Multi-agent collaborative decision-making system and method based on knowledge Token

Abstract

The invention discloses a multi-agent collaborative decision-making system and method based on knowledge Token, and the method comprises the following steps: S1, constructing a knowledge base architecture, executing the content Hash calculation, and generating an original fingerprint in an original fidelity layer; s2, extracting content, packaging the content into knowledge Token, and writing identification, semantics, source, authority, value, contract, association, traceability and health information; s3, activating a knowledge layer, constructing a vector index and mapping knowledge graph nodes; s4, unifying an index layer, matching task requirements, and screening and verifying candidate Tokens to obtain an authorization set; s5, the multiple agents generate local decisions and fuse values and semantics to form a collaborative result; and S6, binding a Token identifier after consistency verification, generating a collaborative decision and recording data consanguinity information. According to the method, multi-agent collaborative decision-making is realized through the knowledge Token and the knowledge graph, and the decision-making accuracy and traceability are improved.

Description

Technical Field

[0001] This invention relates to the field of multi-agent decision-making technology, and in particular to a multi-agent collaborative decision-making system and method based on knowledge tokens. Background Technology

[0002] Multi-agent collaborative decision-making technology has been widely applied in distributed control, intelligent scheduling, autonomous driving platooning, emergency command, and enterprise-level business process governance. Existing multi-agent collaborative frameworks primarily rely on shared state spaces, centralized training and decentralized execution mechanisms, message broadcasting mechanisms, or policy fusion mechanisms to achieve collaborative behavior. For example, some collaborative architectures achieve information flow between multiple agents by sharing environmental observations, exchanging local policy parameters, or transmitting task intentions; other methods utilize knowledge graphs, rule bases, or ontology structures to provide semantic background information to multiple agents to support reasoning and division of labor. However, existing technologies generally suffer from shortcomings in knowledge credibility, semantic uniformity, collaborative verifiability, and cross-agent permission boundary management.

[0003] Traditional multi-agent knowledge interaction methods rely on static knowledge bases or offline-built knowledge graphs, lacking verification of the authenticity of knowledge sources and effective management of knowledge granularity, timeliness, and traceability attributes. This leads to ambiguity or accumulated errors during knowledge transmission. Existing knowledge representation methods cannot provide fine-grained identification for each piece of knowledge, nor can they accurately describe the knowledge generation process, update process, and invocation relationships. This makes it difficult for multi-agents to trace knowledge sources and verify the decision-making derivation chain when making collaborative decisions. Furthermore, in multi-agent collaboration scenarios, different agents have different access permissions. Existing permission control mechanisms mostly adopt coarse-grained role-permission models, lacking fine-grained knowledge-level access control measures. This fails to ensure that every knowledge invocation meets the authorization level, thus posing potential security risks to the collaborative process.

[0004] Meanwhile, existing collaborative decision-making frameworks lack knowledge-oriented contractual constraint mechanisms. When multiple agents invoke external knowledge, they often lack parameter consistency verification mechanisms, leading to incompatibility and unreproducibility issues when knowledge content is transferred between different agents. Knowledge quality assessment mechanisms are often based on static scoring or manual review, lacking quantitative value labels and dynamic confidence measurements, and cannot reflect changes in knowledge reliability in real time. Furthermore, existing structured knowledge mostly remains at the semantic node and relation level, lacking dynamic monitoring of knowledge health status, lifecycle, usage frequency, and validity period, resulting in the erroneous invocation of outdated knowledge, thereby reducing the stability of collaborative decision-making results.

[0005] In practical applications, multi-agent collaboration often relies on rapid knowledge retrieval after a task is triggered. However, traditional vector retrieval and graph retrieval typically operate independently, lacking a unified indexing framework. This makes it impossible for knowledge selection to simultaneously consider semantic similarity and contextual relationships, resulting in biased decision-making references. Furthermore, the lack of verifiable legitimacy verification mechanisms during knowledge selection makes it difficult to ensure that the knowledge set ultimately used for multi-agent decision-making has permission compliance, parameter compatibility, and source credibility, making the decision-making process difficult to audit and reproduce.

[0006] Therefore, how to provide a multi-agent collaborative decision-making system and method based on knowledge tokens is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose a multi-agent collaborative decision-making system and method based on knowledge tokens. This invention utilizes content hashing at the original document fidelity layer, structured encapsulation at the activated knowledge layer, and vector retrieval and knowledge graph fusion mechanisms at the unified index layer to construct a traceable, verifiable, and controllable knowledge token system. It then drives multi-agent collaboration to generate reliable decision results through an authorized knowledge token set. This invention achieves a processing flow that ensures transparency of knowledge sources, refined knowledge retrieval, verifiable decision-making processes, and reliable cross-agent collaboration. It possesses advantages such as strong semantic consistency, high collaborative reliability, excellent decision traceability, and outstanding cross-scenario transferability.

[0008] A multi-agent collaborative decision-making method based on knowledge tokens according to an embodiment of the present invention includes the following steps:

[0009] S1. Construct a knowledge base architecture, acquire multi-source data and perform content hash calculation, anchor the hash result to an immutable storage in the original document fidelity layer, generate the original document fingerprint and record the original document fingerprint identifier;

[0010] S2. Based on the original document fingerprint, perform knowledge extraction, encapsulate the extracted structured content into a knowledge token, and write a unique identifier, semantic vector, source fingerprint, access control field, value tag and confidence level, calling contract, association relationship, data traceability information, health status and lifecycle information into it.

[0011] S3. Write the knowledge token into the activated knowledge layer, construct a vector index based on the semantic vector, and map the association relationship of the knowledge token to knowledge graph nodes to form a bidirectional link with the original fingerprint identifier.

[0012] S4. In the unified index layer, based on the vector index and knowledge graph, the task requirements are matched, the candidate knowledge token set is filtered, and the legality is verified based on the permission control field and the calling contract to obtain the authorized knowledge token set.

[0013] S5. The multi-agent system generates local decision results based on the authorized knowledge token set, and then integrates the local decision results with the value tags and semantic vectors of the knowledge tokens to form a collaborative fusion result.

[0014] S6. Perform consistency verification on the collaborative fusion results. After the verification is passed, bind the unique identifier of the knowledge token used, generate collaborative decision results and record data lineage information.

[0015] Optionally, the knowledge base architecture consists of an original document fidelity layer, an activated knowledge layer, and a unified index layer, specifically including:

[0016] The original document fidelity layer performs integrity calibration on multi-source data based on the content hashing algorithm and anchors the hash result to tamper-proof storage to generate the original document fingerprint.

[0017] The multi-source data consists of structured data collected from business databases, sensor terminals, log records, external knowledge bases, and document storage. After synchronous capture, format parsing, and time calibration are performed by the data acquisition engine, the data is aggregated to the original document fidelity layer to generate the original document fingerprint.

[0018] The activated knowledge layer performs structured parsing and semantic segmentation based on the original fingerprint identifier, and encapsulates the parsing result into a knowledge token with a unique identifier, semantic vector, source fingerprint, access control field, value tag and confidence level, calling contract, association relationship, data traceability information, health status and life cycle information;

[0019] The unified index layer establishes a vector index based on the semantic vector of the knowledge token and constructs knowledge graph nodes based on the association relationship, realizing bidirectional linking between the knowledge token and the original fingerprint and cross-layer traceable retrieval.

[0020] Optionally, the original document fidelity layer is used to solidify and reliably identify the content of multi-source data, specifically including:

[0021] The hash engine receives raw data from sensors, logs, business databases, and file storage, and performs content hash calculations to generate hash digests.

[0022] The hash digest is concatenated and encoded with the timestamp, data source identifier, and data type label to form the original document fingerprint;

[0023] The original document fingerprint is anchored to an immutable storage medium via a blockchain or distributed file system, and a corresponding storage address index is generated, which is the original document fingerprint identifier.

[0024] Write the original document fingerprint and original document fingerprint identifier into the authenticity index table for knowledge extraction and traceability operations;

[0025] Upon completion, a hash verification task is triggered to compare the latest hash digest with the original record to confirm that the original content has not been modified.

[0026] Optionally, the activated knowledge layer is used to realize the structured encapsulation and dynamic management of knowledge, specifically including:

[0027] Based on the original fingerprint reading of the corresponding original content, text parsing, entity recognition, attribute extraction and semantic segmentation operations are performed to generate structured knowledge units;

[0028] Structured knowledge units are encoded into knowledge tokens, and unique identifiers, semantic vectors, source fingerprints, access control fields, value tags and confidence levels, invocation contracts, relationships, data traceability information, health status and lifecycle information are written into the knowledge tokens.

[0029] Establish an association table based on the relationships between knowledge tokens, and cluster and store related knowledge tokens.

[0030] Perform a content signing operation on each knowledge token, and reverse-bind the signature hash with the original document fingerprint to ensure that the encapsulated content is traceable;

[0031] A token index list is generated at the activation knowledge layer for the unified index layer to read.

[0032] Optionally, the unified index layer is used to achieve efficient retrieval and related reasoning of knowledge tokens, specifically including:

[0033] Receive a list of knowledge tokens from the activated knowledge layer, extract the semantic vector of each knowledge token, and construct a vector index space using a vectorized retrieval algorithm;

[0034] Parse the association table information between knowledge tokens, construct a knowledge graph in the form of nodes and edges, and write semantic similarity, association relationship and calling contract into the edge weight field of the graph.

[0035] After index initialization, a bidirectional mapping between the Token index and graph nodes is executed, enabling knowledge tokens to be located through semantic vector retrieval and to track associated content through graph structure relationships.

[0036] When a task query request is received, the semantic similarity score is calculated in the vector index based on the task description, and an association expansion search is performed in the knowledge graph to filter out a set of knowledge tokens that meet the permission control and contract constraints, and the authorization result is returned for multi-agent calls.

[0037] Optionally, the unique identifier is used to identify the single identity of the knowledge token and associate it with the original document fingerprint record; the semantic vector is generated by embedding and encoding structured knowledge units to support vector retrieval; the source fingerprint is calculated by content hashing to trace the origin of the original data; the access control field defines the callable scope and operation level of the knowledge token according to the access policy; the value tag and confidence level are generated by a dynamic scoring mechanism to quantify the quality and reliability of the knowledge token; the calling contract constrains the usage conditions and interaction parameters of the knowledge token in the form of interface specifications; the association relationship is constructed based on semantic dependency mapping to connect related knowledge token nodes to form a knowledge graph; the data traceability information records the complete link of the knowledge token generation and calling process through a fidelity index table; and the health status and lifecycle information are generated by the system monitoring task to mark the validity and expiration status of the knowledge token.

[0038] Optionally, the authorized knowledge token set is generated by the unified index layer after the task request is triggered, based on a joint selection of semantic vectors and knowledge graphs, specifically including:

[0039] Read the task description and calculate the semantic similarity score with each knowledge token in the vector index space, and filter the candidate knowledge token set according to the score;

[0040] The access control field is invoked to verify the access level matching degree, and the interaction parameters are checked for compliance according to the invocation contract, filtering out knowledge tokens that do not meet the access policy.

[0041] For knowledge tokens that pass the legality verification, a value label and confidence weighted calculation is performed, and knowledge tokens with scores higher than the threshold are selected to form the target set;

[0042] The relationships between target sets are expanded into a graph, and semantic dependency nodes are added to form a complete contextual knowledge chain;

[0043] Generate a set of authorized knowledge tokens and assign them unique identifiers. Record the original document fingerprints, source hashes, and permission verification results involved, and use them for collaborative decision-making by multiple agents.

[0044] Optionally, S5 specifically includes:

[0045] S51. Analyze the semantic vector and value tag information in the authorized knowledge token set, combine the task objective to execute strategy reasoning and action selection, and output the preliminary decision result that conforms to the call contract constraint;

[0046] S52. Each intelligent agent binds the primary decision result with the corresponding knowledge token unique identifier to form a local decision result set;

[0047] S53. Perform feature normalization and temporal alignment operations on the local decision result set, calculate the similarity matrix based on the semantic vector of the knowledge token, and use the attention weighting mechanism to weight and aggregate the local results with high similarity to generate a comprehensive feature representation.

[0048] S54. The comprehensive feature representation is jointly fused with the value label, confidence level and correlation of the knowledge token to form a collaborative fusion result, and the fusion weight and the corresponding knowledge token mapping table are recorded to provide input data for consistency verification and collaborative decision output.

[0049] According to an embodiment of the present invention, a multi-agent collaborative decision-making system based on knowledge tokens includes:

[0050] The knowledge base construction module is used to build the knowledge base architecture, acquire multi-source data and perform content hash calculation, anchor the hash result to immutable storage in the original document fidelity layer, generate original document fingerprints and record original document fingerprint identifiers;

[0051] The knowledge encapsulation module is used to perform knowledge extraction based on the original document fingerprint identifier, encapsulate the extracted structured content into a knowledge token, and write a unique identifier, semantic vector, source fingerprint, access control field, value tag and confidence level, calling contract, association relationship, data traceability information, health status and lifecycle information.

[0052] The knowledge mapping module is used to write knowledge tokens into the activated knowledge layer, construct vector indexes based on semantic vectors, and map the association relationships of knowledge tokens to knowledge graph nodes, forming a bidirectional link with the original fingerprint identifier.

[0053] The knowledge filtering module is used to filter the candidate knowledge token set based on the vector index and knowledge graph matching task requirements in the unified index layer, and perform legality verification based on the permission control field and the calling contract to obtain the authorized knowledge token set.

[0054] The decision generation module is used to generate local decision results based on the authorized knowledge token set, and to collaboratively integrate the local decision results with the value tags and semantic vectors of the knowledge tokens to form a collaborative integration result;

[0055] The decision verification module is used to perform consistency verification on the collaborative fusion results. After the verification is passed, it binds the unique identifier of the knowledge token used, generates collaborative decision results, and records data lineage information.

[0056] The beneficial effects of this invention are:

[0057] First, by performing content hash calculations on multi-source data at the original fidelity layer and anchoring it to immutable storage, this invention ensures reliable consistency of the source fingerprint of the knowledge token, fundamentally improving the credibility of knowledge input and providing verifiable foundational support for multi-agent collaborative decision-making.

[0058] Secondly, in the activated knowledge layer, this invention generates knowledge tokens with unique identifiers, access control fields, calling contracts, value tags, confidence levels, and lifecycle information through structured parsing, semantic vector encoding, and relational mapping. In the unified index layer, a vector index and knowledge graph fusion retrieval are constructed, enabling multiple agents to obtain a precise, controllable, and traceable set of authorized knowledge tokens in a unified semantic space, thereby achieving high consistency collaboration and high-quality inference in the decision-making stage.

[0059] Finally, this invention drives multi-agents to generate local decision results and perform collaborative fusion of semantic vectors and value tags by authorizing a set of knowledge tokens. This makes the decision-making process verifiable, interpretable, and trustworthy across agents, effectively overcoming the problems of opaque knowledge, inconsistent semantics, and crude authorization mechanisms in traditional multi-agent collaboration. As a result, it significantly improves the stability, security, and feasibility of collaborative decision-making. Attached Figure Description

[0060] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0061] Figure 1 This is a flowchart of a multi-agent collaborative decision-making method based on knowledge tokens proposed in this invention;

[0062] Figure 2 This is a flowchart of the activation of the knowledge layer and the construction of the knowledge graph in a multi-agent collaborative decision-making method based on knowledge tokens proposed in this invention.

[0063] Figure 3 This is a block diagram of a multi-agent collaborative decision-making system based on knowledge tokens proposed in this invention. Detailed Implementation

[0064] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0065] refer to Figure 1-2 A multi-agent collaborative decision-making method based on knowledge tokens includes the following steps:

[0066] S1. Construct a knowledge base architecture, acquire multi-source data and perform content hash calculation, anchor the hash result to an immutable storage in the original document fidelity layer, generate the original document fingerprint and record the original document fingerprint identifier;

[0067] S2. Based on the original document fingerprint, perform knowledge extraction, encapsulate the extracted structured content into a knowledge token, and write a unique identifier, semantic vector, source fingerprint, access control field, value tag and confidence level, calling contract, association relationship, data traceability information, health status and lifecycle information into it.

[0068] S3. Write the knowledge token into the activated knowledge layer, construct a vector index based on the semantic vector, and map the association relationship of the knowledge token to knowledge graph nodes to form a bidirectional link with the original fingerprint identifier.

[0069] S4. In the unified index layer, based on the vector index and knowledge graph, the task requirements are matched, the candidate knowledge token set is filtered, and the legality is verified based on the permission control field and the calling contract to obtain the authorized knowledge token set.

[0070] S5. The multi-agent system generates local decision results based on the authorized knowledge token set, and then integrates the local decision results with the value tags and semantic vectors of the knowledge tokens to form a collaborative fusion result.

[0071] S6. Perform consistency verification on the collaborative fusion results. After the verification is passed, bind the unique identifier of the knowledge token used, generate collaborative decision results and record data lineage information.

[0072] In this embodiment, the knowledge base architecture consists of an original document fidelity layer, an activated knowledge layer, and a unified index layer, specifically including:

[0073] The original document fidelity layer performs integrity calibration on multi-source data based on the content hashing algorithm and anchors the hash result to tamper-proof storage to generate the original document fingerprint.

[0074] The multi-source data consists of structured data collected from business databases, sensor terminals, log records, external knowledge bases, and document storage. After synchronous capture, format parsing, and time calibration are performed by the data acquisition engine, the data is aggregated to the original document fidelity layer to generate the original document fingerprint.

[0075] The activated knowledge layer performs structured parsing and semantic segmentation based on the original fingerprint identifier, and encapsulates the parsing result into a knowledge token with a unique identifier, semantic vector, source fingerprint, access control field, value tag and confidence level, calling contract, association relationship, data traceability information, health status and life cycle information;

[0076] The unified index layer establishes a vector index based on the semantic vector of the knowledge token and constructs knowledge graph nodes based on the association relationship, realizing bidirectional linking between the knowledge token and the original fingerprint and cross-layer traceable retrieval.

[0077] In this embodiment, the original document fidelity layer is used to realize the content solidification and trust identification of multi-source data, specifically including:

[0078] The hash engine receives raw data from sensors, logs, business databases, and file storage, and performs content hash calculations to generate hash digests.

[0079] The hash digest is concatenated and encoded with the timestamp, data source identifier, and data type label to form the original document fingerprint;

[0080] The original document fingerprint is anchored to an immutable storage medium via a blockchain or distributed file system, and a corresponding storage address index is generated, which is the original document fingerprint identifier.

[0081] Write the original document fingerprint and original document fingerprint identifier into the authenticity index table for knowledge extraction and traceability operations;

[0082] Upon completion, a hash verification task is triggered to compare the latest hash digest with the original record to confirm that the original content has not been modified.

[0083] In this embodiment, the activated knowledge layer is used to realize the structured encapsulation and dynamic management of knowledge, specifically including:

[0084] Based on the original fingerprint reading of the corresponding original content, text parsing, entity recognition, attribute extraction and semantic segmentation operations are performed to generate structured knowledge units;

[0085] Structured knowledge units are encoded into knowledge tokens, and unique identifiers, semantic vectors, source fingerprints, access control fields, value tags and confidence levels, invocation contracts, relationships, data traceability information, health status and lifecycle information are written into the knowledge tokens.

[0086] Establish an association table based on the relationships between knowledge tokens, and cluster and store related knowledge tokens.

[0087] Perform a content signing operation on each knowledge token, and reverse-bind the signature hash with the original document fingerprint to ensure that the encapsulated content is traceable;

[0088] A token index list is generated at the activation knowledge layer for the unified index layer to read.

[0089] In this embodiment, the corresponding original fingerprint identifier and storage address index are read from the original fidelity layer, the parsing engine is called to perform hierarchical parsing on the original data content, the named entity recognition method is used to extract key entities and attribute relationships, and the dependency parsing model is used to identify semantic dependency structures.

[0090] The entity and semantic dependencies are vectorized by the feature encoder to generate a fixed-dimensional semantic vector representation;

[0091] The parsing engine aggregates content fragments based on semantic similarity to form structured knowledge units;

[0092] During the encapsulation phase, structured knowledge units are written with a unique identifier, source fingerprint, generation timestamp, data type label, and access control field.

[0093] Simultaneously calculate the value tags and confidence scores of structured knowledge units and write them into the lifecycle start and expiration times;

[0094] The generated semantic vector is then normalized and embedded into the semantic vector field of the knowledge token.

[0095] The signature algorithm is called to calculate the content hash of the knowledge token and bind it bidirectionally with the original fingerprint. The encapsulated knowledge token is then stored in the activated knowledge layer database, and an index record is generated for use by the unified index layer.

[0096] In this embodiment, the unified index layer is used to achieve efficient retrieval and related reasoning of knowledge tokens, specifically including:

[0097] Receive a list of knowledge tokens from the activated knowledge layer, extract the semantic vector of each knowledge token, and construct a vector index space using a vectorized retrieval algorithm;

[0098] Parse the association table information between knowledge tokens, construct a knowledge graph in the form of nodes and edges, and write semantic similarity, association relationship and calling contract into the edge weight field of the graph.

[0099] After index initialization, a bidirectional mapping between the Token index and graph nodes is executed, enabling knowledge tokens to be located through semantic vector retrieval and to track associated content through graph structure relationships.

[0100] When a task query request is received, the semantic similarity score is calculated in the vector index based on the task description, and an association expansion search is performed in the knowledge graph to filter out a set of knowledge tokens that meet the permission control and contract constraints, and the authorization result is returned for multi-agent calls.

[0101] In this embodiment, the unique identifier is used to identify the single identity of the knowledge token and associate it with the original document fingerprint record; the semantic vector is generated by embedding and encoding structured knowledge units to support vector retrieval; the source fingerprint is calculated by content hashing to trace the origin of the original data; the access control field defines the callable scope and operation level of the knowledge token according to the access policy; the value tag and confidence level are generated by a dynamic scoring mechanism to quantify the quality and reliability of the knowledge token; the calling contract constrains the usage conditions and interaction parameters of the knowledge token in the form of interface specifications; the association relationship is constructed based on semantic dependency mapping to connect related knowledge token nodes to form a knowledge graph; the data traceability information records the complete link of the knowledge token generation and calling process through a fidelity index table; and the health status and lifecycle information are generated by the system monitoring task to mark the validity and expiration status of the knowledge token.

[0102] In this embodiment, the authorized knowledge token set is generated by the unified index layer based on a joint selection of semantic vectors and knowledge graphs after a task request is triggered, specifically including:

[0103] Read the task description and calculate the semantic similarity score with each knowledge token in the vector index space, and filter the candidate knowledge token set according to the score;

[0104] The access control field is invoked to verify the access level matching degree, and the interaction parameters are checked for compliance according to the invocation contract, filtering out knowledge tokens that do not meet the access policy.

[0105] For knowledge tokens that pass the legality verification, a value label and confidence weighted calculation is performed, and knowledge tokens with scores higher than the threshold are selected to form the target set;

[0106] The relationships between target sets are expanded into a graph, and semantic dependency nodes are added to form a complete contextual knowledge chain;

[0107] Generate a set of authorized knowledge tokens and assign them unique identifiers. Record the original document fingerprints, source hashes, and permission verification results involved, and use them for collaborative decision-making by multiple agents.

[0108] In this embodiment, during the generation of the authorized knowledge token set, the calculation of semantic similarity scores is performed through the vector retrieval engine of the unified index layer, specifically including:

[0109] The task description text is converted into a fixed-dimensional vector representation and normalized. Then, all semantic vectors registered in the vector index space are called to perform cosine similarity calculation to generate a similarity matrix.

[0110] The vector retrieval engine performs Top-K sorting based on the matrix results and then performs a weighted sorting based on distance decay on the sorted results to obtain a candidate knowledge token sequence.

[0111] The permission control field of each knowledge token in the candidate knowledge token sequence is compared with the access level table to generate a permission matching boolean mask, and the boolean mask is used to filter out knowledge tokens that do not meet the access conditions.

[0112] For the retained knowledge token, the contract verifier is invoked to parse the parameter specifications, field structure and constraint rules in the invocation contract, and the consistency between the parameter set generated by the task description vector and the parameter set of the invocation contract is verified by the structure matching algorithm.

[0113] All knowledge tokens that pass the contract verification will be input into the value evaluation unit, which will construct a multi-dimensional feature vector using value tags, confidence levels, timestamps, and usage frequency, and calculate a comprehensive score using a linear weighted scoring function.

[0114] Knowledge tokens with scores higher than a preset threshold will be organized into a target list. Then, the graph interface will be used to query the relationships between each knowledge token in the list, retrieve relevant nodes, and perform a one-time semantic closure expansion, appending the directly dependent nodes to the end of the list.

[0115] The expanded list is written to the authorization set cache, an authorization knowledge token set is generated, and the corresponding session identifier, original fingerprint path, permission verification result and contract verification record are recorded.

[0116] In this embodiment, S5 specifically includes:

[0117] S51. Analyze the semantic vector and value tag information in the authorized knowledge token set, combine the task objective to execute strategy reasoning and action selection, and output the preliminary decision result that conforms to the call contract constraint;

[0118] S52. Each intelligent agent binds the primary decision result with the corresponding knowledge token unique identifier to form a local decision result set;

[0119] S53. Perform feature normalization and temporal alignment operations on the local decision result set, calculate the similarity matrix based on the semantic vector of the knowledge token, and use the attention weighting mechanism to weight and aggregate the local results with high similarity to generate a comprehensive feature representation.

[0120] S54. The comprehensive feature representation is jointly fused with the value label, confidence level and correlation of the knowledge token to form a collaborative fusion result, and the fusion weight and the corresponding knowledge token mapping table are recorded to provide input data for consistency verification and collaborative decision output.

[0121] refer to Figure 3 A multi-agent collaborative decision-making system based on knowledge tokens, comprising:

[0122] The knowledge base construction module is used to build the knowledge base architecture, acquire multi-source data and perform content hash calculation, anchor the hash result to immutable storage in the original document fidelity layer, generate original document fingerprints and record original document fingerprint identifiers;

[0123] The knowledge encapsulation module is used to perform knowledge extraction based on the original document fingerprint identifier, encapsulate the extracted structured content into a knowledge token, and write a unique identifier, semantic vector, source fingerprint, access control field, value tag and confidence level, calling contract, association relationship, data traceability information, health status and lifecycle information.

[0124] The knowledge mapping module is used to write knowledge tokens into the activated knowledge layer, construct vector indexes based on semantic vectors, and map the association relationships of knowledge tokens to knowledge graph nodes, forming a bidirectional link with the original fingerprint identifier.

[0125] The knowledge filtering module is used to filter the candidate knowledge token set based on the vector index and knowledge graph matching task requirements in the unified index layer, and perform legality verification based on the permission control field and the calling contract to obtain the authorized knowledge token set.

[0126] The decision generation module is used to generate local decision results based on the authorized knowledge token set, and to collaboratively integrate the local decision results with the value tags and semantic vectors of the knowledge tokens to form a collaborative integration result;

[0127] The decision verification module is used to perform consistency verification on the collaborative fusion results. After the verification is passed, it binds the unique identifier of the knowledge token used, generates collaborative decision results, and records data lineage information.

[0128] Example 1:

[0129] To verify the feasibility of this invention in practice, it was applied to a complex multi-source business collaboration environment. This environment involves multiple types of intelligent agents collaborating, with diverse data sources and tasks requiring real-time performance and high reliability. Previous methods for sharing knowledge among multiple agents suffer from problems such as opaque knowledge sources, inconsistent semantic interpretations, and difficulty in strictly controlling access permissions, leading to decision delays, collaboration conflicts, and accumulated inference biases, significantly impacting overall collaboration efficiency. In this environment, this invention uses knowledge tokens to reliably encapsulate knowledge and leverages a component fidelity layer, an activated knowledge layer, and a unified index layer working together to ensure the flow of knowledge among multiple agents is verifiable, traceable, and has controllable access boundaries, thereby addressing several core pain points of traditional technologies.

[0130] In this application environment, multi-source data is first sent to the original document fidelity layer, where content hashing is performed and the data is written to immutable storage to generate original document fingerprints. Taking one data stream as an example, this data source generates approximately 180,000 structured and semi-structured records daily on average. After hashing, the data fingerprint consistency remained at 100% for 30 consecutive days, with no fingerprint anomalies observed. These original document fingerprints are then read by the activated knowledge layer, where semantic extraction, entity recognition, and attribute encapsulation are performed to generate knowledge tokens with unique identifiers, semantic vectors, source fingerprints, access control fields, value tags, confidence levels, invocation contracts, relationships, data traceability information, and lifecycle information. In continuous operation testing, the activated knowledge layer generates an average of approximately 54,000 knowledge tokens daily, with the semantic vector redundancy rate remaining below 3%, indicating that the knowledge encapsulation maintains high independence and information concentration.

[0131] In practical applications, multiple agents submit decision-making requests for collaborative tasks. After receiving these requests, the unified index layer performs semantic similarity calculations based on vector indexes and filters candidate knowledge tokens using the knowledge graph structure. Following verification by permission control fields and the invocation contract, a set of authorized knowledge tokens is generated and assigned a session identifier. Taking a specific collaborative task as an example, the average similarity between the task vector and the semantic vector of the knowledge tokens is 0.83. After permission and contract verification, approximately 14.6% of the tokens that do not meet the requirements are removed, resulting in a stable set of authorized knowledge tokens. Agents read the tokens from this set, generate local decision results, and then perform collaborative fusion using value tags and semantic weights to form the final collaborative decision result.

[0132] In continuous deployment tests, the application effect of this invention has significant advantages. First, the collaborative consistency of multi-agents is significantly improved. Traditional methods achieve an average decision consistency of 78.4% under high load conditions, while the consistency is improved to 93.1% after adopting this invention. Second, decision latency is significantly reduced. Under the same conditions, the average latency of knowledge token retrieval based on this invention is 47 milliseconds, which is about 41% less than that of traditional structured database retrieval. In addition, since the knowledge token has source fingerprints and invocation contracts, the verifiability of decisions is significantly improved. In a random sampling of 980 multi-agent collaborative events, the success rate of decision chain tracing reached 100%, which is 18 percentage points higher than that of traditional solutions.

[0133] More importantly, this invention demonstrates greater robustness in complex conflict scenarios. Under high-concurrency knowledge invocation, the knowledge parsing error rate of traditional methods reaches 2.7%, while the knowledge token parsing error rate of this invention is only 0.3% under the same conditions, a reduction of 89%. This indicates that the structured encapsulation of knowledge tokens and semantic vector management effectively reduce the risk of knowledge mismatch. Furthermore, through a lifecycle management mechanism, expired knowledge tokens are automatically removed, ensuring that the knowledge set used by the agent remains up-to-date, significantly reducing decision bias.

[0134] Table 1. Performance Statistics of Multi-Agent Collaborative Decision-Making Based on Knowledge Tokens

[0135] Indicator Name Traditional numerical methods The numerical value of this invention Increase ratio Test sample size Remark Decision consistency (%) 78.4 93.1 +18.8% 1200 Multi-agent consensus assessment Average retrieval latency (ms) 80 47 -41.3% 35000 Semantic Vector Retrieval Load Test Success rate of decision chain tracing (%) 82 100 +22.0% 980 Trustworthy traceability capability assessment Knowledge analysis error rate (%) 2.7 0.3 -89% 64000 High concurrency call conditions Token redundancy rate (%) 9.2 3.0 -67.4% 54000 Knowledge encapsulation efficiency statistics Illegal call interception rate (%) 71 98 +38.0% 2100 Permission verification module statistics

[0136] As can be seen from the above embodiments, the present invention can significantly enhance knowledge credibility, improve semantic consistency, reduce decision delay and parsing error risk in complex multi-agent collaborative scenarios, and realize a collaborative decision-making link with full traceability, thus demonstrating comprehensive and significant technical advantages in practical applications.

[0137] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multi-agent collaborative decision-making method based on knowledge tokens, characterized in that, Includes the following steps: S1. Construct a knowledge base architecture, acquire multi-source data and perform content hash calculation, anchor the hash result to an immutable storage in the original document fidelity layer, generate the original document fingerprint and record the original document fingerprint identifier; S2. Based on the original document fingerprint, perform knowledge extraction, encapsulate the extracted structured content into a knowledge token, and write a unique identifier, semantic vector, source fingerprint, access control field, value tag and confidence level, calling contract, association relationship, data traceability information, health status and lifecycle information into it. S3. Write the knowledge token into the activated knowledge layer, construct a vector index based on the semantic vector, and map the association relationship of the knowledge token to knowledge graph nodes to form a bidirectional link with the original fingerprint identifier. S4. In the unified index layer, based on the vector index and knowledge graph, the task requirements are matched, the candidate knowledge token set is filtered, and the legality is verified based on the permission control field and the calling contract to obtain the authorized knowledge token set. S5. The multi-agent system generates local decision results based on the authorized knowledge token set, and then integrates the local decision results with the value tags and semantic vectors of the knowledge tokens to form a collaborative fusion result. S6. Perform consistency verification on the collaborative fusion results. After the verification is passed, bind the unique identifier of the knowledge token used, generate collaborative decision results and record data lineage information.

2. The multi-agent collaborative decision-making method based on knowledge tokens according to claim 1, characterized in that, The knowledge base architecture consists of an original document fidelity layer, an active knowledge layer, and a unified index layer, specifically including: The original document fidelity layer performs integrity calibration on multi-source data based on the content hashing algorithm and anchors the hash result to tamper-proof storage to generate the original document fingerprint. The multi-source data consists of structured data collected from business databases, sensor terminals, log records, external knowledge bases, and document storage. After synchronous capture, format parsing, and time calibration are performed by the data acquisition engine, the data is aggregated to the original document fidelity layer to generate the original document fingerprint. The activated knowledge layer performs structured parsing and semantic segmentation based on the original fingerprint identifier, and encapsulates the parsing result into a knowledge token with a unique identifier, semantic vector, source fingerprint, access control field, value tag and confidence level, calling contract, association relationship, data traceability information, health status and life cycle information; The unified index layer establishes a vector index based on the semantic vector of the knowledge token and constructs knowledge graph nodes based on the association relationship, realizing bidirectional linking between the knowledge token and the original fingerprint and cross-layer traceable retrieval.

3. The multi-agent collaborative decision-making method based on knowledge tokens according to claim 2, characterized in that, The original document fidelity layer is used to solidify and reliably identify the content of multi-source data, specifically including: The hash engine receives raw data from sensors, logs, business databases, and file storage, and performs content hash calculations to generate hash digests. The hash digest is concatenated and encoded with the timestamp, data source identifier, and data type label to form the original document fingerprint; The original document fingerprint is anchored to an immutable storage medium via a blockchain or distributed file system, and a corresponding storage address index is generated, which is the original document fingerprint identifier. Write the original document fingerprint and original document fingerprint identifier into the authenticity index table for knowledge extraction and traceability operations; Upon completion, a hash verification task is triggered to compare the latest hash digest with the original record to confirm that the original content has not been modified.

4. The multi-agent collaborative decision-making method based on knowledge tokens according to claim 2, characterized in that, The activated knowledge layer is used to realize the structured encapsulation and dynamic management of knowledge, specifically including: Based on the original fingerprint reading of the corresponding original content, text parsing, entity recognition, attribute extraction and semantic segmentation operations are performed to generate structured knowledge units; Structured knowledge units are encoded into knowledge tokens, and unique identifiers, semantic vectors, source fingerprints, access control fields, value tags and confidence levels, invocation contracts, relationships, data traceability information, health status and lifecycle information are written into the knowledge tokens. Establish an association table based on the relationships between knowledge tokens, and cluster and store related knowledge tokens. Perform a content signing operation on each knowledge token, and reverse-bind the signature hash with the original document fingerprint to ensure that the encapsulated content is traceable; A token index list is generated at the activation knowledge layer for the unified index layer to read.

5. A multi-agent collaborative decision-making method based on knowledge tokens according to claim 2, characterized in that, The unified index layer is used to achieve efficient retrieval and related reasoning of knowledge tokens, specifically including: Receive a list of knowledge tokens from the activated knowledge layer, extract the semantic vector of each knowledge token, and construct a vector index space using a vectorized retrieval algorithm; Parse the association table information between knowledge tokens, construct a knowledge graph in the form of nodes and edges, and write semantic similarity, association relationship and calling contract into the edge weight field of the graph. After index initialization, a bidirectional mapping between the Token index and graph nodes is executed, enabling knowledge tokens to be located through semantic vector retrieval and to track associated content through graph structure relationships. When a task query request is received, the semantic similarity score is calculated in the vector index based on the task description, and an association expansion search is performed in the knowledge graph to filter out a set of knowledge tokens that meet the permission control and contract constraints, and the authorization result is returned for multi-agent calls.

6. The multi-agent collaborative decision-making method based on knowledge tokens according to claim 4, characterized in that, The unique identifier is used to identify the single identity of the knowledge token and associate it with the original document fingerprint record. The semantic vector is generated by embedding and encoding structured knowledge units to support vector retrieval. The source fingerprint is calculated by content hashing to trace the origin of the original data. The access control field defines the callable scope and operation level of the knowledge token according to the access policy. The value tag and confidence level are generated by a dynamic scoring mechanism to quantify the quality and reliability of the knowledge token. The calling contract constrains the usage conditions and interaction parameters of the knowledge token in the form of interface specifications. The association relationship is constructed based on semantic dependency mapping to connect related knowledge token nodes to form a knowledge graph. The data traceability information records the complete link of the knowledge token generation and calling process through a fidelity index table. The health status and lifecycle information are generated by the system monitoring task to mark the validity and expiration status of the knowledge token.

7. The multi-agent collaborative decision-making method based on knowledge tokens according to claim 1, characterized in that, The authorized knowledge token set is generated by the unified index layer after a task request is triggered, based on a joint selection of semantic vectors and knowledge graphs. Specifically, it includes: Read the task description and calculate the semantic similarity score with each knowledge token in the vector index space, and filter the candidate knowledge token set according to the score; The access control field is invoked to verify the access level matching degree, and the interaction parameters are checked for compliance according to the invocation contract, filtering out knowledge tokens that do not meet the access policy. For knowledge tokens that pass the legality verification, a value label and confidence weighted calculation is performed, and knowledge tokens with scores higher than the threshold are selected to form the target set; The relationships between target sets are expanded into a graph, and semantic dependency nodes are added to form a complete contextual knowledge chain; Generate a set of authorized knowledge tokens and assign them unique identifiers. Record the original document fingerprints, source hashes, and permission verification results involved, and use them for collaborative decision-making by multiple agents.

8. The multi-agent collaborative decision-making method based on knowledge tokens according to claim 1, characterized in that, S5 specifically includes: S51. Analyze the semantic vector and value tag information in the authorized knowledge token set, combine the task objective to execute strategy reasoning and action selection, and output the preliminary decision result that conforms to the call contract constraint; S52. Each intelligent agent binds the primary decision result with the corresponding knowledge token unique identifier to form a local decision result set; S53. Perform feature normalization and temporal alignment operations on the local decision result set, calculate the similarity matrix based on the semantic vector of the knowledge token, and use the attention weighting mechanism to weight and aggregate the local results with high similarity to generate a comprehensive feature representation. S54. The comprehensive feature representation is jointly fused with the value label, confidence level and correlation of the knowledge token to form a collaborative fusion result, and the fusion weight and the corresponding knowledge token mapping table are recorded to provide input data for consistency verification and collaborative decision output.

9. A multi-agent collaborative decision-making system based on knowledge tokens, executing the multi-agent collaborative decision-making method based on knowledge tokens as described in any one of claims 1 to 8, characterized in that, include: The knowledge base construction module is used to build the knowledge base architecture, acquire multi-source data and perform content hash calculation, anchor the hash result to immutable storage in the original document fidelity layer, generate original document fingerprints and record original document fingerprint identifiers; The knowledge encapsulation module is used to perform knowledge extraction based on the original document fingerprint identifier, encapsulate the extracted structured content into a knowledge token, and write a unique identifier, semantic vector, source fingerprint, access control field, value tag and confidence level, calling contract, association relationship, data traceability information, health status and lifecycle information. The knowledge mapping module is used to write knowledge tokens into the activated knowledge layer, construct vector indexes based on semantic vectors, and map the association relationships of knowledge tokens to knowledge graph nodes, forming a bidirectional link with the original fingerprint identifier. The knowledge filtering module is used to filter the candidate knowledge token set based on the vector index and knowledge graph matching task requirements in the unified index layer, and perform legality verification based on the permission control field and the calling contract to obtain the authorized knowledge token set. The decision generation module is used to generate local decision results based on the authorized knowledge token set, and to collaboratively integrate the local decision results with the value tags and semantic vectors of the knowledge tokens to form a collaborative integration result; The decision verification module is used to perform consistency verification on the collaborative fusion results. After the verification is passed, it binds the unique identifier of the knowledge token used, generates collaborative decision results, and records data lineage information.

Images