Dynamic AI Knowledge Graph System

Abstract

The application provides a dynamic AI knowledge graph system, and relates to the technical field of artificial intelligence and knowledge graph. The system aims to solve the problems of static information, missing relationship and context break in the existing AI memory scheme, and provides an efficient, dynamic and long-term evolving "memory system" for AI agents through automatic knowledge construction, dynamic time perception, real-time response change and mixed retrieval reasoning ability. Through the core processes of data encapsulation, knowledge pipeline processing, time sequence addition and fusion, multi-modal retrieval and context empowerment, the application realizes the real-time conversion of unstructured and structured information into a dynamic knowledge network with time perception ability, significantly improving the performance of AI agents in multi-round interaction, state reasoning and task planning.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a dynamic AI knowledge graph system. Background Technology

[0002] In the development of AI agents, building an efficient, dynamic, and long-term evolving "memory system" is one of the core challenges. Currently, mainstream AI memory solutions (such as RAG (Retrieval Augmentation) based on vector databases) perform well when processing static documents, but have significant limitations in dynamic interactive environments.

[0003] On the one hand, traditional RAGs tend to process static information and struggle to handle dynamic information that changes over time or comes in real time. For example, a customer service agent cannot "remember" the process of an order status changing from "pending" to "shipped," and can only query the latest status document.

[0004] On the other hand, traditional RAGs often suffer from missing relationships. In traditional RAGs, vector retrieval focuses on semantic similarity and cannot effectively represent complex, multidimensional relationships between entities. For example, it is difficult to effectively model the multiple relationships between "User A" and "Product B," such as "browsed," "purchased," and "returned."

[0005] On the other hand, traditional RAGs are always plagued by contextual fragmentation. In multi-turn dialogues or long-term interactions in large language models, AI agents often need to reason about historical information, but traditional RAGs cannot effectively absorb newly generated knowledge, and their performance tends to be mechanical and rigid, lacking depth. It is almost impossible for humans to maintain dynamic knowledge relationships.

[0006] Therefore, there is an urgent need for an AI knowledge graph framework that can handle dynamic information, model entity relationships, and maintain contextual continuity. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a dynamic AI knowledge graph system that effectively solves the problems of static information, missing relationships, and fragmented context that are common in traditional RAGs, enabling dynamic memory and efficient reasoning for AI agents.

[0008] Specifically, this invention provides a dynamic AI knowledge graph system, including a data encapsulation module, wherein target data is input into the data encapsulation module and encapsulated into standardized plot objects, the target data including unstructured data and structured data; a knowledge pipeline processing module, wherein the knowledge pipeline processing module calls a large language model and a user-defined data logic structure, the user-defined data logic structure defining extraction rules and data structure templates for the standardized plot objects, and the large language model extracting upper-level entity relationships, internal attribute facts, and special cross-domain relationship associations from the standardized plot objects; a temporal appending and fusion module, wherein the temporal appending and fusion module appends timestamps to the upper-level entity relationships, internal attribute facts, and special cross-domain relationship associations to create an initial knowledge subgraph; a hybrid retrieval interface and a storage layer, wherein the hybrid retrieval interface matches the initial knowledge subgraph with multiple existing knowledge graphs in the storage layer; and a context empowerment module, wherein if a knowledge graph matching the initial knowledge subgraph exists in the storage layer, the matching knowledge graph is input into the context empowerment module, and the context empowerment module makes decisions based on the input knowledge graph.

[0009] Preferably, if the initial knowledge subgraph cannot be matched with the corresponding knowledge graph in the storage layer, the entities in the initial knowledge subgraph are extracted, and a unique node is created for each entity.

[0010] Preferably, the attributes of the unique node are appended with the timestamp generated by the timing appending and fusion module.

[0011] Preferably, the upper-level entity relationships and special cross-domain relationships in the initial knowledge subgraph are extracted, and associated edges are established between the created nodes based on the upper-level entity relationships and special cross-domain relationships.

[0012] Preferably, after the new node and associated edge are created, they are not directly written to the storage layer. Instead, they are first combined with the existing knowledge graph in the storage layer, and the consistency between the existing knowledge graph in the storage layer and the new knowledge is checked according to preset rules. The new knowledge refers to the newly created node and associated edge.

[0013] Preferably, when rule checks detect a conflict between new knowledge and the knowledge graph, a preset conflict resolution strategy is activated.

[0014] Preferably, if the conflict lies in the attribute differences of nodes or associated edges of the same entity at different timestamps, the information corresponding to the latest timestamp shall prevail.

[0015] Therefore, this invention provides a dynamic AI knowledge graph system, relating to the fields of artificial intelligence and knowledge graph technology. This system aims to solve the problems of static information, missing relationships, and fragmented context in existing AI memory solutions. Through automated knowledge construction, dynamic time awareness, real-time response to changes, and hybrid retrieval and reasoning capabilities, it provides AI agents with an efficient, dynamic, and long-term evolving "memory system." This invention achieves real-time transformation of unstructured and structured information into a dynamic knowledge network with time awareness through core processes such as data encapsulation, knowledge pipeline processing, temporal appending and fusion, multimodal retrieval, and context empowerment. This significantly improves the performance of AI agents in multi-turn interactions, state reasoning, and task planning. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be discussed below. Obviously, the technical solutions described in conjunction with the accompanying drawings are only some embodiments of the present invention. For those skilled in the art, other embodiments and their accompanying drawings can be obtained based on the embodiments shown in these drawings without creative effort.

[0017] Figure 1 A modular block diagram of a dynamic AI knowledge graph system according to the present invention is shown;

[0018] Figure 2 A flowchart illustrating the operation of the dynamic AI knowledge graph system according to the present invention is shown. Detailed Implementation

[0019] The technical solutions of various embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments described in the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] In summary, to address the aforementioned problems commonly found in existing technologies, this invention provides a dynamic AI knowledge graph system, including a data encapsulation module, a knowledge pipeline processing module, a temporal appending and fusion module, a multimodal retrieval module, and a context empowerment module. These modules work together to construct and apply a dynamic knowledge graph.

[0021] Figure 1 A general modular block diagram of the dynamic AI knowledge graph system according to the present invention is shown. The following will be described according to... Figure 1 This invention introduces the basic operation of the dynamic AI knowledge graph system provided by this invention.

[0022] like Figure 1 As shown, the target data (or data to be processed) is input into the data encapsulation module. In this module, the target data is uniformly encapsulated into standardized episode objects, thus achieving data standardization. The target data includes both unstructured and structured data, such as text, JSON, and messages. The episode object contains metadata such as the data body, source, and reference time.

[0023] The standardized plot object is input into the knowledge pipeline processing module. The knowledge pipeline processing module calls a large language model (e.g., an LLM large language model) and a user-defined data logic structure (Pydantic Schema). The user-defined data logic structure defines extraction rules and data structure templates for the input standardized plot object to ensure the consistency of the extraction results of the large language model on the standardized plot object. The large language model then performs content parsing, extracting upper-level entity relationships, internal attribute facts, and special cross-domain relationship associations from the input standardized plot object, completing the automatic extraction and structuring of entities, relationships, and facts.

[0024] Extracting upper-level entity relationships refers to identifying and extracting the associations between different data entities from the input scenario object data. Taking the processing of user interactions by an "AI customer service robot" as an example, from the dialogue "Has my order A been shipped?", the robot uses a Large Language Model (LLM) and a custom data logic structure to parse the content, extracting two data entities (i.e., entities) "User x" and "Order A", and determining the association between them as "User x -- {query status] --> Order A". This process transforms unstructured natural language information into structured entity relationship data, laying the foundation for subsequent knowledge graph construction and helping the AI ​​agent understand the interaction patterns of different entities in specific scenarios.

[0025] The extraction of so-called internal attribute facts focuses on individual data entities, mining their inherent attribute information and related facts. For example, if the delivery department returns the information "Order A was shipped at 3 PM today," the system extracts the entity "Order A" and the attributes "Status: Shipped, Time: 3 PM today." These attribute facts enrich the entity's description in the knowledge graph, making the AI ​​agent's understanding of the entity more specific and accurate, enabling more precise decisions and responses based on this detailed information.

[0026] Special cross-domain relationship association refers to the specific relationships that may exist between data entities (entities) from different domains within complex business scenarios or knowledge systems. Special cross-domain relationship association involves discovering and establishing these connections between data entities across different domains. For example, in an e-commerce recommendation system, when user Y browses product B and then product C, the system analyzes and discovers that "product B" and "product C" belong to the same type of electronic product, thus establishing a cross-domain relationship of "product B - same type → product C". This discovery of cross-domain relationships expands the relational dimensions of knowledge graphs, enabling AI agents to reason and analyze from a broader perspective, providing users with more relevant and valuable services.

[0027] The three aspects extracted above lay the foundation for adding precise timestamps to the timing appending and fusion module in this invention.

[0028] The plot object data is then input into the temporal attachment and fusion module, which adds timestamps to the extracted upper-level entity relationships, internal attribute facts, and special cross-domain relationship associations to create an initial knowledge subgraph.

[0029] In this invention, the module may have two ways of adding timestamps depending on the situation.

[0030] One approach involves adding timestamps based on the data's origin time. Specifically, when scenario object data enters the time-series appending and fusion module, this module uses the reference time recorded during data encapsulation as the basis for timestamps. Timestamps are then added to the extracted upper-level entity relationships, internal attribute facts, and special cross-domain relationship associations. For example, the interaction time recorded during data encapsulation in the dialogue information generated by a user interacting with a customer service robot becomes the time stamp for the subsequently extracted upper-level entity relationships, internal attribute facts, and special cross-domain relationship associations. Similarly, when the delivery department returns order status information, the recorded information generation time is used to mark the relevant internal attribute facts, ensuring that the timestamps are closely related to the actual time the data was generated and accurately reflecting the timeliness of the information.

[0031] Another approach is to combine event sequence with timestamps. Specifically, the plot object data corresponds to events. For events with a logical sequence, the time-series appending and fusion module adjusts and confirms the timestamps based on the order in which the events occur. In a smart scheduling assistant scenario, if a user first schedules a meeting and then changes the meeting time, the timestamp of the scheduling operation will be earlier than the timestamp of the change operation in the plot object data generated by these two operations. When processing this data, the module will record the development process of the events in the knowledge graph according to this sequence, ensuring the accuracy of the time series for subsequent time-series querying and analysis.

[0032] The system provided by this invention also includes a storage layer, such as a graph database like Neo4j or FalkorDB, to store the knowledge graph, especially the nodes, relationships, and timestamp information of the knowledge graph, ensuring efficient data access and graph traversal performance.

[0033] Furthermore, the present invention also includes a hybrid retrieval interface. The term "hybrid" means that this interface can convert natural language queries into composite queries that combine semantics, keywords, and graph traversal, accurately locating the required context. The composite query combines semantic similarity retrieval, precise keyword matching, and entity relationship path graph traversal to improve retrieval accuracy.

[0034] The hybrid retrieval interface matches the initial knowledge subgraph with timestamps with multiple existing knowledge graphs in the storage layer. Specifically, it can perform semantic similarity retrieval, precise keyword matching, and entity relationship path graph traversal.

[0035] After matching, if a knowledge graph that matches the initial knowledge subgraph exists, then this matching knowledge graph is a structured knowledge context with a time dimension. It is input into the context empowerment module, which makes decisions based on the input knowledge graph, such as state reasoning, multi-turn dialogue, and task planning.

[0036] If the initial knowledge subgraph cannot be matched with a corresponding knowledge graph in the storage layer, nodes are created in the storage layer based on the initial knowledge subgraph. Data bodies / entities (such as users, products, orders, meetings, etc.) are extracted from the initial knowledge subgraph, and a unique node is created for each entity. Node attributes include the core features of the entity (such as the ID and name of "User Y"; the category and model of "Product B"), and are appended with a timestamp generated by the time-series appending and fusion module (i.e., the time when the entity was first extracted).

[0037] Next, the upper-level entity relationships and special cross-domain relationships in the initial knowledge subgraph are extracted, and associated edges are established between the created nodes based on these relationships. Relationship attributes include the relationship type (e.g., "viewed," "appointed," "same type") and the corresponding timestamp (the time the relationship occurred). For example, when user Y views product C for the first time, the system creates a "viewed" relationship between the "User Y" node and the "Product C" node, with the attribute "Time: 2025-08-19 10:00".

[0038] Next, if the newly created node or associated edge has an indirect connection with an existing knowledge graph in the storage layer (even if it is not a complete match), the system will first establish a basic anchor point. For example, if the newly created "Product C" node is identified as "electronic product", it will be temporarily associated with "electronic product" category nodes in the storage layer to prepare for subsequent fusion.

[0039] After new nodes and associated edges are created, they are not directly written to the storage layer. Instead, they are first combined with the existing knowledge graph in the storage layer, and the consistency between the new knowledge and the existing knowledge is checked through preset rules to avoid logical contradictions or conflicts.

[0040] Then, using graph traversal techniques, the newly created nodes / associated edges are compared with the existing knowledge graphs associated with them in the storage layer (such as the historical behavior of the same user and the attribute characteristics of similar entities) through rule checks.

[0041] For example, if the new node is "user Z", then check whether its attributes (such as ID) are duplicates of existing user nodes in the storage layer (excluding the normal case of duplicate names but different IDs).

[0042] For example, if the new associated edge is "User Z - Purchased → Product D", then check if the relationship "User Z - Did not purchase - Product D" exists in the storage layer, and whether there is a logical contradiction in the timestamp.

[0043] For example, if the new entity attribute is "Product D Price: 200 yuan", then check whether there are abnormal fluctuations in the price of similar products in the storage layer (such as prices that are much higher / lower than similar products and have no reasonable timestamp explanation).

[0044] When rule checks detect conflicts between new knowledge (new nodes and new associated edges) and existing knowledge graphs, a pre-defined conflict resolution strategy is activated, prioritizing the retention of more reliable and logically sound information. The conflict resolution strategy needs to be flexibly configured based on the business scenario. Several conflict resolution strategies are possible; two will be described below.

[0045] One conflict resolution strategy is "latest information priority." Specifically, if a conflict stems from attribute differences of the same entity node or related edge at different timestamps (i.e., a "time-series conflict"), the information corresponding to the latest timestamp takes precedence. This strategy is suitable for dynamically changing attributes (such as order status, meeting time, and product price). For example, a user first creates "Meeting C Time: 10:00" (timestamp 9:00) and then modifies it to "Meeting C Time: 11:00" (timestamp 9:30). If the newly created knowledge subgraph shows "Meeting C Time" as 11:00, which conflicts with the old time 10:00 in the storage layer, the system uses the latest timestamp of 9:30 and retains the 11:00 attribute.

[0046] Another conflict resolution strategy is "high-trust source priority." Specifically, if a conflict stems from differences in descriptions of the same entity / related edge from different sources (i.e., "source conflict"), the system prioritizes the information from the source with higher trustworthiness. Source trustworthiness can be preset by the user through the data logical structure (Pydantic Schema) (e.g., "System Log > User Input > Third-Party Interface"). For example, a user inputs "Order A has been shipped" (source trustworthiness: medium), but the delivery system log shows "Order A has not been shipped" (source trustworthiness: high). When these two conflict, the system prioritizes the information from the delivery system log.

[0047] After a conflict is resolved, the new nodes, relationships, and conflict resolution results are not immediately written into the storage layer as the knowledge graph of the storage layer. Instead, these changes are first encapsulated into a "graph change set to be written" for temporary storage.

[0048] In fact, even if the conflict itself does not exist, the newly created nodes and associated edges are not directly written into the knowledge graph of the storage layer. Instead, they are encapsulated as "graph change sets to be written" for temporary storage.

[0049] The so-called "graph change set to be written" is a temporary storage area allocated within the storage layer.

[0050] The main considerations for using temporary storage instead of direct storage are as follows: First, security: avoiding the direct writing of unverified information into the knowledge graph to prevent the spread of errors that could affect the AI ​​agent's decision-making (such as incorrect order status leading to customer service errors); second, traceability: change sets can be traced back and queried, facilitating later auditing (such as investigating why a certain entity's attribute was modified) or fault repair (such as being able to roll back when conflict resolution strategies are misconfigured); third, batch write optimization: for high-frequency dynamic data (such as real-time order updates), temporary change sets can be written to the storage layer in batches according to time windows, reducing the real-time write pressure on the graph database and improving system performance.

[0051] The temporary "graph change set to be written" undergoes conflict checking again as mentioned above, i.e., a second check. If there are no more conflicts, it is officially written into the storage layer as the knowledge graph of the storage layer. If the result of the second check is that there are conflicts, it is directly discarded.

[0052] Then, new nodes and relationships are created, and data consistency is checked according to temporal rules based on existing knowledge graph data. If a consistency conflict occurs, a conflict resolution strategy is executed, and the graph change set to be written is temporarily stored.

[0053] Figure 2 A flowchart illustrating the operation of a dynamic AI knowledge graph system according to the present invention is shown. The above-described flow of the present invention follows... Figure 2 Execute as shown.

[0054] As described above, this invention provides a dynamic AI knowledge graph system, relating to the fields of artificial intelligence and knowledge graph technology. This system aims to address the problems of static information, missing relationships, and fragmented context in existing AI memory solutions. Through automated knowledge construction, dynamic time awareness, real-time response to changes, and hybrid retrieval and reasoning capabilities, it provides AI agents with an efficient, dynamic, and long-term evolving "memory system." This invention achieves real-time transformation of unstructured and structured information into a dynamic knowledge network with time awareness through core processes such as data encapsulation, knowledge pipeline processing, temporal appending and fusion, multimodal retrieval, and context empowerment. This significantly improves the performance of AI agents in multi-turn interactions, state reasoning, and task planning.

[0055] The above description is merely an exemplary embodiment of the present invention and is not intended to limit 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 dynamic AI knowledge graph system, characterized in that, The system comprises: a data packaging module, wherein target data is input into the data packaging module and packaged into a standardized plot object by the data packaging module, the target data including unstructured data and structured data; a knowledge pipeline processing module, wherein the knowledge pipeline processing module calls a large language model and calls a user-defined data logic structure, the user-defined data logic structure defining extraction rules and data structure templates for the standardized plot object, the large language model extracting upper entity relationships, internal attribute facts, and special cross-domain relationship associations from the standardized plot object to form an extraction result, and the extraction rules and data structure templates ensuring consistency of the extraction result of the standardized plot object; a time sequence addition and fusion module, wherein the time sequence addition and fusion module adds a timestamp to the upper entity relationships, internal attribute facts, and special cross-domain relationship associations to create an initial knowledge subgraph; a hybrid retrieval interface and storage layer, wherein the hybrid retrieval interface matches the initial knowledge subgraph with a plurality of knowledge graphs in the storage layer; a context empowerment module, if there is a knowledge graph in the storage layer that matches the initial knowledge subgraph, the matched knowledge graph is input into the context empowerment module, and the context empowerment module makes a decision based on the input knowledge graph, if the initial knowledge subgraph cannot match a corresponding knowledge graph in the storage layer, entities in the initial knowledge subgraph are extracted, unique nodes are created for each entity, attributes of the unique nodes are added with the timestamp generated by the time sequence addition and fusion module, upper entity relationships and special cross-domain relationships in the initial knowledge subgraph are extracted, and associated edges are established between the created nodes based on the upper entity relationships and special cross-domain relationships, after the new nodes and associated edges are created, they are not directly written into the storage layer, but are first combined with the existing knowledge graphs in the storage layer, and the consistency of the existing knowledge graphs in the storage layer and the new knowledge, i.e., the newly created nodes and associated edges, is checked through preset extraction rules.

2. The system of claim 1, wherein, When the rule check finds that the new knowledge conflicts with the knowledge graphs, a preset conflict resolution strategy is started.

3. The system of claim 2, wherein, If the conflict is in the attribute difference of the nodes or associated edges of the same entity at different timestamps, the information corresponding to the latest timestamp is used as the reference.

Images