An agent-driven automated knowledge base maintenance method

Abstract

The application discloses an agent-driven automatic knowledge base maintenance method and system. The method comprises the following steps: generating a knowledge card by an agent, giving a confidence score differentiated according to the authority degree of a knowledge source and a decay rate set according to the characteristics of a knowledge field, and establishing a bidirectional link between strong-type knowledge; performing duplicate search and conflict analysis before writing new knowledge, merging and fusing knowledge with logical consistency and increasing the confidence, and creating a replacement card and establishing a replacement bidirectional link for knowledge with logical conflicts; awakening the agent to automatically perform archiving, including incremental fusion, context compression and source file backup, by a timing trigger or an external scheduling system. Through the combination of the above mechanisms, the application realizes the automatic maintenance of the knowledge base and the continuous increase of the knowledge precision.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and knowledge management technology, specifically involving an automated knowledge base maintenance method and system driven by multi-agent collaboration, and particularly involving closed-loop knowledge maintenance technology such as knowledge confidence scoring and time decay, anti-island bidirectional linking, incremental fusion and knowledge accuracy increase, context high-concentration compression, immutable original cold backup, and timed triggering automatic archiving. Background Technology

[0002] With the rapid development of Large Language Models (LLM) and Multi-Agent Systems (MAS), the automatic collection, organization, and maintenance of knowledge bases by intelligent agents has become an important application scenario for artificial intelligence. However, existing knowledge base maintenance technologies have the following significant drawbacks.

[0003] Deficiency 1: Lack of confidence assessment and timeliness management of knowledge leads to outdated knowledge continuously misleading decision-making. Existing knowledge management systems do not quantitatively assess the credibility of knowledge entries, nor do they track their timeliness. All knowledge entries are considered equivalently valid once created, regardless of the authority of their source, the number of times they have been verified, or the time of their creation. When knowledge becomes outdated or is overturned, the old entries still coexist with the latest knowledge, lacking an automatic decay and elimination mechanism. This causes downstream decision-makers to make judgments based on outdated or erroneous knowledge, resulting in information pollution. In particular, knowledge in the technical field (such as API changes, framework version upgrades, and security vulnerability patches) has a high time sensitivity, and its effectiveness decays rapidly over time. Existing systems cannot identify and handle this decay.

[0004] In existing technologies, CN120950700A discloses a method for periodically traversing the validity periods of different knowledge data in an enterprise knowledge base and determining the timeliness verification of expired knowledge data. However, this method adopts a binary judgment (valid / expired) model, similar to food shelf-life management, which cannot quantify the gradual change process of knowledge credibility; it adopts uniform validity period management for all knowledge, lacking differentiated decay based on knowledge type; expired knowledge is directly pruned and deleted from the knowledge graph, which is irreversible, and there is no verification and refresh mechanism. CN120030067A discloses an evaluation module to score the knowledge graph, but it is only an overall quality score, not a confidence quantification at the knowledge item level, and the "cleaning of expired content" relies on periodic manual operation without mathematical model support.

[0005] Defect 2: The problem of knowledge silos is severe, with a lack of structured connections between knowledge items. Existing knowledge management tools allow users to create independent knowledge items, but they do not enforce constraints on the relationships between items. A large number of knowledge items exist in isolation, lacking semantic links and logical dependencies. As the knowledge base grows, the silo problem worsens exponentially: users cannot find similar existing knowledge and create duplicates, leading to information redundancy; contradictory relationships between knowledge are not labeled, resulting in multiple conflicting versions of the same topic; and the pre-dependency chains of knowledge are broken, making it impossible for users to trace a piece of knowledge back to its theoretical basis or related concepts. The tagging system and manual linking in existing solutions rely on user self-maintenance, which becomes ineffective in large-scale scenarios.

[0006] In existing technologies, CN120030067A establishes relationships between knowledge graph nodes through multi-ontology mapping and entity alignment. However, this relationship is an entity-level triple relationship, without type constraints, mandatory verification before writing, or a replacement double-chain mechanism. CN120872995A uses a three-level hierarchical structure and a directed index graph to organize the knowledge structure, but its directed edges are retrieval path direction graphs, not semantic relationships between knowledge items, and lack anti-island enforcement constraints.

[0007] Defect 3: Knowledge updates are primarily append-based, lacking an incremental fusion mechanism, leading to knowledge base expansion and accuracy degradation. Current solutions primarily use an append-based knowledge update model: when new knowledge is generated, the system creates new entries without merging them with existing knowledge. This model results in a continuously expanding knowledge base and a continuous degradation in knowledge accuracy: knowledge on the same topic is scattered across multiple entries, each containing only partial information and potentially contradictory; users need to read numerous redundant entries to piece together complete understanding, resulting in extremely low cognitive efficiency. Current solutions lack a mandatory incremental fusion mechanism—a binding rule that requires retrieving and merging existing knowledge before writing new knowledge.

[0008] In the prior art, CN120872995A discloses a method for detecting index conflicts based on a Bayesian simulation question-answering algorithm and outputting a resolution strategy through an adversarial model. However, this method addresses the index compatibility problem of ambiguity in index keywords and contradictions in retrieval paths, rather than the consistency analysis of assertions at the knowledge content level; its conflict resolution method is index standardization, the fusion or replacement of non-knowledge content; and it lacks the knowledge accuracy increment effect, with its feedback iterative optimization objective being the conflict reduction rate and the improvement of non-knowledge accuracy.

[0009] Defect 4: Lack of context compression mechanism leads to uncontrolled computing power and costs due to large-scale knowledge bases. When the knowledge base reaches hundreds or even thousands of entries, the cost of presenting all knowledge content to downstream agents or users increases dramatically. Existing solutions lack effective context compression mechanisms. A common practice is to directly send the original content of all relevant files into the context window of the language model, causing token consumption to increase linearly or even superlinearly with the size of the knowledge base, resulting in uncontrolled inference costs. The effective context window of the language model is occupied by a large amount of low-information-density content, key information is buried, and the quality of the response decreases.

[0010] In existing technologies, CN120030067A's data visualization presentation transforms data into charts for human users, but it does not compress knowledge content for the LLM context window; CN120950700A's data feedback unit is data packaging and compression at the network transmission layer; CN120872995A's feature distillation is machine learning feature dimension compression (dimensionality reduction), but it is not knowledge document content compression. None of the three competing products have a context compression mechanism for token cost control.

[0011] Defect 5: Lack of immutability guarantee for original data, leading to a broken knowledge traceability chain. Existing knowledge management systems allow users to freely edit any document, including original references. When original data is modified, knowledge entries derived from it cannot be traced back to their original source, resulting in a broken knowledge traceability chain. In research and compliance scenarios, the immutability of original data is a rigid requirement—any modification to original experimental data, legal texts, or contract terms must be traceable and documented. Existing solutions lack cold backup and isolation mechanisms for original data, and also lack an automatic format parsing and conversion pipeline from binary formats (PDF, DOCX, XLSX, etc.) to operable text formats.

[0012] Defect 6: Knowledge archiving relies on manual triggering and lacks an automated closed loop. Existing knowledge management systems generally rely on manual triggering for archiving operations: users need to actively determine which materials need archiving, manually perform classification and organization, and manually establish knowledge associations. In multi-agent collaborative scenarios, the amount of knowledge generated daily far exceeds human processing capacity. Knowledge not archived in a timely manner is scattered throughout the workspace and ultimately lost when the session ends. The lack of a scheduled, automatically triggered archiving mechanism makes the cycle from knowledge generation to storage uncontrollable, and the incremental effect of knowledge accuracy cannot be realized.

[0013] In existing technologies, CN120030067A's "periodic update" relies on manual triggering; CN120950700A uses real-time monitoring of data source changes to trigger graph updates, but only involves operations at the graph structure level (branch expansion / pruning), not a complete knowledge archiving pipeline. Both lack a fully automated archiving closed loop that is centrally executed by a dedicated knowledge maintenance agent and includes incremental fusion, context compression, immutable cold backup, metadata updates, and dual-chain maintenance. Summary of the Invention

[0014] This invention aims to address the aforementioned deficiencies in existing knowledge base maintenance technologies by providing an agent-driven automated knowledge base maintenance method and system. Through a six-in-one combination of knowledge confidence scoring and time decay mechanism, anti-island bidirectional linking, incremental fusion and knowledge accuracy enhancement, context high-concentration compression, immutable original cold backup, and timed triggering automatic archiving, it achieves automated knowledge base maintenance and continuous improvement of knowledge accuracy.

[0015] In this invention, "knowledge-generating intelligent agent" and "knowledge-maintaining intelligent agent" are both functional definitions used to refer to intelligent agents that perform corresponding functional steps, rather than limiting the name, identity, number, or implementation method of the intelligent agents: any intelligent agent capable of performing knowledge card generation and corresponding metadata assignment functions belongs to the knowledge-generating intelligent agent described in this invention; any intelligent agent capable of performing knowledge maintenance functions such as incremental fusion, context compression, immutable cold backup, and automatic archiving belongs to the knowledge-maintaining intelligent agent described in this invention. The knowledge generation and knowledge maintenance functions can be performed by two different intelligent agents respectively, or by the same intelligent agent at different operational stages, or by multiple intelligent agents collaboratively; this invention does not impose any restrictions on this. The "intelligent agent" described in this invention includes, but is not limited to, autonomous program entities driven by a large language model, scripted automated programs, service processes, or any combination thereof. Any naming of intelligent agents in specific embodiments is merely an exemplary label for ease of description and does not constitute a limitation on the scope of protection of this invention.

[0016] A first aspect of the present invention provides an agent-driven automated knowledge base maintenance method, comprising the following steps: S1. Knowledge generation steps: The agent generates knowledge cards, assigns each knowledge card a confidence score based on the authority of the knowledge source and a decay rate metadata based on the characteristics of the knowledge domain, and establishes strong-typed anti-island bidirectional links between knowledge cards. Each knowledge card must contain at least one strong-typed link, and cards that do not meet the constraints are rejected from being written into the knowledge base.

[0017] According to a first aspect of the invention, in step S1, the confidence score is assigned a differentiated initial value based on the authority of the knowledge source. The initial value ranges from 0.1 to 0.9, and increases with the increase of the authority of the knowledge source. The metadata of each knowledge card includes the current confidence level, initial confidence level, most recent verification date, decay rate level, knowledge status, substitution reference field, and number of verifications.

[0018] According to a first aspect of the present invention, in step S1, the confidence level of knowledge decreases over time according to an exponential decay model. The decay calculation includes the following sub-steps: determining the decay coefficient—determining the decay coefficient λ based on the decay rate level of the knowledge card, wherein the decay coefficient range for basic theoretical knowledge is 0.001~0.003 / day, the decay coefficient range for framework version knowledge is 0.003~0.008 / day, and the decay coefficient range for API interface and market data knowledge is 0.008~0.015 / day; calculating the current confidence level after decay—C_current = C_last_verified × e^(-λ× Δd), where C_last_verified is the confidence level at the time of the most recent verification, and Δd is the number of days since the most recent verification; Determine the knowledge status - when the current confidence level is lower than the active threshold, the knowledge status automatically changes from active to expired. When the current confidence level is lower than the elimination threshold and the number of verifications is the initial value, the knowledge card is marked as pending elimination. The active threshold ranges from 0.25 to 0.35, and the elimination threshold ranges from 0.08 to 0.12.

[0019] According to a first aspect of the present invention, in step S1, when knowledge is independently verified, the confidence level is increased according to the following formula: C_verified = C_current + α × (1 - C_current) × (1 - 1 / n), where C_current is the current confidence level before verification, α is the adjustment coefficient ranging from 0.2 to 0.4, and n is the cumulative number of independent verifications including this one; when n=1, the adjustment item is zero and no adjustment is performed; when n approaches infinity, the adjustment magnitude approaches α×(1-C_current); the confidence level after adjustment does not exceed a preset upper limit. Independent verification must simultaneously meet the following conditions: the verification source is different from the original source, the verifier is different from the original creator, and the time interval between the verification and the last verification is not less than 7 days.

[0020] According to a first aspect of the invention, in step S1, the anti-island bidirectional link includes four types of strong links: source links, used to mark the source of knowledge; comparison links, used to establish comparison relationships with related concepts; dependency links, used to establish dependency relationships with prior knowledge; and contradiction links, used to mark contradictions and evolution with other knowledge. Each knowledge card must contain at least one strong link, and the first knowledge card created must contain a source link. When old knowledge is replaced by new knowledge, the old card's status is changed to expired and a replacement reference field is set. The new card's text contains a replacement link. The replacement reference field and the replacement link constitute a physical bidirectional link, ensuring that it is possible to trace from the old card to the new card and from the new card back to the old card.

[0021] According to a first aspect of the present invention, based on a bidirectional link network between knowledge cards, each knowledge card is mapped to a graph node, and each bidirectional link is mapped to an undirected edge and weighted according to the link type, wherein dependent links have the largest weight, source links have the second largest weight, contrasting links have the third largest weight, and contradictory links have the smallest weight; the Louvain community detection algorithm is executed on the constructed knowledge graph to maximize the modularity, and the modularity calculation formula is Q = (1 / 2m) × Σ_ij [A_ij- (k_i × k_j) / (2m)] × δ(c_i, c_j), where m is the sum of the weights of all edges in the graph, A_ij is the edge weight between nodes i and j, k_i is the degree of node i, c_i is the community to which node i belongs, and δ is the Kronecker function; the algorithm terminates when the change in modularity is lower than a preset threshold or the maximum number of iterations is reached; each community output by the algorithm is a latent knowledge cluster, and the system automatically generates a cluster summary card for each knowledge cluster and stores it in the corresponding knowledge partition directory.

[0022] S2, Incremental Fusion Step: Before writing new knowledge into the knowledge base, perform deduplication retrieval in the target partition. When an existing knowledge card with the same or highly similar topic as the new knowledge is found, determine the logical consistency between the new knowledge and the existing knowledge card. For logically consistent knowledge, merge the content of the new knowledge into the existing knowledge card, and increase the confidence score of the existing knowledge card, update the verification date, and increase the number of verifications when the preset independent verification conditions are met. For logically conflicting knowledge, create a new knowledge card and mark the existing knowledge card as expired. At the same time, establish a physical bidirectional link between the new and old knowledge cards to replace each other.

[0023] According to a first aspect of the present invention, the incremental fusion process in step S2 includes the following sub-steps: receiving knowledge to be archived—the knowledge maintenance agent receives the knowledge content to be archived, including the knowledge text, source information, and metadata; determining the target partition—the target archive partition directory is determined according to the domain affiliation of the knowledge content; performing deduplication retrieval—keyword retrieval and semantic similarity retrieval are performed in the target partition directory, wherein the keyword similarity is calculated using the Jaccard coefficient, that is, the intersection size of the keyword set of the knowledge to be archived and the existing card keyword set divided by the union size, the keywords are extracted from the knowledge title and the first paragraph of the text using the TF-IDF algorithm, with a preset threshold range of 0.55~0.65, and the semantic similarity is calculated using the cosine similarity of the pre-trained sentence embedding vectors, with a preset threshold range of 0.70~0.80. When both reach the preset threshold, it is determined that the deduplication has been hit, and when the semantic similarity model is unavailable, it is downgraded to using only keyword similarity for deduplication determination.

[0024] According to a first aspect of the present invention, the incremental fusion process further includes a conflict analysis sub-step: when a duplicate is detected, the knowledge maintenance agent reads the complete content of the existing card that was detected, performs a logical consistency analysis between the knowledge to be archived and the existing content, and classifies the results into two categories: mergeable and fundamentally conflicting. Merging is defined as having consistent core arguments but different details or perspectives, while fundamentally conflicting is defined as having contradictory core arguments that cannot be reconciled through supplementary explanations. The logical consistency analysis is performed by the knowledge maintenance agent through reasoning using a large language model. The determination results record the determination basis and confidence level. For determination results with a confidence level lower than a preset threshold, manual confirmation is required and the authorized reviewer is notified.

[0025] According to a first aspect of the invention, when it is determined that the content can be merged, content fusion is performed—new information in the knowledge to be archived is merged into the existing card text, while retaining the original paragraph structure and all anti-island bidirectional links. The new content is inserted in a way that does not disrupt the original text logic; metadata dimensionality is upgraded—the confidence score is increased, the most recent verification date is updated, and the number of verifications is increased according to the verification refresh mechanism. If the source authority of the knowledge to be archived is higher than that of the source of the existing card record, the initial confidence is updated to the higher of the two; dual-link maintenance is performed—all anti-island bidirectional links of the existing card are retained. If the knowledge to be archived introduces new relationships, the corresponding links are added to the existing card text.

[0026] According to a first aspect of the present invention, when a fundamental conflict is determined, a new knowledge card is created and the complete content of the knowledge to be archived is written in it. The status of the old card is changed to expire and a replacement reference field is set to point to the new card. Replacement links are added to the text of the new card, and a back link is established in the old card through the replacement reference field to form a physical bidirectional link. If the conflict is an academic viewpoint or a technical approach, contradictory links are added to the new card to preserve the historical trajectory of knowledge evolution.

[0027] According to a first aspect of the present invention, through an incremental fusion mechanism, the accuracy of each piece of knowledge in the knowledge base increases with the number of independent verifications n, and the knowledge accuracy gain ratio KPGR = [C0 + α × (1 - C0) × (1 - 1 / n)] / C0, which increases monotonically with n and approaches the upper limit, so that each piece of knowledge in the knowledge base will only become more and more accurate and will not degrade due to the writing of new knowledge.

[0028] S3. Automatic Archiving Steps: The agent is automatically awakened by a timed trigger according to preset time rules, or by an external scheduling system. Upon receiving the archiving instruction, the awakened agent automatically reviews its work content and performs the incremental fusion step on the knowledge to be archived. When the number of files in the target partition exceeds a preset threshold, the file content of that partition is extracted to generate a highly condensed summary file with a volume constrained by a preset compression ratio. When a binary file is identified, it is parsed and converted into a text copy, and the conversion quality is verified. After successful verification, the binary file is moved to a read-only cold backup area, and a bidirectional link is established between the text copy and the original cold backup file. No manual intervention is required throughout the process, and an archiving report is generated upon completion. This step is limited to the specific archiving operation method. The triggering timing and scheduling logic of archiving can be managed by the upper-level scheduling system, such as the scheduling method defined in the applicant's application filed on the same day, "A Construction and Scheduling Method and System for a Multi-Agent Collaborative System."

[0029] According to a first aspect of the present invention, the automatic archiving step includes the following sub-steps: Scheduled task scheduler configuration—configuring scheduled task scheduling rules, setting the wake-up time and execution parameters of the knowledge maintenance agent, including trigger time, time zone, inspection scope, and activation status; Knowledge maintenance agent wake-up—after the trigger time is reached, the scheduled task scheduler automatically wakes up the knowledge maintenance agent and transmits the inspection scope, target partition, and execution mode parameters. The knowledge maintenance agent can also be woken up by an external scheduling system after the project has passed full acceptance; Work content inventory—the knowledge maintenance agent automatically scans each partition of the knowledge base and the work area directory, identifies newly added or modified files in the previous time period to generate a list to be archived, determines the archiving status of each newly added file, and simultaneously scans all active knowledge cards to calculate their current confidence level according to the time decay model, marking cards below the active threshold as expired and generating a list to be verified.

[0030] According to a first aspect of the invention, the automatic archiving step further includes the following sub-steps: performing an immutable cold backup step on the binary files in the cold backup queue; performing an incremental fusion step on the knowledge content in the archiving queue; updating metadata for all archived knowledge cards; checking the integrity of anti-island bidirectional links for all archived knowledge cards to ensure that each card contains at least one strongly typed link, and automatically supplementing source links or dependent links for cards found to be missing links. After archiving is completed, the number of files in each partition is checked, and for partitions with more than a preset threshold of files, a context compression process is performed to update the status summary file. Finally, an archiving report is generated, which includes the number of knowledge entries processed in this archiving, the number of incremental fusions, the number of cold backup files, the number of confidence level increases, the number of expired knowledge warnings, and the number of compressed product updates.

[0031] According to a first aspect of the present invention, in the automatic archiving step, when the number of files in the target partition exceeds a preset threshold (the threshold ranges from 8 to 15), core documents are read one by one and a highly condensed state summary file is generated. The state summary file contains one or more of the following: a physical asset list, a summary of core decisions and conclusions, knowledge confidence statistics, and downstream action guidelines. The volume of the compressed product does not exceed a preset proportion of the total number of original files, the proportion ranging from 12% to 18%, and the compression ratio is calculated based on the number of bytes. When reading core documents, they are executed according to priority—architecture specification files take precedence over core code interface definitions, core code interface definitions take precedence over knowledge card files, and knowledge card files take precedence over auxiliary files. After compression, when a downstream agent accesses the directory, it is preferentially guided to read the state summary file as a global entry point.

[0032] According to a first aspect of the present invention, in the automatic archiving step, when a binary format original data file is identified, the binary file format is parsed and converted into a text format. The conversion types include PDF to text retaining heading levels and paragraph structure, DOCX to text retaining headings, body text, and lists, and XLSX to text converting worksheets into table format. After verifying that the integrity of the core content is retained reaches a preset threshold, the verification is passed. The preset threshold ranges from 75% to 85%. After the verification is passed, the original binary file is moved to a read-only cold backup storage area. The file system permissions of the cold backup area are set to read-only, and the permissions are temporarily elevated only when the file is moved in. After the operation is completed, the read-only state is restored immediately. A bidirectional link is established between the text copy and the original cold backup file. All subsequent knowledge operations are performed on the text copy, and the original binary file remains immutable.

[0033] According to a first aspect of the present invention, the knowledge generation step is performed by a knowledge generation agent, and the incremental fusion step and the automatic archiving step are performed by a knowledge maintenance agent; the knowledge generation agent and the knowledge maintenance agent are execution roles divided by function, and the two are implemented by different agents, or by the same agent at different operating stages.

[0034] A second aspect of the present invention provides an agent-driven automated knowledge base maintenance system, comprising: a knowledge generation agent module for generating knowledge cards, assigning each knowledge card a confidence score differentiated according to the authority of the knowledge source and a decay rate set according to the characteristics of the knowledge domain, and establishing strong typed bidirectional links; a knowledge maintenance agent module for performing deduplication and incremental fusion before writing new knowledge, performing context compression on partitions with more than a threshold number of files to generate state summary files, performing format parsing and conversion on binary files and moving them into a read-only cold backup area, and performing automatic archiving in response to a timed trigger or wake-up from an external scheduling system; a knowledge base storage layer including a knowledge partition directory, a binary file cold backup area, and state summary files; a routing index layer including a bidirectional link network between knowledge cards, supporting retrieval and associated traversal; and a timed trigger for automatically waking up the knowledge maintenance agent module according to a preset time rule.

[0035] According to a second aspect of the present invention, the knowledge cards generated by the knowledge generation intelligent agent module are transmitted to the knowledge maintenance intelligent agent module for incremental fusion and archiving. After archiving, the knowledge maintenance intelligent agent module generates an archiving report. The knowledge generation intelligent agent module and the knowledge maintenance intelligent agent module are execution roles divided by function, and they are implemented by different intelligent agents or by different functional modules of the same intelligent agent. The cold backup area in the knowledge base storage layer is set with read-only file system permissions. The files in the cold backup area are the only source of facts for the system. Any modification to the original data is written in the form of a new knowledge card and the evolution relationship is marked by replacing the bidirectional links. The knowledge maintenance intelligent agent module can also be woken up by an external scheduling system.

[0036] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the agent-driven automated knowledge base maintenance method described in the first aspect of the present invention.

[0037] In a fourth aspect, the present invention provides a computer program product comprising a computer program / instructions which, when executed by a processor, implement the agent-driven automated knowledge base maintenance method described in the first aspect of the present invention.

[0038] Compared with the prior art, the present invention has the following beneficial effects: Effect 1: Automated management of knowledge timeliness, eliminating misleading information from expired knowledge. Through confidence scoring and a time decay mechanism, the system can automatically identify and mark expired knowledge. The identification of expired knowledge is shortened from manual inspection to automatic system marking, reducing the identification delay to zero days. API knowledge based on a rapid decay rate smoothly decays from its initial value to a warning level after 30 days, triggering an automatic system warning. Basic theoretical knowledge based on a slow decay rate maintains an active confidence level after 365 days, consistent with actual cognitive patterns. Compared with the binary validity period determination of existing technologies, this invention achieves a smooth decrease in knowledge credibility through continuous probability decay and supports reactivating knowledge about to expire by independently verifying and refreshing the confidence level.

[0039] Effect 2: The knowledge silo rate is significantly reduced, and the knowledge relevance is greatly improved. Through the mandatory constraint of anti-silo bidirectional links, the knowledge silo rate drops to zero under mandatory constraints, and the average number of outgoing links per knowledge card increases from less than one in the existing system to more than two. The Louvain community discovery algorithm can automatically identify implicit knowledge clusters that have not been manually labeled. In particular, the "contradictory" link type explicitly models knowledge conflict relationships as physical links for the first time, making the history of knowledge evolution traceable.

[0040] Effect 3: The knowledge precision increasing effect, where knowledge precision monotonically increases with usage. Through the incremental fusion mechanism, the number of duplicate entries on the same topic is reduced from multiple to one, and the knowledge base volume growth rate decreases from linear growth in a pure addition mode to sublinear growth. The knowledge precision gain ratio, after multiple independent verifications, monotonically increases with the number of verifications, approaching its upper limit, ensuring that each piece of knowledge in the knowledge base only becomes more accurate and will not degrade with the addition of new knowledge.

[0041] Effect 4: Token consumption is significantly reduced, and inference costs are controllable. Through a context-rich compression mechanism, when a knowledge partition contains a large number of files, the compressed state summary file size is approximately 10% to 15% of the original file size. This significantly reduces the token consumption of downstream agents, while the compressed output retains high-information-density content such as core architectural decisions, component health, and action guidelines. The knowledge acquisition cost for downstream agents is reduced from being proportional to the number of files to a constant level.

[0042] Effect 5: Complete and reliable source traceability chain for original data. Through the immutable original cold backup protocol, the modification rate of original data is zero—an inevitable result of the system design of physical isolation and read-only constraints. Any knowledge card can be traced back to its original data via bidirectional links, supporting automatic format parsing and conversion of mainstream binary formats such as PDF, DOCX, and XLSX, with conversion completeness reaching a preset threshold. This protocol is the first in a knowledge management system to achieve the rigid constraint of immutable original data.

[0043] Effect Six: Fully automated knowledge archiving with zero human intervention. Through an automatic archiving mechanism driven by a scheduled task, the delay from knowledge generation to storage is shortened from an average of several days for manual triggering to no more than one archiving cycle. Archiving operations require no human intervention, and archiving reports are automatically generated and sent to the central scheduling node for archiving, allowing the incremental effect of knowledge accuracy to continue to operate. Figure 1 It is a complete closed-loop overview of the database knowledge lifecycle. Figure 2 This is a diagram illustrating the decay of knowledge confidence. Figure 3 This is a diagram of the knowledge activation verification and confidence refresh model; Figure 4 This is a flowchart of knowledge increment fusion and plagiarism detection.

Claims

1. An agent-driven automated knowledge base maintenance method, characterized in that, Includes the following steps: S1. Knowledge generation steps: The agent generates knowledge cards, assigns each knowledge card a confidence score based on the authority of the knowledge source and a decay rate metadata based on the characteristics of the knowledge domain, and establishes strong-typed anti-island bidirectional links between knowledge cards. Each knowledge card must contain at least one strong-typed link, and cards that do not meet the constraints are rejected from being written into the knowledge base. S2, Incremental Fusion Step: Before writing new knowledge into the knowledge base, perform deduplication retrieval in the target partition. When an existing knowledge card with the same or highly similar topic as the new knowledge is found, determine the logical consistency between the new knowledge and the existing knowledge card. For logically consistent knowledge, merge the content of the new knowledge into the existing knowledge card, and when the preset independent verification conditions are met, increase the confidence score of the existing knowledge card, update the verification date, and increase the number of verifications. For logically conflicting knowledge, create new knowledge cards and mark the existing knowledge cards as expired, while establishing a physical two-way link between the new and old knowledge cards to replace each other; S3. Automatic archiving step: The intelligent agent is automatically woken up by a timer trigger according to a preset time rule, or by an external scheduling system. After receiving the archiving instruction, the woken-up intelligent agent automatically takes stock of the work content and performs the incremental fusion step on the knowledge to be archived obtained from the stock. When the number of files in the target partition exceeds a preset threshold, the file content of the partition is extracted to generate a highly condensed state summary file whose volume is constrained by a preset compression ratio. When a binary file is identified, the binary file format is parsed and converted into a text copy and the conversion quality is verified. After the verification is passed, the binary file is moved into the read-only cold backup area, and a bidirectional link is established between the text copy and the original cold backup file. No manual intervention is required throughout the process, and an archive report is generated upon completion of archiving.

2. The method according to claim 1, characterized in that, In step S1, the confidence score is assigned a differentiated initial value based on the authority of the knowledge source. The initial value ranges from 0.1 to 0.9, and increases with the increase of the authority of the knowledge source.

3. The method according to claim 1, characterized in that, In step S1, the confidence level of knowledge decreases exponentially over time according to the following model: C_current = C_last_verified × e^(-λ × Δd), where C_last_verified is the confidence level at the time of the most recent verification, λ is the decay coefficient, and Δd is the number of days since the most recent verification. The decay coefficient is set to three levels according to the knowledge type, with the decay coefficient for basic theoretical knowledge ranging from 0.001 to 0.003 / day, the decay coefficient for framework version knowledge ranging from 0.003 to 0.008 / day, and the decay coefficient for API interface and market data knowledge ranging from 0.008 to 0.015 / day.

4. The method according to claim 1, characterized in that, In step S1, when knowledge is independently verified, the confidence level is increased according to the following formula: C_verified = C_current + α × (1 - C_current) × (1 - 1 / n) where C_current is the current confidence level before verification, α is the adjustment coefficient, which ranges from 0.2 to 0.4, and n is the cumulative number of independent verifications; The increased confidence level does not exceed the preset upper limit.

5. The method according to claim 1, characterized in that, In step S1, the anti-island bidirectional linking includes four types of strong links: source link, contrast link, dependency link, and contradiction link; each knowledge card must contain at least one strong link, and the first knowledge card created must contain a source link; when old knowledge is replaced by new knowledge, the old card status is changed to expired and a replacement reference field is set, and the new card text contains a replacement link. The replacement reference field and the replacement link constitute a physical bidirectional link, ensuring that it is possible to trace from the old card to the new card and from the new card back to the old card.

6. The method according to claim 5, characterized in that, Based on the bidirectional link network between knowledge cards, each knowledge card is mapped to a graph node, and each bidirectional link is mapped to an undirected edge and weighted according to the link type. The Louvain community detection algorithm is executed on the constructed knowledge graph to maximize the modularity. Each community output by the algorithm is a latent knowledge cluster. The system automatically generates a cluster summary card for each knowledge cluster and stores it in the corresponding knowledge partition directory.

7. The method according to claim 1, characterized in that, In step S2, the incremental fusion plagiarism detection adopts a dual judgment standard of keyword similarity and semantic similarity. When both reach the preset threshold, it is judged as a plagiarism hit. Logical consistency analysis divides knowledge conflicts into two categories: mergeable and fundamental conflict. If the core arguments are consistent but the details are different, they are mergeable. If the core arguments are contradictory and irreconcilable, they are fundamental conflict.

8. The method according to claim 7, characterized in that, The keyword similarity is calculated using the Jaccard coefficient, which is the intersection of the set of keywords to be archived and the set of keywords in the existing cards divided by the union. Keywords are extracted from the knowledge title and the first paragraph of the main text using the TF-IDF algorithm, with a preset threshold range of 0.55 to 0.

65. The semantic similarity is calculated using the cosine similarity of the pre-trained sentence embedding vectors, with a preset threshold range of 0.70 to 0.

80. When the semantic similarity model is unavailable, the system is downgraded to using only keyword similarity for plagiarism detection.

9. The method according to claim 1, characterized in that, In the automatic archiving step, when the number of files in the target partition exceeds a preset threshold (the threshold ranges from 8 to 15), the core documents are read one by one and a highly condensed status summary file is generated. The status summary file contains one or more of the following: physical asset list, core decision and conclusion summary, knowledge confidence statistics, and downstream action guidelines. The volume of the compressed product does not exceed a preset proportion of the total number of original files, for example, the proportion ranges from 12% to 18%, and the compression ratio is calculated based on the number of bytes.

10. The method according to claim 1, characterized in that, In the automatic archiving step, when a binary format original data file is identified, the binary file format is parsed and converted into text format. The core content integrity is verified to be within a preset threshold, and the verification is passed. The preset threshold ranges from 75% to 85%. After the verification is passed, the original binary file is moved to a read-only cold backup storage area. The file system permissions of the cold backup area are set to read-only, and permissions are temporarily elevated only when the file is moved in. A bidirectional link is established between the text copy and the original cold backup file. All subsequent knowledge operations are performed on the text copy, and the original binary file remains immutable.

11. The method according to claim 1, characterized in that, In step S2, when it is determined that the knowledge cards can be merged, the fusion operation retains all anti-island bidirectional links of the original knowledge cards, inserts new information in a way that does not destroy the original structure, and increases the confidence score of the existing knowledge cards, updates the most recent verification date and increases the number of verifications when the preset independent verification conditions are met. When a fundamental conflict is identified, if the conflict pertains to opposing academic viewpoints or technical approaches, a contradictory link is added to the new card to preserve the history of knowledge evolution.

12. The method according to claim 1, characterized in that, The knowledge generation step is executed by the knowledge generation agent, and the incremental fusion step and the automatic archiving step are executed by the knowledge maintenance agent. The knowledge generation agent and the knowledge maintenance agent are execution roles divided according to function. They are implemented by different agents or by the same agent at different operating stages.

13. An agent-driven automated knowledge base maintenance system, characterized in that, include: The knowledge generation intelligent agent module generates knowledge cards, assigns each knowledge card a confidence score differentiated based on the authority of the knowledge source, and sets a decay rate based on the characteristics of the knowledge domain, and establishes strong typed bidirectional links. The knowledge maintenance intelligent agent module performs deduplication and incremental fusion before writing new knowledge, performs context compression on partitions with more than a threshold number of files to generate state summary files, performs format parsing and conversion on binary files before moving them to a read-only cold backup area, and performs automatic archiving in response to timed triggers or external scheduling system wake-ups. The knowledge base storage layer includes a knowledge partition directory, a binary file cold backup area, and state summary files. The routing index layer includes a bidirectional link network between knowledge cards, supporting retrieval and associated traversal. A timed trigger automatically wakes up the knowledge maintenance intelligent agent module according to preset time rules. The knowledge cards generated by the knowledge generation intelligent agent module are passed to the knowledge maintenance intelligent agent module for incremental fusion and archiving. After archiving, the knowledge maintenance intelligent agent module generates an archiving report. The knowledge generation and knowledge maintenance intelligent agents are functionally divided execution roles, implemented by different agents or by different functional modules of the same agent.

14. The system according to claim 13, characterized in that, The cold backup area in the knowledge base storage layer is set with read-only file system permissions. The files in the cold backup area are the only source of facts for the system. Any modification to the original data is written in the form of a new knowledge card and the evolution relationship is marked by replacing the bidirectional links. The timed trigger automatically wakes up the knowledge maintenance intelligent agent module to execute the automatic archiving process according to the preset time rules. The knowledge maintenance intelligent agent module can also be woken up by an external scheduling system.

15. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the agent-driven automated knowledge base maintenance method as described in any one of claims 1 to 12.

16. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the agent-driven automated knowledge base maintenance method as described in any one of claims 1 to 12.

Images