A large model-oriented data storage and dynamic updating system and method

Abstract

The application provides a large model-oriented data storage and dynamic updating system and method, and belongs to the technical field of data updating and storage. The system comprises a data aggregation module, a storage module, a large model interaction module and a dynamic updating module. The original interaction data of a user is classified according to attributes, capabilities, environment and intention by a semantic recognition model, a single natural language user portrait document is generated by aggregation, and is stored in the form of a key-value pair. When a large model interacts, the system directly reads the document as complete context input, greatly improving the data acquisition efficiency. In addition, the system can update and maintain the document in real time according to interaction feedback. The application realizes the change from a "relational" to a "text" storage paradigm, significantly reduces the complexity of data query and processing and the Token consumption of large model interaction, guarantees the timeliness and accuracy of the user portrait, and thus improves the quality and efficiency of the personalized service of the large model.

Description

Technical Field

[0001] This invention relates to the field of data update and storage technology, and more specifically to a data storage and dynamic update system and method for large models. Background Technology

[0002] With the widespread application of large language models in fields such as intelligent assistants and personalized recommendations, these models are increasingly reliant on contextual information such as user history, preferences, and states. Currently, to support personalized interactions with large models, application systems typically store various user information in traditional relational databases. When a large model needs to serve a user, the system must query and extract relevant information from multiple related data tables and concatenate them into a natural language description, which serves as context input to the large model. This multi-table join query and information concatenation based on relational databases is currently the main technical path supporting large models in understanding users.

[0003] However, the aforementioned existing technical solutions have significant drawbacks, making it difficult to meet the needs of large-scale models for efficient, real-time, and low-cost interaction. First, query efficiency is low: complex multi-table join operations are time-consuming and labor-intensive, making it difficult to guarantee the real-time response of large-scale models. Second, data updates are lagging: dynamic changes in user profiles (such as intent and environment) are difficult to update and synchronize in real-time and at low cost in relational databases, causing large-scale models to make decisions based on outdated information. Third, interaction costs are high: each interaction requires multi-table queries and manual information concatenation, which not only increases system complexity but also significantly increases the number of tokens input to the large-scale model due to the generation of lengthy context descriptions, driving up API call costs. Finally, there is a data paradigm mismatch: the structured storage designed for join queries is at odds with the input paradigm of large-scale models directly processing natural language text, resulting in cumbersome data processing flows. Summary of the Invention

[0004] This invention provides a data storage and dynamic update system and method for large models, which solves the problems of low efficiency of multi-table join queries, delayed updates of user profile data, and high cost of context token construction in the prior art.

[0005] To achieve the above objectives, one embodiment of the present invention provides a data storage and dynamic update system for large models, comprising: a data aggregation module, used to acquire users' original interaction data, and call a semantic recognition model to perform entity and relationship recognition on the original interaction data, classify and structure the recognized information according to a preset user information classification model, and generate a user profile document; a storage module, used to store the user profile document in key-value pair format; a large model interaction module, used to, upon receiving a user's service request, read the corresponding user profile document from the storage module according to the user identifier, and input the user profile document as context into the service large model; and a dynamic update module, used to update and maintain the corresponding user profile document in the storage module according to the new data generated by the interaction between the service large model and the user.

[0006] Optionally, the dynamic update module includes: an information confidence management unit, used to assign confidence weights that decay over time to information items in the user profile document to maintain its timeliness; a semantic conflict detection unit, used to call a semantic recognition model to determine whether there is a logical conflict between the new information and existing information in the user profile document before writing new information, and to perform state iteration operations or semantic fusion operations based on the judgment result to maintain its consistency; and a document compression unit, used to periodically call the semantic recognition model to perform summarization and redundancy removal processing on the user profile document, generate a compressed user profile document, and update it to the storage module to maintain its conciseness.

[0007] On the other hand, a data storage and dynamic update method for large models is also provided, applied to the aforementioned data storage and dynamic update system, including: Step S1: Obtaining the user's original interaction data, calling a semantic recognition model to perform entity and relationship recognition on the original interaction data, classifying and structuring the recognized information according to a preset user information classification model, generating a user profile document, and storing the user profile document in key-value pair format; Step S2: Receiving the user's service request, reading the corresponding user profile document according to the user identifier, and inputting the user profile document as context into the service large model; Step S3: Updating and maintaining the corresponding user profile document according to the new data generated by the interaction between the service large model and the user.

[0008] Optionally, the preset user information classification model is specifically a classification model based on four dimensions: attributes, capabilities, environment, and intent.

[0009] Optionally, updating and maintaining the corresponding user profile document includes: calculating the confidence weight that decays over time for each information item in the user profile document, and retaining valid information and removing invalid information based on the confidence weight.

[0010] Optionally, step S3 further includes: before writing new information, calling a semantic recognition model to determine whether there is a logical conflict between the new information and the existing information in the user profile document; if there is a conflict, performing a state change operation to transfer the existing information to the history record and use the new information as the current information; if there is no conflict, performing a semantic fusion operation to merge the new information with the existing information.

[0011] Optionally, the data storage and dynamic update method further includes: periodically calling the semantic recognition model to perform summary reconstruction and redundancy removal on the user profile document, generating a refined user profile document and updating the storage.

[0012] Optionally, the semantic recognition model is a large model used to perform named entity recognition and relation extraction.

[0013] This invention provides a data storage and dynamic update system and method for large-scale models. It aggregates scattered, multi-dimensional user information into a single, natural language-based "user profile document" for storage, and introduces a dynamic update mechanism based on confidence management, semantic conflict detection, and timed compression, achieving a shift from relational to text-based storage paradigms. This significantly improves the efficiency of large-scale models in obtaining context, simplifying complex multi-table join queries to a single key-value read; it significantly reduces the token consumption and cost of large-scale model interactions by providing concise and complete user documents directly as prompt input, avoiding redundant information splicing; and it ensures the timeliness, consistency, and accuracy of user profiles, enabling large-scale models to interact based on a continuously evolving, real-time updated "live" profile, thereby providing more accurate and consistent personalized services. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0015] Figure 1 This is a general flowchart of the data storage and dynamic update method provided in the embodiments of the present invention; Figure 2 This is a flowchart of the core process for dynamic updating provided in an embodiment of the present invention; Figure 3 This is a flowchart of the semantic conflict detection sub-process provided in an embodiment of the present invention. Detailed Implementation

[0016] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0017] It should be noted that the acquisition, transmission, storage, use, and processing of data in the technical solution of this application all comply with the relevant provisions of national laws and regulations. In the embodiments of this application, certain existing industry solutions such as software, components, and models may be mentioned. These should be considered exemplary, intended only to illustrate the feasibility of implementing the technical solution of this application, and do not imply that the applicant has already used or necessarily used such solutions.

[0018] The following is combined Figures 1-3 This invention is described in detail.

[0019] Example 1: This invention provides a data storage and dynamic update system for large-scale models, comprising: a data aggregation module for acquiring users' raw interaction data, calling a semantic recognition model to identify entities and relationships in the raw interaction data, classifying and structuring the identified information according to a preset user information classification model, and generating a user profile document; a storage module for storing the user profile document in key-value pair format; a large-scale model interaction module for reading the corresponding user profile document from the storage module based on the user identifier when receiving a user's service request, and inputting the user profile document as context into the service large-scale model; and a dynamic update module for updating and maintaining the corresponding user profile document in the storage module based on new data generated by the interaction between the service large-scale model and the user.

[0020] Specifically, the dynamic update module includes: an information confidence management unit, used to assign confidence weights that decay over time to information items in the user profile document to maintain its timeliness; a semantic conflict detection unit, used to call a semantic recognition model to determine whether there is a logical conflict between the new information and existing information in the user profile document before writing new information, and to perform state iteration operations or semantic fusion operations based on the judgment result to maintain its consistency; and a document compression unit, used to periodically call the semantic recognition model to perform summarization and redundancy removal processing on the user profile document, generate a compressed user profile document and update it to the storage module to maintain its conciseness.

[0021] The original interaction data refers to the data generated when users communicate with applications or services that have natural language interaction capabilities. Its main forms include natural language dialogue text actively entered by the user, as well as user behavior logs and event tracking data recorded by the system in the background. The semantic recognition model, in this system, refers to a large-scale language model with specific prompts and task definitions. The core function of this model is set to perform named entity recognition and relation extraction. Named entity recognition refers to automatically detecting and classifying predefined categories of real-world objects from text, such as names of people, places, and skill names. Relation extraction refers to identifying and structuring the semantic relationships or action associations between these identified entities. The user information classification model is a predefined conceptual framework for classifying user information across all dimensions. This framework divides user information into four mutually exclusive and jointly exhaustive dimensions: the attribute dimension covers the user's objective, static characteristics and long-term preferences; the ability dimension describes the skills and knowledge the user possesses; the environmental dimension refers to the user's physical location, social relationships, and the IoT device environment they belong to; and the intent dimension captures the user's recent goals, plans, or desires. A user profile document is a unique data carrier about a specific user, generated after the aforementioned process. Its content is essentially a structured natural language text. Based on the user information classification model framework, this text organizes and describes the extracted user information in readable sentences, forming a machine-readable and human-understandable user summary. Key-value pair storage is a storage method distinct from traditional relational database table structures. In this format, the system uses each user's unique identifier as the retrieval key and the corresponding complete user profile document text as the value, forming a one-to-one mapping that is persistently saved. This storage paradigm allows the system to directly obtain all aggregated information about a user through a simple key-value query, thus completely avoiding the operations of performing join queries and data concatenation between multiple data tables. A service request specifically refers to an instruction or question initiated by a user to an application that integrates large-scale language model functionality, requiring the model to provide a personalized response based on its understanding of the user. A user identifier is a string used by the system to uniquely locate and retrieve a specific user profile document in the storage module, such as a user account ID or session ID. Using user profile documents as contextual input means that when the system submits a user's current service request to a large language model for processing, it first appends the complete user profile document text, read from storage without any formatting conversion, to the user's question as input prompts for the model. This allows the model to directly understand the user's complete background in a single call, without needing to trace back through the dialogue history or actively ask for information. New data generated from the interaction between the service model and the user refers to information fragments parsed from each completed dialogue round that may indicate changes in the user's state, preferences, or goals. This new data is the source that triggers the evolution of document content.

[0022] Specifically, the information confidence management unit records the precise timestamp of each structured information item in the user profile document when it was extracted and stored, and calculates a confidence weight that dynamically decreases over time. This weight is quantified using a decay model based on a natural exponential function. The calculation involves multiplying a fixed initial weight of the information item by a negative power of the natural constant e. The value of this power is the product of a specific decay coefficient preset for the category to which the information item belongs and its storage duration in days. For example, volatile intent-type information is assigned a larger decay coefficient, causing its weight to decrease rapidly, while stable attribute-type information is assigned a smaller decay coefficient, causing its weight to decrease slowly. When information validity judgment or conflict resolution is required, the system directly compares the real-time confidence weights of each piece of information and adopts the information with the higher weight as the more reliable and timely currently valid information.

[0023] Specifically, before writing any newly extracted user information item into storage, the semantic conflict detection unit again invokes the semantic recognition model to act as a logical reviewer. Its core task is to determine whether there are factual logical contradictions between the new information item and existing related information items in the user profile document. The specific method is that the unit submits both the natural language description of the existing information and the description of the new information to the model, instructing it to analyze whether their states in the objective world are mutually exclusive. Based on the model's output, the unit performs two different operations: if a conflict is determined, a state iteration operation is performed, marking the existing information as a historical version and archiving it, while establishing the new information as the only currently valid information. This is typically used to handle updates such as changes in residence or shifts in points of interest; if no conflict is determined, a semantic fusion operation is performed, merging the new information with the existing information into a more comprehensive description. For example, adding "hobbies like jogging" to a user's existing "hobbies like fishing" results in "hobbies like fishing and jogging," thereby continuously enriching the dimensions of the user profile. This mechanism is a key technical guarantee for maintaining the logical consistency within the user profile document.

[0024] Specifically, the document compression unit is not triggered after each interaction, but rather automatically starts running according to a preset fixed cycle, such as once a day. Its specific workflow is as follows: when the scheduling cycle arrives, the system retrieves the complete profile document of the specified user from storage, including all its current structured information entries; then, it calls the semantic recognition model and instructs it to perform global analysis, summarization, and rewriting of this information based on a preset user information classification model framework. The model's summarization and deredundancy removal mainly include two aspects: first, semantic summarization, which merges frequently occurring and similarly expressed fragmented descriptions into a concise summary statement, such as summarizing "likes fishing," "frequently goes fishing," and "went fishing last weekend" into "experienced fishing enthusiast"; second, logical deredundancy removal, which removes historical state descriptions that have been covered by updated information, or deletes secondary and outdated information whose confidence has decayed below the threshold. Finally, the model outputs a new, coherent, and concise natural language summary, which the system uses to completely replace the old version of the document content in storage, thus ensuring that the context serving each interaction with the large model is highly condensed and has optimal information density.

[0025] like Figure 1 As shown, this embodiment of the invention also provides a data storage and dynamic update method for large models, applied to the aforementioned data storage and dynamic update system, including: Step S1: Obtaining the user's original interaction data, calling a semantic recognition model to perform entity and relationship recognition on the original interaction data, classifying and structuring the recognized information according to a preset user information classification model, generating a user profile document, and storing the user profile document in key-value pair format; Step S2: Receiving a user's service request, reading the corresponding user profile document according to the user identifier, and inputting the user profile document as context into the service large model; Step S3: Updating and maintaining the corresponding user profile document according to the new data generated by the interaction between the service large model and the user.

[0026] Preferably, the preset user information classification model is specifically a classification model based on four dimensions: attributes, capabilities, environment, and intent. The semantic recognition model is a large model used to perform named entity recognition and relation extraction.

[0027] This method continuously monitors and captures users' raw interaction data, processes it using a semantic recognition model, and ultimately generates and stores structured user profile documents. Its core lies in building an automated pipeline from data ingestion and processing to persistence. The service-wide model interaction assistance step, upon receiving a request, quickly retrieves the stored profile document based on the user identifier and uses it as the integrated context input to the service-wide model. Subsequent update and maintenance steps asynchronously analyze new data after each interaction, ensuring continuous optimization of the profile document through a series of sub-processes.

[0028] Specifically, the complete implementation process of this method is as follows: Data Ingestion and Document Construction: The system acquires raw dialogue text or behavioral tracking data bound to user identifiers in real time by listening to message queues or polling APIs. For each piece of data, the system constructs a system prompt containing explicit instructions, calls the semantic recognition model to perform named entity recognition and relation extraction, and forces it to classify and structure the results according to four dimensions: attributes, capabilities, environment, and intent. After receiving this structured fragment, the system adds metadata such as timestamps to it, forming a standardized user information record. Then, using the user identifier as the key, this record is appended to the corresponding user's structured information list in the database, completing the initial profile construction and storage. Context Provision During Interaction: When the application frontend sends a service request that requires a personalized response, the system parses out the user identifier and the current question. Immediately afterwards, using the user identifier as the key, a precise query is initiated to the database, directly reading the compressed, final version of the natural language profile summary text for the corresponding user. Subsequently, the system assembles this summary text and the user's current question according to a preset, optimized prompt template to form a complete prompt, which is then directly submitted to the service big model. This enables real-time and efficient injection of context. Post-interaction profile maintenance and evolution: After the service's main model completes the interaction, the system asynchronously initiates an update process. First, it parses new information clues that may reflect changes in user state from the logs of this interaction. Next, the confidence management sub-process recalculates the real-time weights based on time decay for all relevant new and old information items. Then, the semantic conflict detection and resolution sub-process is triggered, which again calls the semantic recognition model to determine the relationship between new clues and existing information, and performs state iteration (replacing old with new) or semantic fusion (information merging) based on the judgment result (conflict or no conflict), referring to the confidence weights during the decision-making process. The processing result updates the underlying structured information list. Furthermore, a timed compression sub-process, independent of individual interactions, runs periodically. It performs global semantic summarization and logical redundancy removal on the information list, generating a completely new, refined summary to overwrite the old display document, thereby ensuring the persistent refinement of the profile.

[0029] This invention achieves significant technical results by aggregating multi-source, dispersed user information into a single natural language document for storage, and combining confidence management, conflict resolution, and periodic compression to achieve dynamic updates: it greatly improves the efficiency of large models in acquiring user context, optimizing complex multi-table join queries into a single efficient key-value read; it significantly reduces the token consumption and cost of interaction by providing concise and complete documents directly as prompt input, avoiding redundant splicing; and it ensures the continuous accuracy and real-time nature of user profiles, enabling large models to interact based on a continuously evolving and internally consistent "live" profile, thereby providing accurate and consistent personalized services.

[0030] like Figure 2 and Figure 3 As shown, preferably, updating and maintaining the corresponding user profile document includes: calculating the confidence weight that decays over time for each information item in the user profile document, and retaining valid information and removing invalid information based on the confidence weight.

[0031] More preferably, step S3 further includes: before writing new information, calling a semantic recognition model to determine whether there is a logical conflict between the new information and the existing information in the user profile document; if there is a conflict, performing a state change operation to transfer the existing information to the history record and use the new information as the current information; if there is no conflict, performing a semantic fusion operation to merge the new information with the existing information.

[0032] Specifically, when new information is to be processed, the system first calculates the real-time confidence weight for all relevant new and old information items. This calculation is based on a standard exponential decay model, whose inputs are the storage duration of the information item and a pre-set decay coefficient according to its category. This outputs a quantitative score for each piece of information that decreases over time, thereby automatically distinguishing between timely and valid information and low-scoring invalid information. Subsequently, the system enters the conflict detection stage, submitting the natural language descriptions of the new and existing information to the semantic recognition model, instructing it to determine whether the two contradict each other at the level of objective facts. The model's analysis results and the confidence weights calculated in the previous step jointly drive the final decision: if a conflict is determined, a state change is performed, that is, the information item with a higher confidence weight (usually the newer one) is used to cover the information item with a lower weight, and the replaced item is transferred to the historical archive; if no conflict is determined, semantic fusion is performed, merging the new and old information semantically into a more comprehensive description. Through this series of sequential evaluations, judgments, and executions, the system automatically completes the information purification and logical consistency maintenance of user profile documents without human intervention.

[0033] The preferred embodiments of this invention constitute an automated mechanism to ensure the information quality and logical consistency of user profile documents during dynamic updates. Specifically, by introducing confidence weight calculation based on time decay, the system can automatically quantify and filter out more timely and effective information, eliminating outdated information, thereby ensuring the timeliness and accuracy of the profile. Furthermore, by calling a semantic recognition model to perform conflict judgment and accordingly execute precise state changes or semantic fusion operations, the system can intelligently resolve contradictions between information and enrich the dimensions of the profile, thereby ensuring the inherent consistency and completeness of the profile. The synergistic effect of these two mechanisms enables user profile documents to automatically maintain high information quality during continuous and high-frequency updates, providing a reliable and contradictory contextual basis for large models, directly improving the accuracy and credibility of personalized interactions.

[0034] Preferably, the data storage and dynamic update method further includes: periodically calling the semantic recognition model to perform summary reconstruction and redundancy removal on the user profile document, generating a refined user profile document and updating the storage.

[0035] Periodic invocation means that this step is triggered by an independent timed task scheduler within the system, rather than by a single real-time user interaction. Summary reconstruction and redundancy removal refer to a global analysis, semantic summarization, and logical simplification of all valid information entries accumulated at the bottom layer of the user profile document. Generating and updating the refined user profile document means ultimately producing and saving a new version of the document with significantly optimized text length and more concise expression, without information loss, to replace the original version for subsequent interactions.

[0036] Specifically, this step is executed automatically as a separate background batch processing job at fixed intervals. Its complete workflow is as follows: At the set scheduled time, the system iterates through all users or targets a specific user, retrieving a complete user profile document from its storage, particularly the list of all structured information items and the current profile summary text. Subsequently, the system invokes the semantic recognition model and submits a carefully designed instruction, requiring the model to comprehensively analyze all provided information items based on a framework of attributes, capabilities, environment, and intent. This instruction requires the model to perform two core operations: semantic induction, which identifies and merges multiple fragmented descriptions expressing the same or similar semantics into a concise summary statement; and logical deduplication, which identifies and removes historical state descriptions covered by updated information, or filters out less important secondary information according to preset rules. After the model performs in-depth processing and rewriting of the input information according to the instruction, it outputs a completely new, coherent, and non-repeating natural language summary. Finally, the system writes the newly generated summary text to the display field of the corresponding user's stored document, thus completing the compression and refinement update of this document and ensuring that the profile document used for a long time always maintains a high information density.

[0037] Example 2: The present invention will be further described in detail below with reference to embodiments. This embodiment takes the interaction process between a user named "Xiaobai" and a smart home voice assistant as an example to fully demonstrate the specific implementation of the system and method.

[0038] User Xiaobai says to the intelligent assistant, "I am Xiaobai, I live in A, and I like fishing." The system's backend data aggregation module immediately captures this original dialogue text. This module immediately calls the semantic recognition model, submitting a structured instruction to it, requesting the model to identify entities and relationships from the dialogue and output results according to four categories: attributes, abilities, environment, and intent. After analysis, the model identifies "Xiaobai" as a name (attribute), "likes fishing" as a hobby (attribute), and "lives in A" as a residence (environment). The system packages these structured information fragments, along with the current timestamp and session number, into a standard user information record. Next, the system uses "user_001" as Xiaobai's unique user identifier, creates a corresponding storage location in the storage module's database, stores this initial record, and generates the first version of the user profile document, which can be summarized as "User Xiaobai, likes fishing, lives in A."

[0039] When Xiaobai interacts with the assistant again, the large-scale model interaction module begins to work. For example, if Xiaobai asks, "Any suggestions for the weekend?" the system first parses the user identifier "user_001" and uses it to directly query the database, retrieving the complete user profile document stored in one go. Then, the system assembles the document content with the current question in the format "Known user information: [Profile document content] Please answer: [User's current question]" to form a complete prompt, which is then submitted directly to the service large-scale model. Based on the context "Xiaobai enjoys fishing and lives in A," the service large-scale model can generate a personalized recommendation such as "You can go fishing at the lake," without needing to ask the user's interests and location beforehand.

[0040] After each interaction, the dynamic update module is activated to maintain the freshness of the user profile. For example, in a subsequent conversation, Xiaobai reveals, "But I've also started to get interested in programming." The system extracts the new information "interested in programming" from this. First, the information confidence management unit calculates the real-time confidence weight for all relevant information (including newly extracted and stored "hobby fishing"), with new information having the highest weight because it has just been generated. Next, the semantic conflict detection unit is invoked, submitting "interested in programming" and "hobby fishing" together to the semantic recognition model to inquire whether they contradict each other. The model determines that they do not contradict each other, so the system performs a semantic fusion operation, merging the old and new information and updating the user's intent or interest to "hobby fishing, and also interested in programming." Subsequently, Xiaobai says, "Actually, I'm a programmer." The system extracts the ability information "profession is programmer." The conflict detection unit determines that this may be inconsistent with the previous intent of "interested in programming." Combining the model's judgment with the principle of higher confidence in new information, the system performs a state change operation, replacing the previous intent description of "interested in programming" with the more definitive current ability information of "profession is programmer," and archiving the latter. In addition, the document compression unit, as an independent background task, performs global optimization of the user profile, for example, every day at midnight. It calls the semantic recognition model to summarize and deduplicate all of Xiaobai's currently valid information (such as "enjoys fishing," "professional programmer," "resides in Hangzhou"), generating a concise summary such as "User Xiaobai is a programmer residing in A who enjoys fishing," and updates the version of the main document in storage.

[0041] In summary, the data storage and dynamic update system and method for large-scale models provided by this invention achieves a data paradigm shift from relational to textual data by aggregating scattered, multi-dimensional user information into a single, natural language text-based user profile document. This system and method significantly improve the efficiency of large-scale models in acquiring context, simplifying complex multi-table join queries into a single, efficient key-value read; it significantly reduces token consumption and API call costs for interactions by providing concise, complete documents directly as prompt input, avoiding redundant information splicing; and it introduces a dynamic update mechanism based on confidence management, semantic conflict detection, and periodic compression to ensure the timeliness, consistency, and accuracy of user profiles, allowing them to continuously evolve with interactions. Ultimately, this enables large-scale models to interact based on a high-quality, real-time updated "live" profile, thereby providing more accurate, coherent, and low-cost personalized services.

[0042] The above description is merely a preferred embodiment of the technical solution of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A data storage and dynamic update system for large models, characterized in that, include: The data aggregation module is used to acquire the user's original interaction data, and call the semantic recognition model to identify entities and relationships in the original interaction data. The identified information is then classified and structured according to a preset user information classification model to generate a user profile document. The storage module is used to store the user profile document in key-value pair format; The large model interaction module is used to read the corresponding user profile document from the storage module according to the user identifier when a user's service request is received, and input the user profile document as context into the service large model; The dynamic update module is used to update and maintain the corresponding user profile documents in the storage module based on new data generated from the interaction between the service big model and users.

2. The data storage and dynamic update system according to claim 1, characterized in that, The dynamic update module includes: The information confidence management unit is used to assign confidence weights that decay over time to information items in the user profile document in order to maintain their timeliness. The semantic conflict detection unit is used to call the semantic recognition model to determine whether there is a logical conflict between the new information and the existing information in the user profile document before writing new information, and to perform state change operation or semantic fusion operation according to the judgment result in order to maintain its consistency. The document compression unit is used to periodically call the semantic recognition model to perform summarization and redundancy removal on the user profile document, generate a compressed user profile document, and update it to the storage module to maintain its conciseness.

3. A data storage and dynamic update method for large models, applied to the data storage and dynamic update system described in claims 1-2, characterized in that, include: Step S1: Obtain the user's original interaction data, call the semantic recognition model to perform entity and relationship recognition on the original interaction data, classify and structure the recognized information according to the preset user information classification model, generate a user profile document, and store the user profile document in key-value pair format. Step S2: Receive the user's service request, read the corresponding user profile document according to the user identifier, and input the user profile document as context into the service big model; Step S3: Update and maintain the corresponding user profile document based on the new data generated from the interaction between the service big model and the user.

4. The data storage and dynamic update method for large models according to claim 3, characterized in that, The preset user information classification model is specifically a classification model based on four dimensions: attributes, capabilities, environment, and intent.

5. The data storage and dynamic update method for large models according to claim 3, characterized in that, The updating and maintenance of the corresponding user profile document includes: Calculate the confidence weight that decays over time for each information item in the user profile document, and retain valid information and remove invalid information based on the confidence weight.

6. The data storage and dynamic update method according to claim 5, characterized in that, Step S3 further includes: Before writing new information, the semantic recognition model is called to determine whether there is a logical conflict between the new information and the existing information in the user profile document. If a conflict exists, a state change operation is performed, transferring the existing information to the history record and using the new information as the current information; if no conflict exists, a semantic fusion operation is performed, merging the new information with the existing information.

7. The data storage and dynamic update method according to claim 3, characterized in that, It also includes the following steps: The semantic recognition model is periodically invoked to reconstruct summaries and remove redundancies from user profile documents, generating refined user profile documents and updating their storage.

8. The data storage and dynamic update method according to claim 3, characterized in that, The semantic recognition model is a large model used to perform named entity recognition and relation extraction.

Images