, reduces energy consumption and hardware wear caused by real-time model inference. In addition, through the analysis of user query logs, the present application can more accurately predict user demand, making the preprocessing process more targeted and avoiding the waste of resources caused by blind pre-computation. Therefore, the present application exhibits significant progress in real-time performance, efficiency and resource utilization.

[0025] As an embodiment of the present application, the computing resource state data includes workload data and temperature data of the graphics processor on which the OLLAMA model depends, and the response time of the OLLAMA model to complete the inference task.

[0026] The workload data of the graphics processor refers to the busy degree of the graphics processor in processing tasks within a certain period of time, usually expressed in percentage, reflecting the utilization rate of its computing resources. The temperature data refers to the real-time temperature of the graphics processor during operation, which is a key indicator for evaluating hardware health and overheating risk. The response time of the OLLAMA model to complete the inference task measures the time required for the OLLAMA model to complete and return the results after receiving the inference request, directly reflecting the processing efficiency of the model and the real-time performance of the system. These data together constitute a comprehensive monitoring of the running environment and performance of the OLLAMA model.

[0027] The scheme of the present application can monitor the usage and health status of the hardware resources relied on by the OLLAMA model in real time by collecting workload data and temperature data of the graphics processor. When the workload data indicates that the graphics processor is at a low utilization level and the temperature data is within a normal range, the graphics processor can be more accurately determined to be in a preset low load state. At the same time, by monitoring the response time of the OLLAMA model to complete the inference task, the actual performance of the model under the current load can be evaluated, thereby providing a more reliable basis for triggering the predictive semantic preprocessing. This multi-dimensional data collection method enables the system to more intelligently identify resource idle opportunities and optimize the scheduling of preprocessing tasks.

[0028] As an embodiment of the present application, the step of generating a query response according to the search result comprises: If the search result is that there is a pre-computed semantic analysis result in the semantic cache that meets the user query request, the pre-computed semantic analysis result is extracted as the query response; If the search result is that there is no pre-computed semantic analysis result in the semantic cache that meets the user query request, or the confidence of the result inferred by the OLLAMA model in real time is lower than a preset threshold, the user query request is submitted to multiple OLLAMA model instances in parallel, and the output results of the multiple OLLAMA model instances are received; It should be noted that the multiple OLLAMA model instances can be OLLAMA models configured with different parameters, different versions, or fine-tuned on different data sets, or they can be multiple identical OLLAMA model copies simply deployed on different computing resources to achieve redundancy and diversity. By submitting in parallel, independent inference results of multiple models can be obtained at the same time.

[0029] The output results of the multiple OLLAMA model instances are compared, and a semantic analysis result with high consistency and high confidence is selected as the final query response according to a preset arbitration rule.

[0030] It should be noted that the output results of the multiple OLLAMA model instances are compared. This comparison process aims to evaluate the consistency and difference between the outputs of different models. Based on the comparison result, the system will select a semantic analysis result with high consistency and high confidence from these outputs as the final query response according to a preset arbitration rule. The preset arbitration rule can include but is not limited to a majority voting mechanism, weighted average, weight distribution based on historical evaluation of model performance, or decision-making combined with an expert system. The purpose is to effectively avoid the bias, error or "hallucination" phenomenon that may exist in a single model by combining the wisdom of multiple models, thereby improving the quality and reliability of the final query response.

[0031] The scheme of the present application effectively solves the low confidence problem that may occur during real-time reasoning of a single OLLAMA model and the response quality problem when the semantic cache is not hit, by introducing a multi-OLLAMA model instance parallel reasoning and arbitration mechanism. Specifically, when the semantic cache cannot provide an effective response or the single model reasoning result is unreliable, the system is no longer passive waiting or directly outputs an uncertain result, but actively utilizes the parallel processing capability of multiple OLLAMA model instances. By simultaneously obtaining independent insights from multiple models and comparing and arbitrating these insights, the accuracy of the semantic analysis result can be verified from multiple dimensions. A result with high consistency usually means higher reliability, while a result with high confidence directly reflects the model's confidence in its own judgment. This mechanism is essentially an application of ensemble learning or consensus mechanism, which significantly reduces the risk of error propagation of a single model through the "collective wisdom" approach, thereby ensuring that in complex or uncertain scenarios, a high-quality and highly reliable query response can still be generated.

[0032] Through the above technical solutions, the present application can significantly improve the robustness and accuracy of the data center clothing management method in processing user query requests. Especially in the case of semantic cache miss or low confidence of OLLAMA model real-time reasoning result, by parallel use of multiple OLLAMA model instances and intelligent arbitration, errors or inaccurate responses caused by the limitations of a single model can be effectively avoided. This not only improves the reliability and consistency of query response and reduces the probability of "hallucination" phenomenon, but also provides users with a more stable and reliable query service experience, thereby improving the intelligent level and practical value of the entire data center clothing management system.

[0033] In some preferred embodiments, the following is described by a specific example. Suppose the user query is "Please recommend a breathable and quick-drying T-shirt suitable for summer outdoor sports". The system first searches in the semantic cache, but fails to find a completely matching precomputed semantic analysis result. At this time, if only one OLLAMA model is relied on for real-time reasoning, the model may give a low-confidence recommendation due to training data bias or query complexity, for example, recommending a "cotton T-shirt", which contradicts the "quick-drying" requirement.

[0034] According to the scheme of the present application, the system will submit the user query request to three OLLAMA model instances in parallel.

[0035] The first OLLAMA model instance may recommend "a certain brand CoolMax quick-drying T-shirt" with high confidence.

[0036] The second OLLAMA model instance may recommend "a certain brand sweat-wicking and breathable sports T-shirt" with high confidence.

[0037] The third OLLAMA model instance might recommend a "certain brand cotton casual T-shirt", but with a lower confidence level.

[0038] After receiving these three output results, the system will compare them. Through comparison, it is found that the recommendations of the first two model instances focus on the core semantic features of "quick-drying", "breathable", "sports", etc., and have high confidence levels, showing high consistency. The recommendation of the third model instance deviates significantly from the user's needs, and has a low confidence level. According to the pre-set arbitration rules (for example, selecting the majority of high-confidence and semantically consistent results), the system will select the "quick-drying, breathable sports T-shirt" recommended by the first two model instances as the final query response, and may integrate the advantages of both or provide more detailed recommendation information. As a result, even in the case of cache miss or single model error, the system can provide accurate, reliable and user-expected query responses.

[0039] As an embodiment of the present application, the step of comparing the output results of multiple OLLAMA model instances and selecting a semantic analysis result with high consistency and high confidence as the final query response according to the pre-set arbitration rules comprises: Capturing the semantic analysis output of multiple OLLAMA model instances and converting it into a unified format; It should be noted that it refers to collecting the semantic analysis results generated by each parallel running OLLAMA model instance for the same user query request. Since different OLLAMA model instances may use different internal representations or output structures, in order to effectively compare and analyze later, these original outputs need to be standardized into a unified, parsable data format. For example, all outputs can be converted into JSON, XML or specific semantic triple formats to ensure data consistency and operability. The purpose is to provide a standardized input basis for subsequent consensus and divergence quantification.

[0040] Consensus and divergence quantification of the unified format semantic analysis output to obtain entity consensus degree, sentiment polarity divergence degree and abstract content similarity; It should be noted that it can be understood as a fine-grained semantic comparison of the standardized model output. The entity consensus degree is used to measure the consistency of different models in identifying and mentioning the same entity (for example, a specific clothing brand, style, material, etc.). The sentiment polarity divergence degree evaluates the differences in the judgment of the sentiment tendency (for example, positive, negative, neutral) of a particular entity or concept by different models. The abstract content similarity compares the text abstract or key information of the model output to quantify the similarity in core content expression. The purpose of these quantitative indicators is to deeply understand the similarities and differences of model output from multiple dimensions.

[0041] Based on the consensus and divergence quantification results, a semantic graph is constructed, which presents semantic consensus areas and semantic divergence areas, and associates underlying entities with abstract concepts. It should be noted that this means that the entities, concepts and their relationships in the model output are visualized and structured in the form of a graph using the above quantification indicators. The nodes in the semantic graph can represent entities or abstract concepts, and the edges represent their relationships. The consensus area refers to the highly consistent semantic information in multiple model outputs, while the divergence area highlights the differences or conflicts between models. By associating underlying entities (such as "dress", "cotton and hemp") with abstract concepts (such as "wearing in summer", "casual style"), a more comprehensive understanding of semantic content can be achieved. The purpose is to provide an intuitive and structured view to facilitate the identification and analysis of commonalities and individualities in model outputs.

[0042] The consensus and divergence quantification indicators are attached to the elements of the semantic graph as metadata; It should be noted that this means that the entities, concepts and their relationships in the model output are visualized and structured in the form of a graph using the above quantification indicators. The nodes in the semantic graph can represent entities or abstract concepts, and the edges represent their relationships. The consensus area refers to the highly consistent semantic information in multiple model outputs, while the divergence area highlights the differences or conflicts between models. By associating underlying entities (such as "dress", "cotton and hemp") with abstract concepts (such as "wearing in summer", "casual style"), a more comprehensive understanding of semantic content can be achieved. The purpose is to provide an intuitive and structured view to facilitate the identification and analysis of commonalities and individualities in model outputs.

[0043] According to the preset arbitration rules and the elements of the semantic graph, the semantic analysis results with high consistency and high confidence are extracted from the semantic graph as the final query response.

[0044] It should be noted that this means that in the arbitration stage, the system no longer relies solely on simple text matching or voting mechanisms, but rather considers the consensus areas, divergence areas, associations between underlying entities and abstract concepts, and additional consensus and divergence quantification metadata presented in the semantic graph. The preset arbitration rules can make more intelligent and accurate final decisions based on these rich information, such as prioritizing entities with high consensus, analyzing divergence areas in depth to identify their nature, and combining the model's own confidence.

[0045] The scheme of the present application effectively solves the problems of insufficient robustness and accuracy that may exist when comparing the output results of multiple OLLAMA model instances by introducing multi-dimensional consensus and divergence quantification and semantic graph construction. First, by converting the semantic analysis outputs of different model instances into a unified format, the standardization and comparability of subsequent analysis are ensured. Second, by quantifying the entity consensus degree, sentiment polarity divergence degree, and summary content similarity, the system can deeply understand the similarities and differences between model outputs from multiple fine-grained dimensions, rather than just superficial text matching. It is precisely due to this detailed quantitative analysis that the system can identify the commonalities and differences between models in specific entities, sentiment tendencies, and core content. On this basis, a semantic graph is constructed and the quantitative indicators are attached as metadata to the graph elements, so that the semantic information of the model output is structured, visualized, and rich in contextual information. The semantic graph can clearly present the semantic consensus area and the semantic divergence area, and associate the underlying entities with abstract concepts, thereby providing a more comprehensive and intuitive decision basis for the arbitration process. Finally, by combining the pre-set arbitration rules with these rich semantic graph elements, the system can more accurately identify and extract the semantic analysis results with high consistency and high confidence, thereby avoiding misjudgment or information omission due to simple comparison.

[0046] Through the above technical scheme, the present application can significantly improve the accuracy, robustness, and explainability of the semantic analysis results of the data center clothing management method when processing complex user queries. Compared with simple output comparison, the present application can more deeply understand the subtle differences and potential conflicts between the outputs of different OLLAMA model instances through detailed consensus and divergence quantification, thereby avoiding arbitration errors caused by incomplete information or understanding bias. In addition, the construction of the semantic graph not only provides an intuitive semantic representation, but also associates the underlying entities with abstract concepts, enabling the system to better understand the deep intent of the user query and the internal logic of the model output. The additional metadata further enriches the graph information, enabling the arbitration rules to make decisions based on a more comprehensive context. As a result, the final query response not only has higher consistency and confidence, but also has a more transparent and traceable generation process, greatly enhancing the intelligence level of the system and the user experience.

[0047] As an embodiment of the present application, based on the consensus and divergence quantification results, the step of constructing a semantic graph that presents semantic consensus areas and semantic divergence areas and associates underlying entities with abstract concepts includes: Based on the consensus and divergence quantification results, identifying new abstract concepts or entities appearing in the semantic analysis output, as well as semantic consensus areas and semantic divergence areas; According to the identified emerging abstract concept or entity, as well as the semantic consensus area and the semantic divergence area, the nodes and associated edges of the semantic graph are expanded, and the bottom-level entities are associated with the abstract concept.

[0048] Specifically, in the process of constructing the semantic graph, first, the semantic analysis output of the OLLAMA model instance needs to be analyzed in depth. Among them, "identifying the emerging abstract concept or entity in the semantic analysis output" means that the system discovers and extracts the words or phrases that do not exist in the existing semantic graph and have independent semantic meaning from the text processed by consensus and divergence quantification through natural language processing (NLP) techniques such as named entity recognition (NER), keyword extraction, and topic modeling. These emerging words or phrases can be specific clothing brands, styles, materials, etc. entities, or fashion trends, styles, occasions, etc. abstract concepts. At the same time, "identifying semantic consensus areas and semantic divergence areas" means that according to the previously quantified entity consensus degree, sentiment polarity divergence degree, and abstract content similarity, etc. indicators, determine which parts of the semantic analysis output are agreed upon by multiple OLLAMA model instances (consensus area), and which parts are significantly different or contradictory (divergence area).

[0049] Further, "extending the nodes and associated edges of the semantic graph" means adding the identified emerging abstract concept or entity as a new node to the semantic graph. At the same time, according to the semantic relationship between these new nodes and existing nodes, such as hierarchical relationship, synonymous relationship, causal relationship, attribute relationship, etc., create or update the associated edges in the graph. For example, if the abstract concept "retro style" is identified, and it is found that it has a strong association with "bell-bottom pants", "dot skirt", etc. entities, then the corresponding nodes and edges will be added in the graph. In addition, "associating bottom-level entities with abstract concepts" means ensuring that the semantic graph not only contains high-level abstract concepts, but also contains specific, perceptible bottom-level entities, and connecting them through explicit association edges to form a multi-level, structured knowledge network. For example, associate the abstract concept "dress" with the specific material entities "silk dress", "cotton and hemp dress", or associate the abstract concept "summer clothing" with the specific clothing entities "short-sleeved T-shirt", "shorts".

[0050] The scheme of the present application makes the construction process of the semantic graph more dynamic and accurate through in-depth use of consensus and divergence quantization results. By identifying newly emerging abstract concepts or entities in the semantic analysis output, it can ensure that the semantic graph can continuously absorb and integrate the latest fashion information and user query patterns, avoiding the aging and information missing of the graph content. At the same time, by identifying semantic consensus areas and semantic divergence areas, the graph can be expanded in a targeted manner, with knowledge connections being strengthened in the consensus area and potential ambiguity or polysemy being marked in the divergence area, providing a basis for subsequent arbitration and explanation. Thus, by taking these newly identified elements as nodes and expanding associated edges according to their semantic relationships, a more comprehensive, detailed and multi-model output difference reflecting semantic graph can be constructed, providing richer and more accurate semantic context for subsequent query response generation.

[0051] As an embodiment of the present application, based on the consensus and divergence quantization results, the step of constructing a semantic graph presenting semantic consensus areas and semantic divergence areas and associating underlying entities and abstract concepts further comprises: According to the context information contained in the user query, in combination with the confidence of the association between the specific entity and the abstract concept in the current context, the strength of the associated edge in the semantic graph is adjusted; It should be noted that the context information can be understood as additional information such as the specific background, time, place, user preference, historical interaction record or current fashion trend of the user query. For example, when the user queries "summer dress", the context information may include the geographical location of the user (such as tropical regions), the current season (summer) and the user's past preference for light, breathable fabric. The confidence of the association between the specific entity and the abstract concept in the current context refers to the reliability or likelihood of the association between a certain entity (such as "cotton and linen fabric") and a certain abstract concept (such as "breathability") in a specific context. This confidence can be evaluated through historical data analysis, expert knowledge or machine learning models. Adjusting the strength of the associated edge in the semantic graph means dynamically increasing or decreasing the weight of the connection between nodes in the semantic graph according to the above context information and confidence, to reflect the strength of the relevance in the current context. For example, in the summer context, the associated edge strength between "dress" and "breathability" will be enhanced.

[0052] According to the user query intent, by identifying the core abstract concept of the query and performing multi-step semantic association, exploring and explicitating the deep association path connected through intermediate entities or secondary abstract concepts, a multi-level abstract concept association path in the semantic graph is constructed.

[0053] It should be noted that the user query intention refers to the real purpose or demand implied by the user when initiating the query, for example, when the user queries "business casual wear", the intention may be to find clothes suitable for office wear but not too formal. Identifying the core abstract concept of the query refers to extracting the high-level concept that best represents the user's intention from the user query text, for example, identifying the core concepts of "light jacket suitable for spring outing" such as "spring", "outing", "light", and "jacket". Multi-step semantic association refers to exploring and discovering other concepts or entities indirectly associated with a concept or entity in the semantic graph through multiple intermediate nodes and associated edges. For example, from "dress" to "fabric", then to "breathability", and then to "cotton and hemp". Exploring and explicitating the deep association path connected by intermediate entities or secondary abstract concepts refers to explicitly representing these indirect association paths discovered through multi-step association, so that they are visible and can be utilized in the semantic graph. Building multi-level abstract concept association paths in the semantic graph refers to establishing and maintaining connections between concepts at different abstraction levels in the semantic graph, such as from "clothing" to "top" to "shirt", forming a hierarchical association structure.

[0054] The scheme of the present application dynamically adjusts the association edge strength of the semantic graph by introducing context information, solving the problem of fixed association strength and lack of flexibility of traditional semantic graph in processing dynamic context. Specifically, when the user query request contains context information, the system can combine these context information to evaluate the association confidence of a specific entity and an abstract concept under the current context, and dynamically adjust the strength of the corresponding association edge in the semantic graph accordingly. For example, in the "summer" context, if the user queries "dress", the association strength between "dress" and "breathable fabric" will be significantly enhanced, while the association strength with "wool fabric" will be weakened, so that the semantic graph can more accurately reflect the semantic association under the current context.

[0055] In addition, by identifying the core abstract concept of the user query and performing multi-step semantic association, the present application can explore and explicitate deep association paths and build multi-level abstract concept association paths. This enables the system not only to identify direct associations, but also to discover indirect and hidden semantic relationships, thereby solving the problem of insufficient depth in understanding complex user intentions that may exist in traditional schemes. For example, when the user queries "clothing suitable for outdoor sports", the system can not only directly associate with "sports wear", but also through multi-step association, associate "outdoor sports" with "sun protection", "quick drying", and then associate with "sun protection clothes", "quick drying pants", and further associate with "light", "wear-resistant" and other attributes, thereby building a more comprehensive and more in line with the user's deep demand semantic association path.

[0056] As an embodiment of the present application, after the steps of consensus and divergence quantification of the semantic analysis output in unified format, the entity consensus degree, the sentiment polarity divergence degree and the abstract content similarity are obtained, including: According to the entity consensus degree, the sentiment polarity divergence degree and the abstract content similarity, the text segments with divergence in the semantic analysis output are identified; From the divergent text segments, the semantic features leading to divergence are extracted; According to the semantic features, the divergence is classified into preset divergence types; An explanatory description of the divergence type is generated and attached as metadata to the corresponding divergence area in the semantic graph.

[0057] Specifically, after the steps of consensus and divergence quantification of the semantic analysis output in unified format, the entity consensus degree, the sentiment polarity divergence degree and the abstract content similarity are obtained, the system will use these quantitative indicators as the basis to identify the text segments with significant differences or inconsistencies in the output of different OLLAMA model instances. For example, by setting a threshold, when the entity consensus degree of a certain text segment in different model outputs is lower than a certain threshold, or the sentiment polarity divergence degree is higher than a certain threshold, it is marked as a divergent text segment. Further, once the divergent text segments are identified, the specific semantic features leading to divergence are extracted from these segments. These semantic features can be words, phrases, sentence structures, or even implicit contextual information. For example, natural language processing (NLP) techniques such as keyword extraction, dependency syntax analysis, named entity recognition, etc. can be used to locate and extract these key semantic elements. The purpose is to accurately find out the root cause of the difference in understanding of different models. On this basis, according to the extracted semantic features, the identified divergence is classified into preset divergence types. These divergence types can be defined according to actual application scenarios, for example, they can be polysemy ambiguity, context-dependent divergence, regional cultural difference divergence, factual error divergence, and viewpoint difference divergence. Through classification, the divergence can be structured and managed, which is convenient for subsequent analysis and processing. Finally, for the classified divergence types, the corresponding explanatory description is generated. The description aims to clearly explain the nature of the divergence, the possible reasons and the impact on the overall semantic understanding. For example, for the "polysemy ambiguity" type, the explanatory description can point out which word has polysemy and how different models interpret it differently. These explanatory descriptions are then attached as metadata to the corresponding divergence area in the semantic graph, so that the semantic graph not only shows consensus and divergence, but also provides deep reasons and explanations for the divergence, greatly enhancing the diagnostic and explainability of the graph.

[0058] The scheme of the present application effectively solves the limitation that only the divergence can be quantified and the reason for the divergence cannot be understood deeply in the above technical basis by introducing the recognition of divergent text segments, semantic feature extraction, divergence type classification and explanatory description generation after consensus and divergence quantification. Specifically, when multiple OLLAMA model instances produce different outputs for the same user query request, the text segments with inconsistencies in the semantic analysis output are first accurately located by using the quantified entity consensus degree, sentiment polarity divergence degree and summary content similarity. Subsequently, by performing detailed semantic analysis on these divergent text segments, key semantic features that cause the understanding differences between the models are extracted, such as the polysemy of a certain word, different interpretations in a specific context, or differences in regional expressions. Based on these extracted semantic features, the system can induce the complex divergence into preset and easily understood types, thereby concretizing the abstract divergence. Finally, a clear explanatory description is generated for each divergence type and integrated into the semantic graph as additional information, so that the semantic graph not only presents the consensus and divergence areas of the semantics, but also provides a “diagnostic report” for the divergence. This series of steps enables the system to upgrade from “knowing there is divergence” to “knowing what the divergence is and why there is divergence”, providing a more abundant and detailed basis for subsequent arbitration decisions.

[0059] As an embodiment of the present application, the semantic features include polysemous words, context modifiers or regional expressions; and the divergence types include polysemous ambiguity, context-dependent divergence or regional cultural difference divergence.

[0060] Specifically, a polysemous word refers to a word that has multiple meanings in different contexts, such as the word “fashion” which may refer to different styles or trends in different cultural backgrounds. A context modifier refers to a word whose meaning or influence changes with a specific context, such as “leisure” which has different specific requirements in home wear and business casual wear. Regional expression refers to a word, phrase or expression specific to a certain region or cultural circle, such as the understanding of “formal wear” in some regions which may differ from the international standard.

[0061] Among them, polysemous ambiguity refers to divergence caused by different OLLAMA model instances producing different understandings of the same word due to the existence of polysemous words. Context-dependent divergence refers to the difference between different OLLAMA model instances in processing semantics in a specific context due to the existence of context modifiers. Regional cultural difference divergence refers to the inconsistency between different OLLAMA model instances in understanding and interpreting clothing concepts in a specific cultural background due to the existence of regional expressions.

[0062] The solution of the present application classifies the semantic divergence by explicitly identifying specific semantic features that cause the divergence, such as polysemous words, contextual modifiers or regional expressions, and accordingly classifies the divergence as polysemous ambiguity, context-dependent divergence or regional cultural difference divergence, so that the system can make a detailed analysis of the divergence in the OLLAMA model instance output result. It is precisely because of the accurate identification of the cause of the divergence that the subsequent divergence explanation description can more accurately reflect the essence of the problem and provide more insightful information to the user.

[0063] As an embodiment of the present application, the step of adjusting the strength of the associated edge in the semantic graph according to the context information contained in the user query and combining the confidence of the association between the specific entity and the abstract concept in the current context includes: Monitoring the co-occurrence of entities and abstract concepts in the context in the fashion data stream; Specifically, "monitoring the co-occurrence of entities and abstract concepts in the context in the fashion data stream" means that the system continuously collects information related to fashion from various real-time or near real-time data sources, such as social media posts, fashion blog articles, online retail platform product descriptions and user reviews, fashion magazine content, etc. These data streams contain a large amount of text, image and video information, from which specific "entities" (such as "dresses", "jeans", "sneakers", etc.) and "abstract concepts" (such as "retro style", "minimalism", "street fashion", etc.) can be extracted through natural language processing and image recognition techniques. "Context" refers to the specific background in which these entities and concepts appear, such as "summer outfits", "business casual", "party occasions", etc. Co-occurrence means that two or more entities or concepts appear or are closely related in the same context.

[0064] Within the time window, the co-occurrence frequency and recency of entities and abstract concepts in the context are quantified; that is, within a set time period (such as the past 24 hours, a week or a month), the number of times a specific entity and abstract concept appear together in a specific context, i.e. "co-occurrence frequency", is counted. At the same time, "recency" measures the time distance of these co-occurrence events from the current time, and recent co-occurrence events are usually given higher weights to reflect the latest trends. For example, an exponential decay function can be used to calculate the recency score, so that the more recent events contribute more to the total score.

[0065] comparing the co-occurrence frequency and recency with a baseline, identifying changes in the association trend between entities and abstract concepts in a context; refers to comparing the co-occurrence frequency and recency indicators calculated within the current time window with pre-set historical average values, long-term trends or standard values in a specific context (i.e. "baseline"). Through this comparison, it can be judged whether the association strength between the current entity and abstract concept is significantly enhanced, weakened or remains stable, thereby identifying changes in the association trend. For example, if the co-occurrence frequency and recency scores of "oversized suits" and "business casual" in the past week are much higher than their historical baseline, it indicates that the association trend between the two is strengthening.

[0066] According to the context information contained in the user query, combining the confidence of the association between a specific entity and abstract concept in the current context, and adjusting the strength of the association edge in the semantic graph according to the identified trend changes; specifically, when receiving a user query, the system will first analyze the implicit or explicit context information in the query. Then, combined with the existing confidence of the association between a specific entity and abstract concept in the current context (for example, based on historical data or expert knowledge), and taking into account the association trend changes identified through the above steps. If the trend shows that a certain association is strengthening, the strength of the association edge connecting the two entities or concepts in the semantic graph will be increased accordingly; conversely, if the trend is weakening, the strength of the edge will be reduced. This dynamic adjustment ensures that the semantic graph can reflect the latest trends in the fashion field in real time.

[0067] The scheme of the present application can effectively solve the problem of the lag of the existing technology in adjusting the strength of the association edge of the semantic graph, because it introduces a real-time monitoring and trend analysis mechanism for fashion data streams. By continuously "monitoring the co-occurrence of entities and abstract concepts in a context in fashion data streams", the system can capture the budding and evolution of relationships between various elements in the fashion field. Subsequently, "quantifying the co-occurrence frequency and recency of entities and abstract concepts in a context within a time window", the system not only pays attention to the number of co-occurrences, but also values recent co-occurrence events, thereby ensuring sensitivity to the latest trends. It is precisely because of "comparing the co-occurrence frequency and recency with a baseline, identifying changes in the association trend between entities and abstract concepts in a context" that the system can objectively judge which associations are becoming stronger or weaker, rather than relying solely on static experience. Finally, "according to the context information contained in the user query, combining the confidence of the association between a specific entity and abstract concept in the current context, and adjusting the strength of the association edge in the semantic graph according to the identified trend changes", ensures that the association relationships in the semantic graph can dynamically and accurately reflect the real connections in the current fashion context, thereby providing users with more timely and relevant query results.

[0068] As Figure 2An OLLAMA-based data center clothing management system is shown, and the system comprises: The acquisition module 201 is configured to acquire the computing resource state data. The preprocessing module 202 is configured to perform predictive semantic preprocessing, including: When the computing resource state data indicates that the graphic processor is in a preset low load state, according to the query mode information obtained by analyzing the user query log, the unstructured clothing data to be pre-calculated is selected, and the OLLAMA model is driven to perform semantic analysis on the unstructured clothing data to be pre-calculated to obtain a pre-calculated semantic analysis result, and the pre-calculated semantic analysis result is stored in the semantic cache. The query processing module 203 is configured to process a user query request, including: Receiving a user query request and searching in the semantic cache according to the user query request; Generating a query response according to the search result.

[0069] The basic principles, main features and advantages of the present application are shown and described above. It should be understood by those skilled in the art that the present application is not limited by the above embodiments, and the above embodiments and descriptions in the specification are only the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements fall within the scope of the claimed present application.

Claims

1. A data center clothing management method based on OLLAMA, characterized in that, The method includes the following steps: S101. Collect computing resource status data; S102. Perform predictive semantic preprocessing, including: When the computing resource status data indicates that the graphics processor is in a preset low load state, the unstructured clothing data to be pre-computed is selected based on the query pattern information obtained from the user query log analysis, and the OLLAMA model is driven to perform semantic analysis on the unstructured clothing data to be pre-computed to obtain the pre-computed semantic analysis result, and the pre-computed semantic analysis result is stored in the semantic cache. S103. Processing user query requests, including: Receive user query requests and perform a search in the semantic cache according to the user query requests; Generate a query response based on the search results; The step of generating a query response based on the search results includes: If the search result shows that there is a pre-computed semantic analysis result in the semantic cache that satisfies the user's query request, then the pre-computed semantic analysis result is extracted as the query response; If the search result is that there is no pre-computed semantic analysis result in the semantic cache that satisfies the user's query request, or the confidence level of the result inferred in real time by the OLLAMA model is lower than a preset threshold, then the user's query request is submitted to multiple OLLAMA model instances in parallel, and the output results of the multiple OLLAMA model instances are received. When the semantic cache cannot provide a valid response or the inference result of a single model is unreliable, the system no longer passively waits or directly outputs uncertain results. Instead, it actively utilizes the parallel processing capabilities of multiple OLLAMA model instances to simultaneously acquire independent insights from multiple models and compare and arbitrate these insights. The output results of the multiple OLLAMA model instances are compared, and the semantic analysis result with high consistency and high confidence is selected as the final query response according to the preset arbitration rules. The step of comparing the output results of the multiple OLLAMA model instances and selecting the semantic analysis result with high consistency and high confidence as the final query response according to the preset arbitration rules includes: Capture the semantic analysis output of the multiple OLLAMA model instances and convert it into a unified format; Consensus and divergence quantifications are performed on the semantic analysis output of the unified format to obtain entity consensus degree, sentiment polarity divergence degree, and summary content similarity. Based on the consensus and disagreement quantification results, a semantic graph is constructed, which presents semantic consensus regions and semantic disagreement regions, and associates underlying entities with abstract concepts; The consensus and disagreement quantification indicators are appended as metadata to semantic graph elements; Based on the preset arbitration rules and semantic graph elements, extract highly consistent and confident semantic analysis results from the semantic graph as the final query response; After the steps of quantifying consensus and divergence in the semantic analysis output of the unified format to obtain entity consensus degree, sentiment polarity divergence degree, and summary content similarity, the following steps are included: Based on the entity consensus, sentiment polarity divergence, and summary content similarity, identify the text fragments with divergence in the semantic analysis output; Extract the semantic features that lead to the disagreement from the disagreeing text fragments, including polysemous words, contextual modifiers, or regional expressions; Based on the semantic features, the disagreements are classified into preset disagreement types, including polysemous ambiguity, context-dependent disagreement, or regional cultural difference disagreement. An explanatory description of the divergence type is generated and appended as metadata to the corresponding divergence region in the semantic graph.

2. The data center clothing management method based on OLLAMA according to claim 1, characterized in that, The computing resource status data includes: workload and temperature data of the graphics processor on which the OLLAMA model depends for operation, as well as the response time of the OLLAMA model in completing inference tasks.

3. The data center clothing management method based on OLLAMA according to claim 1, characterized in that, The step of constructing a semantic graph based on the consensus and disagreement quantification results, wherein the semantic graph presents semantic consensus regions and semantic disagreement regions, and associates underlying entities with abstract concepts, includes: Based on the consensus and disagreement quantification results, newly emerging abstract concepts or entities in the semantic analysis output are identified, as well as the semantic consensus region and semantic disagreement region. Based on the newly identified abstract concepts or entities, as well as the semantic consensus region and semantic divergence region, the nodes and associated edges of the semantic graph are expanded, and the underlying entities are associated with the abstract concepts.

4. The data center clothing management method based on OLLAMA according to claim 3, characterized in that, The step of constructing a semantic graph based on the consensus and disagreement quantification results, wherein the semantic graph presents semantic consensus regions and semantic disagreement regions, and associates underlying entities with abstract concepts, further includes: Based on the contextual information contained in the user query, and combined with the confidence level of the association between a specific entity and an abstract concept in the current context, the strength of the association edges in the semantic graph is adjusted. Based on the user's query intent, by identifying the core abstract concept of the query and performing multi-step semantic association, the system explores and makes explicit the deep-level association paths connected by intermediate entities or secondary abstract concepts, thus constructing multi-level abstract concept association paths in the semantic graph.

5. A data center clothing management method based on OLLAMA according to claim 4, characterized in that, The step of adjusting the strength of the association edges in the semantic graph based on the contextual information contained in the user query and the confidence level of the association between a specific entity and an abstract concept in the current context includes: Monitoring the co-occurrence of entities and abstract concepts in context within fashion data streams; Within a time window, quantify the co-occurrence frequency and proximity of the entity and the abstract concept in the context; By comparing the co-occurrence frequency and proximity with a benchmark, the changing trend of the association between the entity and the abstract concept in the context can be identified. Based on the contextual information contained in the user query, combined with the confidence level of the association between specific entities and abstract concepts in the current context, and according to the identified trend changes in association, the strength of the association edges in the semantic graph is adjusted.

Images