CN-121981286-A - Intelligent content tracing method and device, computer equipment and storage medium

CN121981286ACN 121981286 ACN121981286 ACN 121981286ACN-121981286-A

Abstract

The invention relates to the technical field of artificial intelligence and discloses an agent content tracing method, device, computer equipment and storage medium, which comprises the steps of constructing a dynamic trusted knowledge graph with version evolution records and reference relation promises by hashing knowledge block contents and linking metadata; when the agent generates the conclusion, the knowledge block hash and the reasoning steps quoted by the agent are recorded to generate the knowledge fingerprint representing the complete reasoning path, and zero knowledge reasoning evidence is constructed based on the knowledge fingerprint, so that the auditor can verify the deduction process and the credibility of the conclusion under the condition of not accessing the original knowledge content and the reasoning internal state. The intelligent content tracing method can be applied to content tracing scenes of financial science and technology and medical health, and achieves full-range tracing, non-falsification and privacy protection of intelligent body content.

Inventors

LIN YANYU
XU WEI
CHEN YOUXIN

Assignees

深圳平安通信科技有限公司

Dates

Publication Date: 20260505
Application Date: 20260408

Claims (10)

1. The intelligent content tracing method is characterized by comprising the following steps of: acquiring a plurality of knowledge blocks cited by an agent in the process of searching enhancement generation, carrying out hash calculation on the content of each knowledge block, and recording the content hash value and corresponding metadata in a license chain network to form a dynamic trusted knowledge graph with version evolution records and reference relation promises; when the intelligent agent generates an output conclusion according to user inquiry, the reasoning steps executed by the intelligent agent are formed into a verifiable calculation paradigm, merkle commitments are carried out on the input and output of each step of reasoning and the hash value of the cited knowledge block, and reasoning path commitments are generated; constructing a zero-knowledge reasoning proof based on the reasoning path promise, wherein the zero-knowledge reasoning proof is used for proving that the output conclusion is derived by the reasoning step strictly based on the quoted knowledge blocks, and the reasoning logic accords with a preset compliance rule; and outputting the output conclusion, the reasoning path promise and the zero knowledge reasoning proof together so as to verify the complete deduction process and the credibility of the conclusion under the condition that the auditor does not access the original knowledge content and the reasoning internal state.
2. The method for tracing content of an agent according to claim 1, wherein the steps of obtaining a plurality of knowledge blocks referenced by the agent in the process of retrieving and enhancing generation, performing hash computation on the content of each knowledge block, and recording the content hash value and the corresponding metadata in a license chain network to form a dynamic trusted knowledge graph with version evolution records and reference relation commitments, include: Acquiring a plurality of knowledge documents for intelligent physical examination cable enhancement generation; Dividing each knowledge document into a plurality of knowledge blocks with independent semantics, distributing a unique identifier for each knowledge block and calculating the content hash value of each knowledge block; packaging the content hash value and the corresponding metadata into transaction data, and submitting the transaction data to a license chain network; The transaction data is written into the blocks through a consensus mechanism to generate a chain identifier of each knowledge block on a license chain network, a global dynamic state commitment tree is constructed based on the transaction data of all the uplink knowledge blocks, the global dynamic state commitment tree root hash uniquely represents the current state of the dynamic trusted knowledge graph, and version change information of each knowledge block is recorded through a chain transaction history; And establishing a mapping relation between the chain identifier of each knowledge block and the local storage path of each knowledge block so that an intelligent agent can locate and acquire the original content of the corresponding knowledge block through the chain identifier during reasoning, thereby forming a dynamic trusted knowledge map capable of dynamically evolving.
3. The method according to claim 1, wherein when the agent generates the output conclusion according to the user query, the step of reasoning performed by the agent is formed into a verifiable calculation paradigm, and Merkle commitment is performed on the input and output of each step of reasoning and the hash value of the referenced knowledge block, and the generation of the reasoning path commitment includes: In the process that the intelligent agent generates an output conclusion according to user inquiry, recording hash addresses of all knowledge blocks referenced by the output conclusion, wherein the hash addresses are obtained from the dynamic trusted knowledge graph; recording an reasoning step of the intelligent agent in the process of generating the output conclusion, wherein the reasoning step comprises a called logic rule, an intermediate reasoning result and a knowledge fusion operation; Carrying out hash calculation on the reasoning step to obtain a reasoning step hash; And combining the hash addresses of all the cited knowledge blocks with the reasoning step hash to generate a knowledge fingerprint.
4. The method according to claim 1, wherein the establishing a zero-knowledge reasoning proof based on the reasoning path commitment, the zero-knowledge reasoning proof being used to prove that the output conclusion is derived by the reasoning step based strictly on the referenced knowledge blocks, and the reasoning logic conforms to a preset compliance rule, includes: taking the reasoning path promise, an output conclusion and preset compliance constraint conditions as public input; The method comprises the steps of taking the content of an original knowledge block used by an agent in the reasoning process, the internal state variable of each step of reasoning and the reasoning step as private input, wherein the reasoning step comprises a logic rule identifier of each step of reasoning, an on-chain identifier of a quoted knowledge block, an intermediate result hash and an output result hash; Based on the zero knowledge proof algorithm, zero knowledge reasoning proof is generated according to public input and private input.
5. The agent content tracing method of claim 4, wherein said zero knowledge reasoning proof is used to verify that all references are identified on a knowledge block chain in said dynamic trusted knowledge graph; The zero knowledge reasoning proof is used for verifying that all the references on the knowledge block chain marks exist in the dynamic trusted knowledge graph; the zero knowledge reasoning proof is used to verify that the output conclusion is uniquely determined by the reasoning path.
6. The method according to claim 1, wherein the outputting the output conclusion, the inference path commitment and the zero knowledge inference proof together for the auditor to verify the complete deducing process and the credibility thereof without accessing the original knowledge content and inferring the internal state comprises: packaging the output conclusion, the reasoning path promise and the zero knowledge reasoning proof into an auditable execution proof package; generating a unique verification identifier for the auditable execution certification package, associating the verification identifier with a storage location of the auditable execution certification package, and optionally storing the verification identifier in a chain; And the auditor acquires the auditable execution proof package through the verification identifier, verifies the zero knowledge reasoning proof by using the public verification key, and confirms the reasoning integrity, the source authenticity and the logic compliance of the output conclusion if the verification is passed, without accessing any original knowledge content or the internal state of the intelligent body.
7. The agent content tracing method of claim 2, further comprising: When the content of the knowledge block in the dynamic trusted knowledge graph is updated, the content hash is recalculated, the updated hash and version information are submitted to the license chain network as new transaction data, a new on-chain identifier is generated after consensus, and meanwhile, the root hash of the global dynamic state promise tree is updated, so that the evolution history of the dynamic trusted knowledge graph is traceable and verifiable.
8. An intelligent content traceability device, which is characterized by comprising: The map construction unit is used for acquiring a plurality of knowledge blocks quoted by the agent in the process of searching, enhancing and generating, carrying out hash calculation on the content of each knowledge block, and recording the content hash value and corresponding metadata in a license chain network to form a dynamic trusted knowledge map with version evolution records and quoted relation promises; The path promise unit is used for forming the reasoning steps executed by the intelligent agent into a verifiable calculation paradigm when the intelligent agent generates an output conclusion according to the user inquiry, and carrying out Merkle promise on the input and output of each step of reasoning and the hash value of the quoted knowledge block to generate a reasoning path promise; The proving construction unit is used for constructing zero-knowledge reasoning proving based on the reasoning path promise, the zero-knowledge reasoning proving is used for proving that the output conclusion is derived by the reasoning step strictly based on the quoted knowledge blocks, and the reasoning logic accords with a preset compliance rule; and the output packaging unit is used for outputting the output conclusion, the reasoning path promise and the zero knowledge reasoning proof together so as to verify the complete deduction process and the credibility of the conclusion under the condition that the auditor does not access the original knowledge content and reason the internal state.
9. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that the processor implements the steps of the agent content tracing method of any one of claims 1 to 7 when the computer program is executed.
10. A computer readable storage medium storing a computer program, wherein the computer program when executed by a processor implements the steps of the agent content tracing method of any one of claims 1 to 7.

Description

Intelligent content tracing method and device, computer equipment and storage medium Technical Field The present invention relates to the field of artificial intelligence technologies, and in particular, to an agent content tracing method, apparatus, computer device, and storage medium. Background In the highly sensitive fields such as financial risk assessment and medical diagnosis assistance, intelligent agent platforms based on large language models are increasingly widely applied to automated conclusion generation, for example, key conclusions such as investment compliance judgment of intelligent agent output or medical image report interpretation need to have verifiable knowledge sources to support the credibility. However, existing intelligent systems have some drawbacks in terms of content traceability and trusted security. First, while current mainstream platforms provide reference links to knowledge segments through retrieval enhanced generation (RAG) techniques, these link information are typically stored in a centralized database, with the risk of tampering or loss in system migration, failing to form a non-repudiated evidence chain with legal effectiveness, especially one that is difficult to satisfy for data regulatory requirements. Secondly, when the intelligent agent generates a conclusion through multi-source knowledge fusion, the internal reasoning process of the intelligent agent presents black box characteristics, and although the reasoning steps can be partially shown through a thinking chain (CoT), the conclusion cannot be formally proved to be completely deduced from the cited specific knowledge document to a supervision organization or a medical review organization, and model illusions are not introduced in the process, so that the transparency of an AI decision is insufficient. In addition, in compliance audit or medical quality control scenarios, to verify the reliability of an agent conclusion, it is often necessary to disclose cited raw knowledge documents, such as customer sensitive information or patient medical record data, thereby introducing a serious risk of privacy leakage. Therefore, the prior art has obvious defects in the aspects of traceability credibility, process verifiability and privacy protection, and is difficult to meet the strict requirements of the fields of finance, medical treatment and the like on the credibility of the intelligent body content. Disclosure of Invention The invention provides an agent content tracing method, an agent content tracing device, computer equipment and a storage medium, which aim to solve the technical problem of how to make up for the defects of the existing agent system in content tracing and credibility guarantee, thereby meeting the strict requirements of the fields of finance, medical treatment and the like on the credibility of the agent content. In a first aspect, an agent content tracing method is provided, including: acquiring a plurality of knowledge blocks cited by an agent in the process of searching enhancement generation, carrying out hash calculation on the content of each knowledge block, and recording the content hash value and corresponding metadata in a license chain network to form a dynamic trusted knowledge graph with version evolution records and reference relation promises; when the intelligent agent generates an output conclusion according to user inquiry, the reasoning steps executed by the intelligent agent are formed into a verifiable calculation paradigm, merkle commitments are carried out on the input and output of each step of reasoning and the hash value of the cited knowledge block, and reasoning path commitments are generated; constructing a zero-knowledge reasoning proof based on the reasoning path promise, wherein the zero-knowledge reasoning proof is used for proving that the output conclusion is derived by the reasoning step strictly based on the quoted knowledge blocks, and the reasoning logic accords with a preset compliance rule; and outputting the output conclusion, the reasoning path promise and the zero knowledge reasoning proof together so as to verify the complete deduction process and the credibility of the conclusion under the condition that the auditor does not access the original knowledge content and the reasoning internal state. In a second aspect, an agent content tracing device is provided, including: The map construction unit is used for acquiring a plurality of knowledge blocks quoted by the agent in the process of searching, enhancing and generating, carrying out hash calculation on the content of each knowledge block, and recording the content hash value and corresponding metadata in a license chain network to form a dynamic trusted knowledge map with version evolution records and quoted relation promises; The path promise unit is used for forming the reasoning steps executed by the intelligent agent into a verifiable calculation paradigm when the intelligent agent generates