Multi-dimensional dynamic AI agent knowledge loading and processing method and system

Abstract

The invention relates to a multi-dimensional dynamic AI agent knowledge loading and processing method and system. The method comprises the following steps: performing intention recognition, disassembly and dynamic context calibration on a question input by a user, generating a plurality of retrieval sub-tasks, and distributing a heterogeneous knowledge base; according to the retrieval sub-task, performing multi-level retrieval in the graph, the vector, the memory knowledge base and the internet data source to obtain candidate knowledge; performing aging, security, correlation and conflict four-dimensional evaluation and dynamic weight optimization on the candidate knowledge to obtain sorted target knowledge; and splicing the target knowledge according to the structured mode and then outputting the spliced target knowledge. According to the scheme, full coverage of professional knowledge, accurate answers, thousands of people and thousands of faces and high credibility are realized through graph-vector-memory three-library collaboration, context sensing and configurable priority strategies, hot plug expansion is supported, and domain question and answer services can be quickly constructed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a multi-dimensional dynamic AI agent knowledge loading and processing method and system, a computer readable storage medium and an electronic device. BACKGROUND

[0002] With the rapid development of large model technology, AI agents have become the core interactive entrance of various online services, and their output quality is highly dependent on the underlying knowledge loading and processing link. However, the existing technical solutions generally have the following significant defects when facing professional scenarios, personalized needs and high-trust applications, which seriously restrict the accuracy, completeness and applicability of agent knowledge acquisition. The specific manifestations are as follows: I. Insufficient completeness of knowledge answer In the existing technology, the knowledge source of AI agents mainly relies on single knowledge base storage or Internet retrieval, lacking a collaborative scheduling mechanism for multi-source heterogeneous knowledge (such as structured databases, unstructured texts, semi-structured knowledge graphs, etc.). Due to the lack of effective integration of the relevance between different sources of knowledge, the coverage of professional domain knowledge is limited, making it difficult to fully cover the multi-dimensional information required for complex problems. At the same time, due to the lack of deduplication and conflict detection mechanisms, knowledge duplication and redundancy or contradictions are prone to occur, ultimately resulting in one-sidedness or errors in knowledge answers (see CN202510285798.7).

[0003] II. Lack of personalized knowledge adaptation ability The existing knowledge loading method usually outputs fixedly based on general knowledge bases, without dynamic adjustment according to the individual characteristics of users (such as identity attributes, professional backgrounds, interest preferences, core concerns, etc.). The agent is difficult to identify the specific demand context and background differences of users, resulting in a "one-size-fits-all" feature of knowledge output, which cannot meet the customized needs of different users in professional scenarios, and the adaptability of knowledge to users is seriously insufficient (see CN202410521733.3).

[0004] III. Lack of graph correlation divergence ability The current knowledge system of AI agents generally does not integrate graph structure knowledge bases (i.e. graph knowledge bases), and the semantic association and logical relationship between knowledge are not explicitly modeled. When dealing with professional problems, the agent is difficult to perform divergent reasoning (such as cross-domain knowledge transfer, upper and lower concept expansion, causal chain derivation, etc.) through the association between knowledge nodes, and can only respond based on isolated fragmented knowledge, unable to form comprehensive and in-depth professional problem solving ability.

[0005] IV. Incomplete knowledge quality control mechanism The prior art lacks a multi-dimensional comprehensive evaluation system for knowledge quality, and does not systematically consider key indicators such as timeliness (such as data update frequency, field frontiers), security (such as sensitive information filtering, compliance review), and relevance (such as the degree of matching with user questions). Due to the lack of dynamic verification and filtering mechanism, low-quality or invalid knowledge may be incorrectly called, directly leading to unreliable conclusions output by the intelligent agent (see CN202411990966.4).

[0006] V. Knowledge weight optimization method is extensive After multi-source knowledge recall, existing methods usually use a simple rerank model to reorder the knowledge, without deeply processing complex relationships such as redundancy (such as different representations of the same fact), conflict (such as contradictory information from different sources), and business priority (such as weight allocation of domain expert experience and general data). This extensive sorting method may result in high-value knowledge being incorrectly down-weighted, while low-relevance or low-reliability knowledge is preferentially called, ultimately affecting the knowledge application effect of the intelligent agent (see patent CN202411990966.4).

[0007] In summary, the existing AI intelligent agent knowledge loading and processing technology has systematic defects in knowledge coverage, individual adaptation, associated reasoning, quality control, and weight optimization, resulting in the knowledge obtained by the intelligent agent often presenting the characteristics of inaccuracy, incompleteness, and inapplicability, which seriously restricts the knowledge application efficiency of the intelligent agent in professional scenarios. An intelligent, multi-dimensional knowledge loading and processing method is urgently needed to solve the above problems. SUMMARY

[0008] To solve the above problems, the present application proposes a new multi-dimensional dynamic AI intelligent agent knowledge loading and processing method and system, more specifically, provides a method and system for precise loading of knowledge in an intelligent question answering system through intent recognition, multi-source data fusion, and dynamic strategy evaluation.

[0009] To achieve the above purpose, the following technical strategies are adopted in the present application: (1) Dynamic multi-dimensional knowledge fusion mechanism: for the first time, graph knowledge base, vectorized knowledge base, and memory knowledge base are jointly included in the retrieval range, dynamic retrieval strategy is realized, and the limitation of knowledge dimension in traditional methods is broken through.

[0010] (2) Dynamic context perception mechanism: develop an intent calibration algorithm based on attention mechanism to solve the misjudgment problem of single intent recognition model.

[0011] (3) Professional association divergent retrieval method: construct a business knowledge graph, based on a multi-hop reasoning framework, realize deep retrieval of knowledge (such as cooperative retrieval of graph knowledge base and vector knowledge base, improve the coverage of knowledge retrieval).

[0012] (4) Security filtering strategy: supplement the scene security, role security strategy on the basis of social security, so that the security filtering is more suitable for business.

[0013] (5) Knowledge weight strategy: propose a configurable business priority strategy scheme, form a dynamic weight distribution algorithm, and solve the problem of knowledge redundancy and conflict.

[0014] Specifically, the application provides the following technical solutions: The first aspect of the application provides a multi-dimensional dynamic AI agent knowledge loading and processing method, as shown in Figure 1 The method comprises the following steps: S1. Question preprocessing: performing intent type identification, intent disassembly and dynamic context calibration on the user input question, and generating a plurality of retrieval sub-tasks based on the calibrated intent, and assigning a corresponding target heterogeneous knowledge base to each retrieval sub-task; S2. Multi-source knowledge retrieval: performing multi-level retrieval in parallel or in series in a heterogeneous knowledge base set comprising at least a graph knowledge base, a vector knowledge base, a memory knowledge base and an Internet data source according to the retrieval sub-tasks, to obtain candidate knowledge fragments; S3. Strategy-based knowledge processing: performing multi-dimensional evaluation, conflict resolution and weight optimization on the candidate knowledge fragments to generate a sorted and filtered target knowledge set; S4. Knowledge output: assembling and returning the target knowledge set to the user according to a preset structured mode.

[0015] Optionally, in the method of the application, step S1 comprises: S11. Intent type identification: dividing the user input question into basic question and answer type, professional analysis type and complex reasoning type by using semantic analysis and domain classification model; S12. Intent disassembly: using a BERT-based intent decomposition algorithm to split the question into sub-intent units in time, space and business scenario dimensions; S13. Dynamic context calibration: introducing historical dialogue context correction or confirmation intent, and triggering user interaction when there is ambiguity; S14. Extracting the core elements of the question, splitting the question into a plurality of different retrieval sub-tasks, and specifying the target heterogeneous knowledge base corresponding to each sub-task.

[0016] Optionally, in the method of the application, step S2 comprises: S21. Quick retrieval: preferentially retrieving the question library (storing system historical high-quality question and answer data) and the memory library (storing user historical interaction data and personalized settings), and if a high-quality historical answer is hit, directly entering step S4; S22. Heterogeneous database dynamic retrieval: if the quick retrieval is not hit, the graph database, vector database (based on Faiss semantic similarity matching), third-party API interface (such as Tools / MCP interface), and Internet data source (integrated with Selenium+BeautifulSoup crawler) are dynamically selected according to the retrieval subtask and the matching degree of the knowledge base for retrieval.

[0017] Optionally, in the method of the application, the graph database in step S2 is stored by Neo4j, and secondary vector retrieval is performed on the retrieval results to expand the associated knowledge.

[0018] Optionally, in the method of the application, the multi-dimensional evaluation in step S3 includes: (1) timeliness evaluation: setting a domain-related timeliness threshold and comparing the knowledge generation time with the current time; (2) security evaluation: multi-layer filtering through sensitive word detection (regular expression + SnowNLP), scene matching verification, and user role characteristic verification; (3) relevance evaluation: using a BiLSTM-CRF model to calculate the matching degree of the knowledge fragment and the detection subtask; (4) conflict resolution: when the knowledge is contradictory, the high-priority knowledge is retained according to the priority order of “user-defined knowledge > expert-verified knowledge > high-confidence Internet data”.

[0019] Optionally, in the method of the application, the weight optimization in step S3 includes: constructing a knowledge feature matrix containing credibility R, timeliness T, relevance C, and business priority P, and using an improved PageRank algorithm to calculate the comprehensive weight W, the calculation formula being W = αR + βT + γC + δP, wherein α, β, γ, and δ are weight coefficients corresponding to each parameter, which are dynamically adjusted according to the domain, and α + β + γ + δ = 1.

[0020] Optionally, in the method of the application, the preset structured mode in step S4 is JSON Schema, and the fields at least include subtask identification, answer body, time record, weight label, source database, and expansion association point.

[0021] The second aspect of the application provides a multi-dimensional dynamic AI intelligent agent knowledge loading and processing system, which comprises: a question preprocessing module for performing intent type recognition, intent disassembly, and dynamic context calibration on the user input question, generating a plurality of retrieval subtasks based on the calibrated intent, and assigning a corresponding heterogeneous knowledge base set to each retrieval subtask; Multi-source knowledge retrieval module: used for performing multi-level retrieval in parallel or in series in a heterogeneous knowledge base set containing at least a graph knowledge base, a vector knowledge base, a memory knowledge base and an Internet data source according to the retrieval subtask, to obtain candidate knowledge segments; Strategic knowledge processing module: used for performing multi-dimensional evaluation, conflict resolution and weight optimization on the candidate knowledge segments to generate a sorted and filtered target knowledge set; Knowledge output module: used for assembling and returning the target knowledge set to the user in a preset structured mode.

[0022] The system realizes the steps of the aforementioned multi-dimensional dynamic AI agent knowledge loading and processing method during operation.

[0023] The third aspect of the present application provides an electronic device, comprising a memory and a processor. Memory: used for storing a computer program; Processor: used for executing the computer program to realize the steps of the aforementioned multi-dimensional dynamic AI agent knowledge loading and processing method.

[0024] The fourth aspect of the present application provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to realize the steps of the aforementioned multi-dimensional dynamic AI agent knowledge loading and processing method.

[0025] In summary, compared with the prior art, the multi-dimensional dynamic AI agent knowledge loading and processing method has the following advantages: (1) Comprehensive knowledge coverage is improved Through the "graph-vector-memory" three-library unified retrieval framework, the graph knowledge base, the vectorized knowledge base and the user private memory library are first jointly included in the dynamic recall range, and the heterogeneous knowledge is cooperatively scheduled based on the multi-hop reasoning and semantic complementary strategy, which significantly expands the professional field knowledge coverage and overcomes the blind area caused by traditional single-source recall.

[0026] (2) Answer accuracy is enhanced The dynamic weight distribution algorithm with configurable business priority is introduced to complete the duplicate detection, conflict resolution and scene relevance re-scoring in the rerank stage, so that the high-value knowledge is always on top, the generation illusion caused by sorting error is effectively reduced, and the final answer accuracy is improved.

[0027] (3) Personalized service capability is enhanced The user-level memory library is constructed and the intention calibration model based on the attention mechanism is matched, the user identity, historical interaction and attention preference are fused in real time, the knowledge recall and output style self-adaptation of thousands of people are realized, and the drawbacks of the general knowledge "same as everyone else" are solved.

[0028] (4) Professional problem handling ability is enhanced Relying on the multi-hop association and version tracking mechanism of the business knowledge graph, the amendment history of legal articles, the evolution path of cases, and the interpretation provisions can be automatically associated in legal consultation scenarios, realizing deep, divergent, and traceable professional-level answers, and making up for the defects of traditional vector retrieval that cannot capture long-chain causality.

[0029] (5) Knowledge reliability guarantee A four-dimensional evaluation system of "timeliness-authority-safety-scenario" is established to filter outdated, illegal, or low-confidence knowledge at each layer before recall, sorting, and generation. The safety module adds role safety and scenario safety strategies to ensure that the output content is both compliant and contextually relevant.

[0030] (6) Improved system scalability and flexibility The overall modular and pluggable architecture supports zero-code access to new libraries and on-demand adjustment of processing procedures, providing a technical foundation for quickly building customized knowledge loading services for different businesses.

[0031] Other features and advantages of the present application will be described in detail in the subsequent description, or can be understood by implementing the relevant technical solutions of the present application. The purpose and other advantages of the present application can be achieved by the technical features and technical means explicitly indicated in the specification, claims and drawings, and obtained by the implementation process of these technical contents. BRIEF DESCRIPTION OF DRAWINGS

[0032] In order to more clearly set forth the technical solutions of the present application, the following will briefly introduce the drawings involved in the embodiment description. It should be pointed out that the drawings only show some embodiments of the present application. For those skilled in the art, other related drawings can be derived from these drawings without creative labor.

[0033] Figure 1 The implementation flowchart of the multi-dimensional dynamic AI agent knowledge loading and processing method of the present application.

[0034] Figure 2 The overall design architecture diagram of the multi-dimensional dynamic AI agent knowledge loading and processing method provided by the embodiment of the present application.

[0035] Figure 3 The composition structure diagram of the multi-dimensional dynamic AI agent knowledge loading and processing system of the present application.

[0036] Figure 4 The structural schematic diagram of the electronic device provided by the embodiment of the present application. DETAILED DESCRIPTION

[0037] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the drawings in the embodiments of the present application. It should be noted that the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative effort fall within the protection scope of the present application.

[0038] In this document, the term “comprising” and any form of the term “comprise”, such as “comprise”, “comprises” and “comprising”, are open-ended terms that are to be interpreted to mean “including but not limited to”, that is, an open-ended list that is not exhaustive. The term “based on” is to be interpreted as “based, at least in part, on”, that is, the basis or condition referred to may not be the only factor, and other related factors may also be involved. The term “one embodiment” is to be interpreted as “at least one embodiment”, that is, the described embodiment is not the only possible implementation, and other similar embodiments may also exist.

[0039] In this document, the terms “one” and “multiple” used to modify relevant elements or features are illustrative rather than restrictive. Unless explicitly stated otherwise in the context, “one” is to be understood as “at least one”, and “multiple” is to be understood as “at least two”. A person of ordinary skill in the art should reasonably interpret these terms according to the semantics and logical relationship of the context, and ensure that they cover the possibility of “one or more”.

[0040] Embodiment: A multi-dimensional dynamic AI agent knowledge loading and processing method The core process of the method is as follows: (I) Problem preprocessing stage 1. Intention type identification: Through semantic analysis and domain classification model, the input question is divided into three types: basic question and answer type, professional analysis type and complex reasoning type.

[0041] 2. Intention decomposition and analysis: The intention decomposition algorithm based on BERT is used to decompose the composite question into multiple sub-intention units (such as time dimension, space dimension, business scenario dimension, etc.).

[0042] 3. Dynamic context calibration: If there is ambiguity in intention analysis, historical dialogue context is introduced for intention correction, or user confirmation interaction is triggered (such as “Do you mean the content of XX aspect?”).

[0043] 4, extract the key elements of the problem and disassemble the retrieval task: abstract the core elements in the question for subsequent auxiliary retrieval and knowledge processing, and split the problem into different retrieval sub-tasks and clearly define the target "heterogeneous knowledge base". Different tasks have different "heterogeneous knowledge bases".

[0044] (II) Multi-source knowledge retrieval stage Fast retrieval: First, retrieve the problem library (store all system users' high-quality historical problems and answers) and memory library (store user's historical interaction data, personalized setting data), confirm whether the problem can be answered directly based on the problem library, if it can be directly answered, it can greatly improve the retrieval and processing efficiency, and directly enter the knowledge assembly stage, and then improve the knowledge response efficiency.

[0045] If it cannot be directly answered, then according to the matching degree of each sub-task to the heterogeneous database, dynamic retrieval is performed, including: retrieving the memory knowledge base (storing the user's historical interaction data); Retrieving the vector knowledge base (based on Faiss semantic similarity matching); Calling Tools / MCP interface (real-time calling third-party API); Internet search (integrating Selenium+BeautifulSoup crawler); Retrieving the graph knowledge base (Neo4j graph database professional relationship network), and performing secondary use of the graph knowledge data (which can further retrieve the vector knowledge base to enrich the associated knowledge retrieved).

[0046] (III) Strategic knowledge processing stage 1. Four-dimensional evaluation strategy of knowledge (1) Timeliness judgment: Set the timeliness threshold of various knowledge (such as the effective period of news knowledge ≤30 days), get the knowledge generation time, record time, and compare with the current time to determine whether the knowledge needs to be discarded.

[0047] (2) Safety filtering: Through sensitive word detection (regular expression + SnowNLP), scene matching verification, user role characteristic verification (such as a student asking a question, there are knowledge points that only college students can learn), and other multiple dimensions to determine whether the knowledge exists. It is not suitable to output.

[0048] (3) Relevance verification: Use BiLSTM-CRF model to preliminarily calculate the matching degree of knowledge fragments and sub-tasks.

[0049] (4) Conflict resolution strategy: When there is a conflict in the knowledge, the knowledge with high priority is preferred according to the priority strategy, such as priority sorting rules: user-defined knowledge > domain expert certified knowledge > high confidence internet data.

[0050] 2. Knowledge weight optimization strategy Build a knowledge feature matrix (including credibility, timeliness, relevance, business priority), use an improved PageRank algorithm to calculate the comprehensive weight (formula: W = aR + bT + gC + dP, where a + b + g + d = 1), dynamically adjust the parameter coefficient (such as medical field d = 0.4, e-commerce field a = 0.6).

[0051] 3. Knowledge output strategy (1) Output control: According to the output control strategy of knowledge, decide whether to perform knowledge interception and knowledge summary and refinement.

[0052] (2) Structured assembly: Organize the filtered knowledge according to the JSON Schema format, and each knowledge contains sub-tasks, answer body, time record, weight label, source knowledge base, and other extended association points, etc. Field information.

[0053] 4. Alternative solution When the graph knowledge base is not hit, the entity relationship reasoning module based on Word2Vec can be enabled.

[0054] In the resource-limited scenario, the Internet retrieval module can be closed, and only local heterogeneous knowledge base retrieval is retained.

[0055] In order to more clearly illustrate the technical solutions of the present application, the following will be further illustrated through specific examples of specific scenarios.

[0056] Figure 2 The overall design architecture and implementation process of the multi-dimensional dynamic AI intelligent agent knowledge loading and processing method provided by the embodiment are shown, including the following steps: Step 1: Problem input First, clarify the user's question content.

[0057] Step 2: Intention type identification The intention type of the user's question is identified, and the input question is identified and classified into three categories through semantic analysis and domain classification model: basic question and answer type, professional analysis type and complex reasoning type, providing an intention framework for subsequent processing.

[0058] Step 3: Intent resolution and analysis The identified intention is "intent resolution and analysis". The intention decomposition algorithm based on BERT is used to decompose the complex question into multiple sub-intention units (such as time, space, business scenario, etc. Dimension), realizing the structured analysis of the question.

[0059] Step 4: Dynamic context calibration Check the intention integrity and certainty, introduce historical dialogue context to assist in correcting the intention, or trigger user confirmation interaction (such as prompting "Do you mean XX aspect?") if necessary, to ensure accurate intention parsing. Then further check and perfect the question and intention through question correction, question rewriting, question expansion, and question checking, etc. independent capabilities.

[0060] Step 5: Extract question elements and decompose retrieval tasks Abstract the core elements in the calibrated question for subsequent knowledge retrieval and processing. At the same time, split the question into different retrieval sub-tasks, clarify the core elements of each sub-task, and clarify the "heterogeneous knowledge base" that each sub-task needs to retrieve (different tasks, different target knowledge bases).

[0061] Step 6: Multi-source knowledge retrieval This stage is divided into two sub-processes: 1. Quick retrieval: First, search the question bank (store all users' high-quality historical question and answer data) and memory bank (store the user's historical interaction data and personalized settings), and determine whether the question can be answered directly. If it can be answered directly, skip the subsequent retrieval and go directly to the knowledge assembly stage. Quick retrieval can significantly improve efficiency.

[0062] 2. Multi-type knowledge base retrieval: If the quick retrieval determines that it cannot be answered directly, perform multi-source retrieval dynamically according to the matching degree of each sub-task and heterogeneous knowledge base, including: (1) Retrieve the memory knowledge base (user's historical interaction data); (2) Retrieve the vector knowledge base (based on Faiss semantic similarity matching); (3) Call Tools / MCP interface (real-time call third-party API); (4) Internet search (integrated Selenium+BeautifulSoup crawler); (5) Retrieve the graph knowledge base (Neo4j graph database professional relationship network), and can further retrieve the vector knowledge base to enrich the associated knowledge.

[0063] Step 7: Strategic knowledge processing All the knowledge retrieved is processed with multiple strategies: 1. Knowledge four-dimensional evaluation strategy: Time effectiveness judgment: Set the time threshold of each type of knowledge (such as news ≤ 30 days), compare the knowledge generation time, record time and current time, and decide whether to discard the expired knowledge.

[0064] Security filtering: sensitive word detection (regular expression + SnowNLP), scene matching verification, user role characteristic verification (such as primary school students accessing university knowledge points), and other multi-dimensional filtering of inappropriate output content.

[0065] Relevance verification: use BiLSTM-CRF model to preliminarily calculate the matching degree of knowledge fragments and subtasks.

[0066] Conflict resolution strategy: when the knowledge is contradictory, the high-priority knowledge is trusted according to the priority rule (user-defined knowledge> domain expert certified knowledge> high-confidence Internet data).

[0067] 2、Weight optimization strategy: Construct a knowledge feature matrix (including credibility, timeliness, relevance, business priority), use an improved PageRank algorithm to calculate the comprehensive weight (formula: W = αR + βT + γC + δP, where α+β+γ+δ=1), and dynamically adjust the parameter coefficients (such as δ=0.4 in the medical field, α=0.6 in the e-commerce field), Finally, all available knowledge weights and rankings are formed.

[0068] Step 8: Knowledge output strategy Processed knowledge is output controlled and structured assembled: Output control: According to the context token control strategy, decide whether to perform knowledge interception and summary refinement.

[0069] Structured assembly: Organize the filtered knowledge according to the JSON Schema format, including subtasks, answer bodies, time records, weight labels, source knowledge bases, and other extended association points.

[0070] Step 9: Knowledge output The processed knowledge is returned to the user through the "knowledge output" module, and the process ends.

[0071] Figure 3 The system includes: Question preprocessing module: used for intent type recognition, intent disassembly and dynamic context calibration of user input questions, and generating a number of retrieval subtasks based on the calibrated intent, and assigning a corresponding heterogeneous knowledge base set to each retrieval subtask; Multi-source knowledge retrieval module: used for multi-level retrieval in parallel or series in a heterogeneous knowledge base set containing at least graph knowledge base, vector knowledge base, memory knowledge base and Internet data source according to the retrieval subtask, to obtain candidate knowledge fragments; A policy-based knowledge processing module is configured to perform multi-dimensional evaluation, conflict resolution and weight optimization on the candidate knowledge fragments to generate a ranked and filtered target knowledge set. A knowledge output module is configured to assemble the target knowledge set into a preset structured mode and return it to the user.

[0072] The above system realizes the steps of the multi-dimensional dynamic AI agent knowledge loading and processing method disclosed in the present application.

[0073] The flowcharts and block diagrams in the drawings show possible implementations of systems, methods and computer program products according to various embodiments of the present application, including architecture, functions and operations. In these diagrams, each block can represent a module, a program segment or a part of code, which contains one or more executable instructions for implementing the specified logic function. It should be noted that each block in the block diagram and / or flowchart and the combination of these blocks can use a dedicated hardware-based system to implement the specified function or operation, or a combination of dedicated hardware and computer instructions.

[0074] As shown in Figure 4 The embodiments of the present application also disclose an electronic device, which comprises a processor 310, a communication interface 320, a memory 330 for storing processor-executable computer programs and a communication bus 340. Wherein the processor 310, the communication interface 320 and the memory 330 complete mutual communication through the communication bus 340. The processor 310 realizes the steps of the multi-dimensional dynamic AI agent knowledge loading and processing method disclosed above by running the executable computer programs.

[0075] It can be understood that, in addition to the memory and the processor, the electronic device can also include input devices (such as a keyboard), output devices (such as a display) and other communication modules. These input devices, output devices and other communication modules all communicate with the processor through an I / O interface (i.e. an input / output interface).

[0076] The operations of the present application can be realized by using one or more program design languages or combinations thereof to write computer program codes. The program design languages include but are not limited to the following types: Object-oriented program design languages, such as Java, Smalltalk, C++ and the like; Conventional procedural program design languages, such as the "C" language or similar program design languages.

[0077] The execution modes of the program codes include but are not limited to: Completely on the user computer; Some portions are executed on the user's computer, some portions are executed on a remote computer; As a standalone software package; Completely on a remote computer or server.

[0078] In scenarios involving a remote computer, the remote computer can be connected to the user's computer through any type of network, including but not limited to a local area network (LAN) or a wide area network (WAN). In addition, the remote computer can also be connected to an external computer through an Internet service provider, such as through the Internet.

[0079] Further, the present application also discloses a computer readable storage medium, when the instructions in the computer readable storage medium are executed by the processor of the electronic device, the electronic device can execute each step of the multi-dimensional dynamic AI agent knowledge loading and processing method disclosed in the present application.

[0080] In the context of the present application, computer readable storage medium refers to a tangible medium capable of storing computer program code and related data. Specific examples include but are not limited to the following: (1) Portable computer disk: such as floppy disk and other removable magnetic storage medium.

[0081] (2) Hard disk: including mechanical hard disk and solid state disk and other fixed storage device.

[0082] (3) Random access memory (RAM): a volatile storage medium used to temporarily store data and program code.

[0083] (4) Read-only memory (ROM): a non-volatile storage medium used to store fixed programs and data.

[0084] (5) Erasable programmable read-only memory (EPROM) or flash memory: a non-volatile storage medium that supports multiple erasing and programming.

[0085] (6) Fiber optic storage device: storage medium based on fiber optic technology.

[0086] (7) Portable compact disc read-only memory (CD-ROM): a read-only medium that stores data in the form of an optical disc.

[0087] (8) Optical storage device: such as DVD, Blu-ray disc and other storage media based on optical principles.

[0088] (9) Magnetic storage device: such as magnetic tape, magnetic disk and other storage media based on magnetic principles.

[0089] (10) Any suitable combination of the above: for example, using a combination of storage media to meet different storage requirements.

[0090] These computer readable storage media can be used for storing the program codes and related data described in this application to support the running of the program and the persistent storage of the data.

[0091] In particular, according to the embodiments of the present application, the processes described in the flowcharts can be implemented as a computer software program. For example, the embodiments of the present application relate to a computer program product comprising a computer program carried on a non-transitory computer readable medium. The computer program contains program codes for executing the multi-dimensional dynamic AI agent knowledge loading and processing method disclosed in the present application. When the computer program is executed by a processing system, the above-mentioned functions defined in the embodiments of the present application can be realized.

[0092] Although the above discussion contains a number of specific implementation details, these details should not be construed as limiting the scope of the present application. The above description is merely a preferred embodiment of the present application and an explanation of the principles of the technology used. Those skilled in the art should understand that the scope of the disclosure involved in the present application is not limited to the technical solutions formed by the specific combination of the above technical features. At the same time, the present application should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above disclosed concept.

[0093] Those skilled in the art should also understand that they can modify the technical solutions recorded in the foregoing embodiments or replace some of the technical features with equivalent features without departing from the spirit and scope of the technical solutions of the embodiments of the present application. These modifications or replacements will not make the corresponding technical solutions deviate from the core spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for multi-dimensional dynamic AI agent knowledge loading and processing, characterized in that, The method comprises: S1. Question preprocessing: intent type recognition, intent disassembly and dynamic context calibration are performed on the user input question, and a plurality of retrieval subtasks are generated based on the calibrated intent, and each retrieval subtask is assigned to a corresponding target heterogeneous knowledge base; S2. Multi-source knowledge retrieval: according to the retrieval subtask, multi-level retrieval is performed in parallel or in series in a heterogeneous knowledge base set including at least a graph knowledge base, a vector knowledge base, a memory knowledge base and an Internet data source, to obtain candidate knowledge fragments; S3. Strategy knowledge processing: multi-dimensional evaluation, conflict resolution and weight optimization are performed on the candidate knowledge fragments to generate a sorted and filtered target knowledge set; S4. Knowledge output: the target knowledge set is assembled according to a preset structured mode and returned to the user.

2. The method of claim 1, wherein, Step S1 includes: S11. Intent type recognition: semantic analysis and domain classification model is used to divide the user input question into basic question and answer type, professional analysis type and complex reasoning type; S12. Intent disassembly: BERT-based intent decomposition algorithm is used to split the question into time, space, business scenario dimension sub-intent units; S13. Dynamic context calibration: historical dialogue context correction or confirmation intent is introduced, and user interaction is triggered in ambiguity; S14. Extract the core elements of the question, split the question into several different retrieval subtasks, and specify the target heterogeneous knowledge base corresponding to each subtask.

3. The method of claim 1, wherein, Step S2 includes: S21. Quick retrieval: preferentially retrieve the problem library and the memory library, and if a high-quality historical answer is hit, go directly to step S4; S22. Dynamic retrieval of heterogeneous databases: if the quick retrieval does not hit, dynamically select the graph knowledge base, the vector knowledge base, the third-party API interface and the Internet data source for retrieval according to the retrieval subtask and the matching degree of the knowledge base.

4. The method of claim 1, wherein, The graph knowledge base in step S2 is stored by Neo4j, and secondary vector retrieval is supported to expand associated knowledge.

5. The method of claim 1, wherein, The multi-dimensional evaluation in step S3 includes: (1) timeliness evaluation: set the domain-related timeliness threshold and compare the knowledge generation time with the current time; (2) safety evaluation: multi-layer filtering is performed through sensitive word detection, scene matching verification and user role characteristic verification; (3) relevance evaluation: BiLSTM-CRF model is used to calculate the matching degree of knowledge fragments and detection subtasks; (4) Conflict resolution: when the knowledge is contradictory, the high-priority knowledge is retained according to the priority order of "user-defined knowledge>expert-certified knowledge>high-confidence Internet data".

6. The method of claim 1, wherein, The weight optimization in step S3 includes: constructing a knowledge feature matrix including credibility R, timeliness T, relevance C and business priority P, and calculating the comprehensive weight W by using the improved PageRank algorithm, the calculation formula is W=αR+βT+γC+δP, wherein α, β, γ, δ are weight coefficients corresponding to each parameter, which are dynamically adjusted according to the field, and α+β+γ+δ=1.

7. The method of claim 1, wherein, The preset structured mode in step S4 is JSONSchema, and the fields at least include subtask identification, answer body, time record, weight label, source knowledge base, and extended association point.

8. A multi-dimensional dynamic AI agent knowledge loading and processing system, characterized in that, The system comprises: A question preprocessing module for performing intent type recognition, intent disassembly, and dynamic context calibration on a question input by a user, and generating a plurality of retrieval subtasks based on the calibrated intent, and assigning a corresponding heterogeneous knowledge base set to each retrieval subtask; A multi-source knowledge retrieval module for performing multi-level retrieval in parallel or in series in a heterogeneous knowledge base set at least including a graph knowledge base, a vector knowledge base, a memory knowledge base, and an Internet data source according to the retrieval subtasks, to obtain candidate knowledge segments; A strategized knowledge processing module for performing multi-dimensional evaluation, conflict resolution, and weight optimization on the candidate knowledge segments to generate a sorted and filtered target knowledge set; A knowledge output module for assembling the target knowledge set according to a preset structured mode and returning it to the user.

9. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the multi-dimensional dynamic AI agent knowledge loading and processing method according to any one of claims 1-7.

10. An electronic device, comprising: Comprise: A memory and a processor; The memory is used for storing a computer program; The processor is used for executing the computer program to implement the steps of the multi-dimensional dynamic AI agent knowledge loading and processing method according to any one of claims 1-7.

Images