Multi-agent based data processing method and system

Abstract

The application provides a kind of data processing method and system based on multi-agent, belong to artificial intelligence technical field, this method includes: obtaining to be processed hypothesis data;In the running context that is isolated from each other and does not contain the historical reasoning data of arrangement equipment, respectively obtain the argument data of first sub-agent output and the attack data of second sub-agent output;According to both, generate revised hypothesis data containing falsification condition;Determine a plurality of evidence dimensions and corresponding call plan;Prioritize execution falsification query operation, when falsification query result meets falsification condition, generate interrupt instruction to terminate the query of remaining evidence dimensions and output negative result.The application blocks the path anchoring between sub-agents by isolating running context, deduces the effectiveness of hypothesis confrontation stress test and revision is improved;Through the short-circuit interruption control of falsification priority triggering, deduce the invalid stacking of positive evidence is avoided, and the system computing resource consumption and query call overhead are reduced.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a data processing method and system based on multiple agents. Background Technology

[0002] As the complexity of scientific research scenarios, such as scientific mechanism exploration and mechanistic analysis, increases, the use of multi-agent collaborative technology to automatically construct testable hypotheses and organize multi-source data for verification has become an important need in this field.

[0003] Existing multi-agent research assistance solutions typically construct a full-scale evidence collection task covering multiple evidence dimensions after generating a hypothesis, and concurrently call various external databases, numerical simulation tools, and literature interfaces to perform the entire process of querying. After all the query and calculation operations for all evidence dimensions are completed, the verification results are summarized and generated.

[0004] The aforementioned existing technologies result in a large amount of invalid data calls and computational consumption, leading to high computational resource overhead and query costs. Summary of the Invention

[0005] This invention provides a data processing method and system based on multiple agents to overcome the shortcomings of the prior art and reduce the consumption of system computing resources and query call overhead.

[0006] This invention provides a multi-agent-based data processing method for use in orchestration equipment, comprising the following steps: Obtain the hypothesis data to be processed; The hypothesis data to be processed is input to the first sub-agent and the second sub-agent respectively, and the argument data output by the first sub-agent and the attack data output by the second sub-agent are obtained; wherein, the first sub-agent and the second sub-agent are configured with mutually isolated running contexts, and the running contexts do not contain the historical inference data of the orchestration device; Based on the argumentation data and the attack data, revised hypothesis data is generated; the revised hypothesis data includes falsification conditions. Based on the revised hypothesis data, multiple evidence dimensions are determined, as well as the invocation plan corresponding to each evidence dimension; wherein, the evidence dimension represents the evidence acquisition category used to verify the revised hypothesis data; The falsification query operation corresponding to the falsification condition is executed based on the call plan; wherein, if the falsification query result satisfies the falsification condition, an interruption instruction is generated to terminate the query operation corresponding to the remaining evidence dimensions, and negative result data is output.

[0007] According to a multi-agent data processing method provided by the present invention, before acquiring the hypothesis data to be processed, the method further includes: Knowledge entries are extracted from historical session files, and each knowledge entry is labeled with a status tag; wherein, the status tag includes at least one of the following: valid status, need to be corrected status, outdated status, pending verification status, and original data status; Perform integrity audit, consistency audit, and boundary audit on each knowledge item labeled with a status tag. After the audit is passed, the data source referenced for inference will be switched to the audited knowledge entries, and the historical session files will be moved to the archive area and configured as read-only.

[0008] According to a multi-agent data processing method provided by the present invention, the step of determining multiple evidence dimensions and a corresponding invocation plan based on the revised hypothesis data includes: Based on the causal chain in the revised hypothesis data, the revised hypothesis data is decomposed into multiple sub-problems; Retrieve skill units from the skill registry that match each of the sub-problems; The matched skill units are sorted according to the rule that fact query operations take precedence over computation query operations; The matched skill units are sorted according to the rule that fact query operations take precedence over calculation query operations and document cross-operations are placed at the end, and duplicate call operations are merged. Send probe query requests to each of the sorted skill units; when the probe query results indicate that no data is available, reduce the coverage of the corresponding evidence dimension; When the number of available evidence dimensions reaches a preset dimension threshold, an execution blueprint containing the dependencies of each skill unit is generated, and the execution blueprint is determined as the invocation plan; If the number of available evidence dimensions is lower than the preset dimension threshold, the modification operation on the revised hypothesis data is rolled back.

[0009] According to a multi-agent data processing method provided by the present invention, before acquiring the hypothesis data to be processed, the method further includes: Receive input data, classify the input data according to intent, and determine the corresponding entry mode; The step of inputting the hypothesis data to be processed into the first sub-agent and the second sub-agent respectively includes: When the entry mode is the first mode, the hypothesis data to be processed is input into the first sub-agent and the second sub-agent in the order of calling the first sub-agent first and then calling the second sub-agent. When the entry mode is the second mode, a correction list is generated based on historical feedback data, and the second sub-agent and the first sub-agent are called in the order of first calling the second sub-agent to launch an attack based on the correction list, and then calling the first sub-agent to perform hypothesis revision.

[0010] A data processing method based on multiple agents according to the present invention further includes: When the entry mode is the third mode, the current working phase is deduced based on the feature information of the session file, and the pre-stored status file is read and cross-validated with the deduced current working phase; if the cross-validation is inconsistent, the current working phase deduced from the feature information shall be taken as the standard and an alarm message shall be output. When the entry mode is the second mode, the number of feedback iterations is counted; when the number of feedback iterations reaches the first threshold, a hypothetical reversal prompt is output; when the number of feedback iterations reaches the second threshold, an interrupt command is generated to block automatic iteration.

[0011] According to a multi-agent data processing method provided by the present invention, the step of generating revised hypothesis data based on the argument data and the attack data includes: Determine the type of conflict between the argument data and the attack data; In the case of a conflict of fact, the data corresponding to the source that meets the preset conditions of timeliness and verifiability will be determined as the conflict resolution result. In the case where the conflict type is an interpretation conflict, the corresponding conflict content is marked as pending verification, and the conflict content marked as pending verification is transferred to the verification process after the falsification query operation. In the case of a logical conflict, if there is supplementary evidence, the supplementary evidence is obtained; if there is no supplementary evidence, the process is reversed and the hypothesis data to be processed is modified. In the case where the conflict type is a methodological conflict, assertions that exceed the capability boundary are restricted according to the capability boundary of the skill unit; The hypothetical data to be processed is revised based on the processing results corresponding to each of the conflict types, and the revised hypothetical data is generated.

[0012] According to a multi-agent data processing method provided by the present invention, after executing the falsification query operation corresponding to the falsification condition based on the invocation plan, the method further includes: If the falsification query result does not meet the falsification condition, at least two independent evidence sources are invoked according to the invocation plan to perform the corresponding evidence acquisition operations, and cross-validation is performed on each of the acquired evidence data. Each claim data generated based on the aforementioned evidence data is associated and bound with its corresponding source identifier; wherein, the source identifier includes tool name, query identifier or document access information; Filter out claim data that are not bound to the source identifier, and generate an evidence synthesis report based on each of the bound claim data.

[0013] A data processing method based on multiple agents according to the present invention further includes: When a rerun instruction is received for a specified evidence dimension, the specified evidence dimension is marked as a rerun state. If there is an associated evidence dimension that has the specified evidence dimension as a prerequisite, the associated evidence dimension is also marked as a rerun state, and the evidence acquisition operation corresponding to the evidence dimension marked as a rerun state is re-executed.

[0014] This invention also provides a multi-agent-based data processing system for use in orchestration equipment, comprising the following modules: The acquisition module is used to acquire hypothesis data to be processed. The processing module is used to input the hypothesis data to be processed into the first sub-agent and the second sub-agent respectively, and to obtain the argument data output by the first sub-agent and the attack data output by the second sub-agent; wherein the first sub-agent and the second sub-agent are configured with mutually isolated running contexts, and the running contexts do not contain the historical inference data of the orchestration device; The adjudication module is used to generate revised hypothesis data based on the argument data and the attack data; the revised hypothesis data includes falsification conditions. The planning module is used to determine multiple evidence dimensions and a corresponding invocation plan for each evidence dimension based on the revised hypothesis data; wherein, the evidence dimension represents the evidence acquisition category used to verify the revised hypothesis data; The verification module is used to execute the falsification query operation corresponding to the falsification condition based on the call plan; wherein, if the falsification query result satisfies the falsification condition, an interruption instruction is generated to terminate the query operation corresponding to the remaining evidence dimensions, and negative result data is output.

[0015] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the multi-agent-based data processing method as described above.

[0016] In summary, one or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: By acquiring hypothesis data to be processed, the data is input into a first sub-agent and a second sub-agent, respectively. The first sub-agent outputs argument data, and the second sub-agent outputs attack data. The first and second sub-agents are configured with isolated runtime contexts, which do not contain historical reasoning data from the orchestration device. This effectively avoids interference from historical data or the opponent's reasoning path anchoring during hypothesis reasoning debates, ensuring the objectivity and effectiveness of the hypothesis adversarial stress test. Based on the argument data and attack data, revised hypothesis data is generated, containing falsification conditions. This allows the hypothesis, refined through multi-agent debate, to possess clear and scientifically verifiable standards. By determining multiple evidence dimensions based on the revised hypothesis data, and the corresponding invocation plan for each evidence dimension, the evidence dimension represents the type of evidence acquired to verify the revised hypothesis data. Based on the invocation plan, the falsification query operation corresponding to the falsification condition is executed. If the falsification query result satisfies the falsification condition, an interrupt instruction is generated to terminate the query operations corresponding to the remaining evidence dimensions, and negative result data is output. This effectively short-circuits the invalid verification operations of the remaining evidence dimensions when the falsification condition is triggered, significantly reducing the system's computational resource consumption and query call overhead. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the overall process of the data processing method based on multiple agents provided by the present invention.

[0019] Figure 2 This is one of the flowcharts of the multi-agent-based data processing method provided by the present invention.

[0020] Figure 3 This is a schematic diagram of the structure of the data processing system based on multiple agents provided by the present invention.

[0021] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0023] It should be noted that in the description of this invention, the terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. The terms "upper," "lower," etc., indicating orientation or positional relationships according to the accompanying drawings, are only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the system or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0024] The terms "first," "second," etc., used in this invention are used to distinguish similar objects, not to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class, without limiting the number of objects; for example, a first object can be one or more. Furthermore, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0025] The following is combined Figures 1 to 4 This invention describes the data processing method, system, electronic device, storage medium, and computer program product based on multi-agent systems.

[0026] In the implementation of the method of this invention, the executing entity involves various computerized processing devices, including but not limited to: server clusters with multi-agent scheduling capabilities, distributed computing platforms deployed in the cloud, workstations configured with multi-process concurrent scheduling modules, and edge computing nodes integrating large language model inference interfaces. In the following detailed descriptions of various embodiments, the orchestration device is used as the unified method executing entity. The orchestration device assumes central scheduling responsibilities throughout the entire method execution process, including dispatching various sub-agents, managing context isolation between sub-agents, maintaining system phase states, and performing unified interface scheduling for external evidence sources and skill units.

[0027] In one possible implementation, this embodiment first describes the overall flow of the multi-agent-based data processing method. (Refer to...) Figure 1 , Figure 1 This is a schematic diagram of the overall process of the data processing method based on multiple agents provided by the present invention.

[0028] The orchestration device first performs an entry mode detection operation. Based on the feature information of the input data, it determines the entry mode of the current task as one of the following: zero-exploration mode, literature signal mode, experimental signal mode, feedback iteration mode, or session recovery mode, and routes the data to the corresponding preprocessing branch according to the determination result.

[0029] In the zero-based exploration mode, the orchestration device sequentially executes capability guidance and scope confirmation operations, as well as literature gap analysis and blank graph generation operations. This structures the literature survey results into partitioned representations such as known facts, open controversies, methodological gaps, and contributing entry points, generating blank graph data. Subsequently, a gate 1 judgment operation is executed to check whether there are any contributing research controversies in the blank graph data. If the gate 1 judgment result is negative, a rollback or reversal operation is triggered, returning to the literature gap analysis stage for re-execution. If the gate 1 judgment result is positive, the hypothesis generation and quality control stage begins, and the first and second sub-agents are scheduled to complete the debate process according to the order of the affirmative side debating first.

[0030] When the input mode is document signal mode, after the orchestration device performs directional document parsing and focused blank map generation operations, it directly enters the hypothesis generation and quality control stage. When the input mode is experimental signal mode, after performing signal structuring and candidate mechanism interpretation operations, it enters the hypothesis generation and quality control stage.

[0031] In the feedback iteration mode at the entry point, the orchestration device performs memory organization and signal classification operations to generate a set of knowledge entries with status labels, followed by a memory audit operation. If the memory audit fails and there is no finalized hypothesis data, it returns to the hypothesis generation and quality control stage for re-execution. If the memory audit passes, the second sub-agent is scheduled to debate with the first sub-agent in the order of the opposing side's first turn to complete the debate process, and then the detection pre-detection and inference chain blueprint generation stage begins.

[0032] When the entry mode is session recovery mode, the orchestration device performs a multi-feature phase recovery operation, derives the current working phase based on file feature information, and performs cross-validation with the pre-stored state file. Then, it continues running according to the recovered working phase. If the hypothesis corresponding to the recovered phase has been finalized but the debate or blueprint has not yet been completed, the debate and blueprint generation steps are completed before entering the detection pre-detection and inference chain blueprint generation stage. If the recovered phase has entered the verification stage, it directly enters the priority execution of the falsification path stage.

[0033] After completing the debate process for each entry mode, the process uniformly enters the detection pre-inspection and reasoning chain blueprint generation stage. Minimal detection query operations are performed on each evidence dimension to generate an execution blueprint containing dependencies. Subsequently, a gate 2 judgment operation is executed, verifying whether the number of available evidence dimensions for the hypothesis data reaches a preset dimension threshold. If the gate 2 judgment result is negative, a hypothesis rollback operation is triggered, returning to the hypothesis modification stage. When consecutive rollbacks reach a preset falsification rollback threshold, a blank rollback operation is triggered, returning to the blank graph analysis stage. If the gate 2 judgment result is positive, the process enters the priority execution of the falsification path stage, performing a falsification query operation. If falsification is triggered, the multi-source evidence synthesis stage is skipped, and the process directly enters the gate 3-dimensional level review operation. After the gate 3-dimensional level review operation is completed, a negative result is output and recorded. If falsification is not triggered, the process enters the multi-source evidence synthesis stage, invoking at least two independent evidence sources to perform cross-validation operations, followed by the gate 3-dimensional level review operation. If Gating 3 determines that there is an evidence dimension that needs to be re-run, a dimension rollback operation is triggered, returning to the multi-source evidence synthesis stage to re-execute the evidence acquisition operation corresponding to the affected evidence dimension. If Gating 3 passes the review, a summary report is executed and saved, generating an evidence synthesis report with source identifiers and saving the current session state, thus ending the process.

[0034] Next, the following embodiments will further describe in detail the specific implementation steps of the multi-agent-based data processing method. (Refer to...) Figure 1 , Figure 1 This is one of the flowcharts illustrating the multi-agent-based data processing method provided by the present invention, such as... Figure 1 As shown, the method includes steps 110 to 150: Step 110: Obtain the hypothesis data to be processed.

[0035] As the initial step in the process, the orchestration device receives externally input hypothesis data to be processed. This hypothesis data represents scientific hypothesis texts that have undergone validation and stress testing, and includes at least the core claims of the hypothesis, causal predictive relationships, and a preliminary description of the hypothesis's mechanism by the researcher. Optionally, after acquiring the hypothesis data, the orchestration device stores it in a structured manner for subsequent use by sub-agents.

[0036] Step 120: Input the hypothesis data to be processed into the first sub-agent and the second sub-agent respectively, and obtain the argument data output by the first sub-agent and the attack data output by the second sub-agent.

[0037] Specifically, the orchestration device initiates call requests to the first sub-agent and the second sub-agent respectively, and injects the hypothesis data to be processed as input into the runtime context space of the first sub-agent and the runtime context space of the second sub-agent respectively.

[0038] It should be noted that the first sub-agent and the second sub-agent are configured with mutually isolated runtime contexts, which do not contain historical inference data from the orchestration device.

[0039] As an optional implementation, the runtime context isolation mechanism is achieved by allocating an independent memory session space for each sub-agent call. When constructing the call requests for the first and second sub-agents, the orchestration device only transmits filtered factual artifacts such as hypothesis data to be processed, blank map data, and feedback lists to each sub-agent, without transmitting intermediate inference process data or internal inference trajectory data generated by the orchestration device in the current session to any sub-agent.

[0040] The first sub-agent assumes the role of argumentation, generating argumentation data that supports the hypothesis based on the hypothesis data to be processed in an independent operating context space. The argumentation data includes verifiable supporting arguments and relevant literature evidence.

[0041] Correspondingly, the second sub-agent assumes the role of attack, generating attack data targeting weaknesses in the hypothesis based on the hypothesis data to be processed in an independent operating context space. The attack data includes rebuttal arguments and potential experimentally falsifiable scenarios.

[0042] It is worth noting that, since the first and second sub-agents operate in their own isolated runtime contexts, the second sub-agent cannot obtain the argument reasoning paths already generated by the first sub-agent when generating attack data. This avoids the path anchoring effect caused by the second sub-agent seeing the argument reasoning paths first, thus ensuring the effectiveness of the adversarial stress test.

[0043] Subsequently, the orchestration device receives the argumentation data returned by the first sub-agent and the attack data returned by the second sub-agent, and aggregates and stores the two types of data.

[0044] Step 130: Generate revised hypothesis data based on the argument data and attack data; the revised hypothesis data contains falsification conditions.

[0045] Furthermore, based on the argumentation data output by the first sub-agent and the attack data output by the second sub-agent, the orchestration device performs revision processing on the hypothesis data to be processed, generating revised hypothesis data. The revised hypothesis data is an evolved version of the hypothesis formed by iteratively optimizing the hypothesis data to be processed after comprehensively analyzing the strengthening points and weaknesses revealed by the argumentation data and attack data.

[0046] It should be noted that the revised hypothesis data must include at least one falsification condition. A falsification condition represents the criteria under which the revised hypothesis data can be refuted based on observable experimental data or query results. The setting of falsification conditions ensures the clear verifiability of the subsequent verification process.

[0047] Step 140: Determine multiple evidence dimensions and the corresponding invocation plan for each evidence dimension based on the revised hypothesis data.

[0048] As an example, the orchestration device identifies and determines multiple evidentiary dimensions required to verify the revised hypothesis data based on the revised hypothesis data. Evidence dimensions characterize the type of evidence acquired to verify the revised hypothesis data; different evidence dimensions correspond to different types of external data sources or computational tools, exemplarily including structured statistics, pathway knowledge, simulation, authoritative database queries, and literature citations.

[0049] Accordingly, the orchestration device generates a corresponding invocation plan for each evidence dimension. The invocation plan contains the specific skill unit identifiers and invocation order relationships required when each evidence dimension is executed.

[0050] Step 150: Execute the falsification query operation corresponding to the falsification condition based on the call plan; if the falsification query result satisfies the falsification condition, generate an interrupt instruction to terminate the query operation corresponding to the remaining evidence dimensions, and output the negative result data.

[0051] Specifically, the orchestration equipment, according to the call plan, prioritizes executing falsification query operations corresponding to the falsification conditions in the revised hypothesis data. The falsification query operation initiates query requests to external evidence sources by calling database retrieval interfaces, computational simulation interfaces, or literature access interfaces, and obtains the falsification query results. The orchestration equipment monitors the falsification query results in real time and compares them with the preset falsification conditions in the revised hypothesis data.

[0052] As an optional implementation, if the falsification query result meets the falsification conditions, the orchestration device immediately generates an interrupt command. This interrupt command is sent to the query processes corresponding to the remaining unexecuted evidence dimensions in the call plan, forcibly terminating the query operations for those remaining evidence dimensions, releasing the corresponding system computing resources, and outputting negative result data. Here, the negative result data is fully preserved as a valid scientific research output, recording the detailed content of the falsification query result and the specific basis for triggering the falsification conditions.

[0053] By setting the falsification query operation as the mandatory first step and immediately generating an interrupt command to terminate the query operations of other evidence dimensions when a falsification condition is triggered, the orchestration device effectively overcomes the problem of biased positive evidence piling up that is common in existing multi-agent verification systems. Since the system can terminate the verification process and output conclusions without waiting for all evidence dimensions to complete their queries when a falsification condition is triggered, the number of invalid database query calls and computational simulation resource consumption are significantly reduced. Furthermore, the independent context isolation mechanism eliminates path anchoring interference between the first and second sub-agents, giving the attack data realistic adversarial value and ensuring the reliability of the revised hypothesis data.

[0054] In one possible implementation, before the main process described above is executed, the method further includes the following steps.

[0055] Step 210: Extract knowledge entries from the historical session files and label each knowledge entry with a status tag.

[0056] As an optional implementation, the orchestration device invokes a memory-organizing agent, which reads the entire contents of the historical session files. These files include hypothesis reports, debate records, verification conclusions generated in previous rounds, as well as experimental observation feedback submitted by researchers, records of literature gaps, and explanations of data errors. Based on this, the memory-organizing agent extracts knowledge entries with independent and verifiable semantics from the historical session files. Each knowledge entry represents a knowledge assertion or data fact that can be cited independently.

[0057] Furthermore, the arrangement device labels each knowledge entry with a status tag, which includes at least one of the following: valid status, need-to-correct status, outdated status, pending verification status, and raw data status. Specifically, a valid status indicates that the knowledge entry has been verified and is still valid at the current research stage; a need-to-correct status indicates that the knowledge entry has known errors or inconsistencies and needs to be corrected before it can be cited; an outdated status indicates that although the knowledge entry was valid in a historical period, it has been replaced by newer evidence or experimental conclusions and is no longer applicable to current reasoning; a pending verification status indicates that the knowledge entry does not yet have sufficient evidence to support its authenticity and needs to be verified in the subsequent evidence synthesis stage; and a raw data status indicates that the knowledge entry is unprocessed raw observation data and has not yet formed a form of inferable conclusion.

[0058] Step 220: Perform integrity audit, consistency audit, and boundary audit on each knowledge item after the status label is marked.

[0059] It is worth noting that the orchestration device performs three types of audit operations on all knowledge entries after they are labeled with status tags. These include: a completeness audit to check whether the coverage of the knowledge entries meets the basic requirements for supporting subsequent hypothesis reasoning, identifying and recording any missing knowledge domains; a consistency audit to check whether there are contradictory assertions between different knowledge entries, marking and recording the specific content and source of any detected contradictions; and a boundary audit to check whether the semantic scope of each knowledge entry exceeds the capability boundaries of the corresponding data source, identifying and marking assertions that exceed the semantic or database coverage of the tool.

[0060] Step 230: After the audit is passed, switch the data source referenced by the inference to the knowledge entries that have passed the audit, and move the historical session files to the archive area and configure them to read-only status.

[0061] If the integrity audit, consistency audit, and boundary audit all pass, the orchestration device performs a knowledge source switching operation, switching the data source referenced in the reasoning of all subsequent sub-agent calls to the audited set of knowledge entries, and using the audited set of knowledge entries as the sole knowledge source for subsequent reasoning.

[0062] Simultaneously, the orchestration device moves the entire historical session file into the archive storage area and configures the access permissions of the historical session files in the archive storage area to read-only. Optionally, the read-only configuration prevents subsequent running sub-agents from using the historical session file as inference input; the historical session file is only retained for future reference.

[0063] In one possible implementation, this embodiment describes in detail the process of determining multiple evidence dimensions and the invocation plan corresponding to each evidence dimension based on revised hypothesis data. The method specifically includes the following steps.

[0064] Step 310: Based on the causal chain in the revised hypothesis data, break down the revised hypothesis data into multiple sub-problems.

[0065] Specifically, the orchestration device analyzes the causal chain structure in the revised hypothesis data. The causal chain represents the logical dependencies and derivation order between various mechanism nodes in the revised hypothesis data. Based on this, the orchestration device decomposes the revised hypothesis data into multiple independently testable sub-problems according to the independence boundaries of each node in the causal chain. Each sub-problem corresponds to an independent mechanism node or a relationship derivation between mechanism nodes in the causal chain.

[0066] Step 320: Retrieve skill units that match each sub-problem from the skill registry.

[0067] As an optional implementation, the orchestration device invokes a skill discovery sub-agent, which uses the semantic descriptions of each sub-question as query input to perform a matching retrieval operation in the skill registry.

[0068] The skill registry stores various pluggable skill units, exemplarily including genome retrieval skill units, pathway query skill units, simulation skill units, and literature retrieval interface skill units. After matching and searching, the skill discovery sub-agent outputs a list of matched skill units and a description of the capability gaps not covered by any skill units for each sub-question.

[0069] Step 330: Sort the matched skill units according to the rule that fact query operations take precedence over calculation query operations and document cross-operations are placed at the end, and merge duplicate call operations.

[0070] Furthermore, the orchestration device performs a dependency sorting operation on all matched skill units, determining the execution order of each skill unit according to the following rules: factual query skill units are arranged before computational query skill units, and cross-validation skill units are placed at the end of the execution sequence. Typically, factual query skill units involve directly retrieving existing records from a structured database, while computational query skill units typically involve operations requiring numerical calculations or simulations of the data. The above sorting rules are established because the results of factual queries can provide a preliminary data foundation for computational queries, and the reliability of cross-validation depends on the support of preceding calculation results. In addition, after completing the sorting, the orchestration device identifies and merges semantically equivalent duplicate calls pointing to the same data source, avoiding redundant transmission of the same query request multiple times.

[0071] Step 340: Send probe query requests to each sorted skill unit; if the probe query results indicate that no data is available, reduce the coverage of the corresponding evidence dimension.

[0072] The orchestration device sends probe query requests to each skill unit after dependency ordering. These probe queries employ a minimal query approach, sending only a few key identifiers to probe the data availability of the corresponding skill unit, rather than performing a full data acquisition operation. Subsequently, the orchestration device receives the probe query results returned by each skill unit. If the probe query result indicates that the corresponding skill unit has no available data, the coverage score of the evidence dimension corresponding to that skill unit is reduced.

[0073] Step 350: If the number of available evidence dimensions reaches the preset dimension threshold, generate an execution blueprint containing the dependencies of each skill unit, and determine the execution blueprint as the invocation plan; if the number of available evidence dimensions is lower than the preset dimension threshold, roll back the modification operation on the revised hypothesis data.

[0074] Furthermore, the orchestration device counts the number of evidence dimensions with usable data confirmed through probing and querying, and compares the counted number with a preset dimension threshold. When the number of usable evidence dimensions reaches the preset dimension threshold, the orchestration device generates an execution blueprint. The execution blueprint includes a dependency graph between each skill unit, an estimated expected execution time for each skill unit, a call cost label, and a reliability label. The orchestration device then uses the execution blueprint as a call plan for subsequent human-machine gating confirmation before execution.

[0075] Optionally, the preset dimension threshold can be set to a range of 2 to 3 evidence dimensions, with the specific value configured based on the resource conditions of the actual deployment environment. Conversely, if the number of available evidence dimensions is lower than the preset dimension threshold, the orchestration device triggers a rollback operation, reverting to the modification of the revised hypothesis data and notifying researchers to make targeted adjustments to the revised hypothesis data to improve the fit between the hypothesis content and available data sources.

[0076] In one possible implementation, this embodiment describes the entry intent routing operation performed before acquiring the hypothesis data to be processed, and the process of dynamically adjusting the calling order of the first and second sub-agents according to the entry pattern. It is performed according to the following steps: Step 410: Receive input data, classify the input data according to intent, and determine the corresponding entry mode.

[0077] As a preliminary step in the process, the orchestration device receives input data submitted by the user. This input data can take the form of natural language text, attachments, experimental observation records, and historical session status files. Specifically, the orchestration device performs intent classification on the input data, extracting file features, semantic features, and structural features, and determines the entry mode based on a combination of these features. Entry modes include at least the following types: Zero-based exploration mode, indicating the user starts with a broad research interest and there are currently no existing hypothesis texts; Literature signal mode, indicating the user starts with a specific literature or set of literature; Experimental signal mode, indicating the user starts with experimental observations or data anomalies; Feedback iteration mode, indicating the user submits iterative feedback including revisions to existing hypotheses; and Session resumption mode, indicating the user wishes to resume the process from a point of interruption in a historical session. It is worth noting that the orchestration device tailors subsequent execution processes based on the identified entry mode, skipping process stages unrelated to the current entry mode, thereby avoiding process redundancy.

[0078] Step 420: When the entry mode is the first mode, the hypothesis data to be processed is input into the first sub-agent and the second sub-agent in the order of calling the first sub-agent first and then the second sub-agent.

[0079] As a specific implementation method, the first mode corresponds to the entry point scenario of exploring from scratch. As an optional implementation method, when the entry mode is the first mode, the orchestration device executes the debate process in the order of calling the first sub-agent first and then the second sub-agent: first, it initiates a call request to the first sub-agent, waits for the first sub-agent to generate argument data based on the hypothesis data to be processed, and then initiates a call request to the second sub-agent, which generates attack data based on the hypothesis data to be processed in an independent operating context space.

[0080] Step 420: When the entry mode is the second mode, a correction list is generated based on historical feedback data, and the second sub-agent and the first sub-agent are called in the order of first calling the second sub-agent to launch an attack based on the correction list and then calling the first sub-agent to execute hypothesis revision.

[0081] As a specific implementation method, the second mode corresponds to the entry point of the feedback iteration. When the entry mode is the second mode, the orchestration device first generates a comprehensive correction list based on the historical feedback data submitted by the user. This comprehensive correction list summarizes the key correction points identified in the historical feedback data, such as experimental observation anomalies, literature gaps, and data errors. Subsequently, the orchestration device executes the debate process in the order of calling the second sub-agent first, then the first sub-agent: first, the comprehensive correction list and the hypothesis data to be processed are input into the second sub-agent, which then launches a targeted attack on the hypothesis data based on the comprehensive correction list and generates attack data; then, the orchestration device sends a call request to the first sub-agent, which then revises the hypothesis data based on the attack data and generates argumentation data.

[0082] Furthermore, this embodiment provides a detailed description of the phase cross-validation operation in the third mode and the iteration number threshold protection mechanism in the second mode. The following steps are performed: Step 510: When the entry mode is the third mode, deduce the current working phase based on the feature information of the session file, and cross-validate the pre-stored status file with the deduced current working phase; if the cross-validation is inconsistent, the current working phase deduced from the feature information shall prevail and an alarm message shall be output.

[0083] As a specific implementation method, the third mode corresponds to the entry point for session resumption. When the entry mode is identified as the third mode, a dual verification operation of the phase state is performed. First, the current session file is read, and feature information in the session file is extracted. Feature information includes records of completed process stages, identifiers of generated workpiece types, and markers of the last successfully saved execution node. Based on the above feature information, a file feature deduction operation is performed to deduce the working phase that should continue running. The working phase corresponds to a specific execution stage in the main process flow of the method.

[0084] Simultaneously, a pre-stored state file is read from persistent storage. This state file records the working phase identifier explicitly saved by the system in the previous session. The current working phase derived from the file features is cross-validated and compared with the working phase recorded in the state file. If they match, the breakpoint resume process is initiated according to the validated working phase. If the cross-validation results are inconsistent, the current working phase derived from the file features takes precedence, overwriting the record in the state file. An alarm message is also output to the user terminal, explaining the differences between the two verification sources and the reason for using the file feature derivation result.

[0085] The execution logic of the continuation process follows these rules: If there is no finalized hypothesis data, the process continues from the hypothesis generation stage; if the hypothesis data is finalized but the debate process or execution blueprint is not yet completed, the incomplete debate and execution blueprint generation steps are completed, and after completion, the process undergoes human-machine gating confirmation before entering the verification stage; if the execution blueprint is confirmed but the verification process is not yet completed, the process continues from the falsification priority verification step, and the dimensional gating steps must not be omitted.

[0086] Step 520: When the entry mode is the second mode, count the number of feedback iterations; when the number of feedback iterations reaches the first threshold, output the assumption reversal prompt; when the number of feedback iterations reaches the second threshold, generate an interrupt command to block automatic iteration.

[0087] Furthermore, when the entry mode is the second mode, after each complete feedback iteration process, the feedback iteration count recorded in persistent storage is incremented by one, and the updated feedback iteration count is compared with the first threshold and the second threshold. The reference value for the first threshold is 3, and the reference value for the second threshold is 5. The specific threshold values ​​are adjusted according to the actual deployment configuration.

[0088] Specifically, when the number of feedback iterations reaches the first threshold, a hypothesis reversal prompt is output to the user. This prompt suggests that researchers reconsider the underlying assumptions upon which the current hypothesis relies, indicating the possibility that the current iteration direction may be trapped in a local optimization bottleneck. When the number of feedback iterations reaches the second threshold, an interrupt command is generated. This interrupt command terminates the system's behavior of automatically starting a new round of feedback iterations. Subsequent new rounds of feedback iterations can only be initiated after the user submits an explicit confirmation command; the system must not automatically enter the next iteration without explicit user confirmation.

[0089] In addition, when the number of consecutive falsification rollbacks reaches the preset falsification rollback threshold, a suggestion message is output to the user, prompting the researcher to consider rolling back to the blank map analysis stage or changing the current research direction.

[0090] In one possible implementation, this embodiment describes a detailed process for generating revised hypothesis data based on argumentation data and attack data, particularly differentiated adjudication rules for different conflict types. The process is performed according to the following steps: Step 610: Determine the type of conflict between the argument data and the attack data.

[0091] Specifically, the orchestration device receives the argumentation data returned by the first sub-agent and the attack data returned by the second sub-agent, performs conflict detection operations on the content segments in the two types of data that are contradictory or mutually questioning, and identifies the conflict type of each conflict pair.

[0092] There are four types of conflicts: factual conflict, interpretational conflict, logical conflict, and methodological conflict. Factual conflict indicates that the argumentative data and the attack data cite objective factual data from different sources and reach opposite conclusions; interpretational conflict indicates that both sides give different mechanistic explanations for the same factual data; logical conflict indicates that one side's reasoning chain has missing premises or incomplete derivation steps; methodological conflict indicates that one side cites assertions that exceed the capabilities of the current skill unit.

[0093] Step 620: In the case of a conflict type of factual conflict, the data corresponding to the source that meets the preset conditions of timeliness and verifiability is determined as the conflict resolution result.

[0094] As an optional implementation, when a factual conflict is identified, the orchestration device performs a timeliness assessment and a verifiability assessment on the data sources cited by both parties in the conflict. The timeliness assessment compares the publication time of each data source, prioritizing the more recent data source. The verifiability assessment checks whether each data source has publicly accessible original data records or endorsement from an authoritative institution.

[0095] Based on this, the orchestration device identifies the data corresponding to the source that simultaneously meets the preset conditions of timeliness and verifiability as the conflict resolution result of the factual conflict, and replaces the corresponding factual assertion in the revised hypothesis data with the conflict resolution result.

[0096] Step 630: If the conflict type is interpretation conflict, mark the corresponding conflict content as pending verification, and transfer the conflict content marked as pending verification to the verification process after the falsification query operation.

[0097] Unlike the aforementioned implementation, when an interpretation conflict is identified, the orchestration device marks the conflicting content corresponding to the interpretation conflict as pending verification in the revised hypothesis data, without making a definitive ruling on the interpretation conflict during the debate stage. Optionally, the orchestration device records the conflicting content marked as pending verification and its corresponding two interpretation paths in their entirety, and transfers the recorded content to the multi-source evidence synthesis verification process after the falsification query operation for further examination, using actual evidence data to determine the relative credibility of the two interpretation paths.

[0098] Step 640: In the case of a logical conflict, if there is supplementary evidence, obtain the supplementary evidence; otherwise, return to modify the hypothesis data to be processed.

[0099] Furthermore, in the event of a logical conflict, the orchestration device checks whether there is evidence that can be supplemented by an external data source for the party in the logical conflict that has missing premises or incomplete derivation steps.

[0100] If supplementary evidence exists, the orchestration device calls the corresponding skill unit to obtain the supplementary evidence, fills the missing position in the reasoning chain with the obtained evidence, improves the argumentation path, and incorporates the revised result into the revised hypothesis data; conversely, if no supplementary evidence exists, the orchestration device determines that the logical conflict cannot be repaired under the current evidence conditions, triggers a rollback operation, returns the hypothesis data to be processed to the modification stage, and requires targeted modification of the hypothesis claims that cause logical conflicts.

[0101] Step 650: In the case of a methodological conflict, restrict assertions that exceed the capability boundary according to the capability boundary of the skill unit.

[0102] When a methodological conflict is identified, the orchestration device reads the capability boundary description of the corresponding skill unit from the skill registry. The capability boundary description records the known limitations of the corresponding skill unit in terms of data type, analysis granularity, prediction range, etc.

[0103] Based on this, the orchestration device identifies assertions in the argumentation data or attack data that exceed the aforementioned capability boundaries, and performs restriction operations on the assertions that exceed the capability boundaries. The specific forms of restriction operations include narrowing the scope of the assertion to the actual capability coverage of the skill unit, adding a scope limitation description, or deleting assertions that completely exceed the capability boundaries.

[0104] Step 660: Revise the hypotheses to be processed based on the processing results corresponding to each conflict type, and generate revised hypothesis data.

[0105] Finally, the orchestration device summarizes the processing results corresponding to each conflict type in steps 620 to 650, and performs revisions on the corresponding positions of the hypothesis data to be processed according to each processing result, generating revised hypothesis data. The revised hypothesis data includes the corrections to the resolved conflicts, the explanations of the conflicts marked in the status to be verified, and the falsification conditions refined based on the conflict resolution results.

[0106] Optionally, there is a maximum limit to the number of debate rounds. The debate process can be terminated early if neither side generates any new independent arguments, and any unresolved issues will be transferred to the subsequent evidence synthesis stage.

[0107] In one possible implementation, this embodiment describes the subsequent multi-source evidence synthesis operation performed when the falsification query result does not meet the falsification conditions. The operation is performed according to the following steps: Step 710: If the falsification query result does not meet the falsification conditions, call at least two independent evidence sources according to the call plan to perform the corresponding evidence acquisition operations, and perform cross-validation on each acquired evidence data.

[0108] Specifically, if the falsification query result is confirmed to not meet the falsification conditions after comparison, the orchestration device continues to execute the subsequent evidence acquisition steps in the call plan. As an optional implementation, the orchestration device calls at least two independent evidence sources according to the call plan to perform evidence acquisition operations respectively. The independence between each evidence source is reflected in the independence of data collection methods, data source institutions, or analysis methodologies, so as to eliminate the influence of internal biases of a single data source on the verification conclusion.

[0109] For example, the types of evidence sources include structured statistical databases, biological pathway knowledge graphs, simulation computing platforms, authoritative professional databases, peer-reviewed literature collections, and manually verifiable computational reasoning modules. Each evidence source does not have a fixed and unique priority. The arrangement device dynamically selects the most suitable combination of evidence sources based on the specific prediction content of the revised hypothesis data.

[0110] Subsequently, the orchestration equipment performs cross-validation on the evidence data obtained from various evidence sources. The cross-validation operation examines the consistency and complementarity of evidence data from different sources in supporting the core predictions of the revised hypothesis data. For evidence data with discrepancies, the cause of the discrepancies is analyzed and the analysis conclusions are recorded.

[0111] Step 720: Associate and bind each claim data generated based on each piece of evidence data with the corresponding source identifier; the source identifier includes the tool name, query identifier or document access information.

[0112] Furthermore, the orchestration device generates corresponding claim data based on the evidence data obtained in step 710. Each claim data represents a support or supplementary explanation made by the evidence data for a certain prediction point in the revised hypothesis data. After generating the claim data, the orchestration device performs a source identifier association and binding operation on each claim data, binding the source identifier on which the claim data is based to each claim data.

[0113] Optionally, the source identifier may take the form of a tool name, a query identifier, or document access information: wherein, the tool name records the name of the skill unit invoked to generate the claim data; the query identifier records the unique identifier of the query request sent when performing the evidence acquisition operation; and the document access information records the traceable access address, such as the digital object identifier or Uniform Resource Locator, of the document cited in the claim data.

[0114] Step 730: Filter claim data that are not bound to source identifiers, and generate an evidence synthesis report based on the bound claim data.

[0115] Finally, the orchestration device performs a filtering operation on all generated claim data, removing claim data that has not been successfully bound to any source identifier from the claim data set. This removal operation ensures that assertive content without a source is not included in the final report. Based on this, the orchestration device generates an evidence synthesis report based on each filtered claim data. The evidence synthesis report includes the content of each claim data, the corresponding source identifier binding record, the cross-validation conclusions of each evidence dimension, and a comprehensive support assessment of the core predictions of the revised hypothesis data.

[0116] In one possible implementation, this embodiment describes the process of performing a dimension-level rerun and dependency collateralization marking operation when a rerun instruction for a specified evidence dimension is received during the evidence synthesis phase. This is performed according to the following steps: Step 810: When a rerun instruction is received for a specified evidence dimension, the specified evidence dimension is marked as a rerun state; if there are related evidence dimensions that have the specified evidence dimension as a prerequisite, the related evidence dimensions are also marked as rerun states, and the evidence acquisition operation corresponding to the evidence dimension marked as a rerun state is re-executed.

[0117] As an optional implementation, after completing the initial evidence synthesis operation, the orchestration device continuously listens for rerun instructions from the user terminal or the system review module. These rerun instructions include dimension identifiers specifying the evidence dimension, which indicates the specific type of evidence that the researcher or review module deems necessary to re-execute the evidence acquisition operation.

[0118] Specifically, upon receiving a rerun instruction, the orchestration device updates the status marker of the specified evidence dimension to the rerun state. Then, the orchestration device reads the dependency graph recorded in the execution blueprint, searching for all associated evidence dimensions in the dependency graph that have the specified evidence dimension as a prerequisite dependency. Here, a prerequisite dependency indicates that the evidence acquisition operation of a certain evidence dimension depends on the output result of the specified evidence dimension as input. If there is at least one associated evidence dimension that has the specified evidence dimension as a prerequisite dependency, the orchestration device marks all identified associated evidence dimensions as rerun states. This collective marking operation recursively propagates along the dependency chain in the dependency graph until all evidence dimensions affected by the output result of the specified evidence dimension are marked as rerun states.

[0119] Subsequently, the orchestration device re-executes the corresponding evidence acquisition operations for all evidence dimensions marked as re-running, according to the dependency sorting order recorded in the execution blueprint. The re-execution process reuses the original call plan configuration in the execution blueprint and replaces the original data of the corresponding dimension in the evidence synthesis report with the newly acquired evidence data. Optionally, when the sufficient skip condition is met, the orchestration device skips redundant evidence dimension re-running operations. The sufficient skip condition is determined by the system's preset evidence coverage evaluation rules. After re-execution, the orchestration device regenerates the evidence synthesis report based on the updated full evidence data. The updated evidence synthesis report includes the triggering reason record for this re-running operation and a list of affected dimensions.

[0120] Reference Figure 3 , Figure 3 This is a schematic diagram of the structure of the multi-agent-based data processing system provided by the present invention. The system includes: The acquisition module is used to acquire hypothesis data to be processed; The processing module is used to input the hypothesis data to be processed into the first sub-agent and the second sub-agent respectively, and to obtain the argument data output by the first sub-agent and the attack data output by the second sub-agent; wherein the first sub-agent and the second sub-agent are configured with mutually isolated running contexts, and the running contexts do not contain the historical reasoning data of the orchestration device; The adjudication module is used to generate revised hypothesis data based on the argument data and attack data; the revised hypothesis data includes falsification conditions. The planning module is used to determine multiple evidence dimensions based on the revised hypothesis data, as well as the corresponding invocation plan for each evidence dimension; among them, the evidence dimension represents the type of evidence obtained to verify the revised hypothesis data; The verification module is used to execute the falsification query operation corresponding to the falsification condition based on the call plan; wherein, if the falsification query result meets the falsification condition, an interruption instruction is generated to terminate the query operation corresponding to the remaining evidence dimensions, and negative result data is output.

[0121] It should be noted that the multi-agent-based data processing system provided by the present invention can execute the multi-agent-based data processing method of any of the above embodiments during specific operation, which will not be elaborated in this embodiment.

[0122] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 4As shown, the electronic device may include a processor 410, a communications interface 420, a memory 430, and a communication bus 440, wherein the processor 410, the communications interface 420, and the memory 430 communicate with each other via the communication bus 440. The processor 410 can call logical instructions in the memory 430 to execute the multi-agent-based data processing methods provided in the above embodiments.

[0123] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0124] On the other hand, the present invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions, and when the program instructions are executed by a computer, the computer is able to execute the multi-agent-based data processing method provided in the above embodiments.

[0125] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the multi-agent-based data processing methods provided in the above embodiments.

[0126] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0127] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0128] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A data processing method based on multi-agent systems, characterized in that, Used in orchestration equipment, including: Obtain the hypothesis data to be processed; The hypothesis data to be processed is input into the first sub-agent and the second sub-agent respectively, and the argument data output by the first sub-agent and the attack data output by the second sub-agent are obtained; wherein, the first sub-agent and the second sub-agent are configured with mutually isolated running contexts; Based on the argumentation data and the attack data, revised hypothesis data is generated; the revised hypothesis data includes falsification conditions. Based on the revised hypothesis data, multiple evidence dimensions are determined, as well as the invocation plan corresponding to each evidence dimension; wherein, the evidence dimension represents the evidence acquisition category used to verify the revised hypothesis data; The falsification query operation corresponding to the falsification condition is executed based on the call plan; wherein, if the falsification query result satisfies the falsification condition, an interruption instruction is generated to terminate the query operation corresponding to the remaining evidence dimensions, and negative result data is output.

2. The method according to claim 1, characterized in that, Before acquiring the hypothesis data to be processed, the following steps are also included: Knowledge entries are extracted from historical session files, and each knowledge entry is labeled with a status tag; wherein, the status tag includes at least one of the following: valid status, need to be corrected status, outdated status, pending verification status, and original data status; Perform integrity audit, consistency audit, and boundary audit on each knowledge item labeled with a status tag. After the audit is passed, the data source referenced for inference will be switched to the audited knowledge entries, and the historical session files will be moved to the archive area and configured as read-only.

3. The method according to claim 1, characterized in that, The process of determining multiple evidence dimensions based on the revised hypothesis data, and the corresponding invocation plan for each evidence dimension, includes: Based on the causal chain in the revised hypothesis data, the revised hypothesis data is decomposed into multiple sub-problems; Retrieve skill units from the skill registry that match each of the sub-problems; The matched skill units are sorted according to the rule that fact query operations take precedence over computation query operations; The matched skill units are sorted according to the rule that fact query operations take precedence over calculation query operations and document cross-operations are placed at the end, and duplicate call operations are merged. Send probe query requests to each of the sorted skill units; when the probe query results indicate that no data is available, reduce the coverage of the corresponding evidence dimension; When the number of available evidence dimensions reaches a preset dimension threshold, an execution blueprint containing the dependencies of each skill unit is generated, and the execution blueprint is determined as the invocation plan; If the number of available evidence dimensions is lower than the preset dimension threshold, the modification operation on the revised hypothesis data is rolled back.

4. The method according to claim 1, characterized in that, Before acquiring the hypothesis data to be processed, the following steps are also included: Receive input data, classify the input data according to intent, and determine the corresponding entry mode; The step of inputting the hypothesis data to be processed into the first sub-agent and the second sub-agent respectively includes: When the entry mode is the first mode, the hypothesis data to be processed is input into the first sub-agent and the second sub-agent in the order of calling the first sub-agent first and then calling the second sub-agent. When the entry mode is the second mode, a correction list is generated based on historical feedback data, and the second sub-agent and the first sub-agent are called in the order of first calling the second sub-agent to launch an attack based on the correction list, and then calling the first sub-agent to perform hypothesis revision.

5. The method according to claim 4, characterized in that, Also includes: When the entry mode is the third mode, the current working phase is deduced based on the feature information of the session file, and the pre-stored status file is read and cross-validated with the deduced current working phase; if the cross-validation is inconsistent, the current working phase deduced from the feature information shall be taken as the standard and an alarm message shall be output. When the entry mode is the second mode, the number of feedback iterations is counted. When the number of feedback iterations reaches a first threshold, a hypothetical reversal prompt message is output. When the number of feedback iterations reaches the second threshold, an interrupt command is generated to block automatic iteration.

6. The method according to claim 1, characterized in that, The step of generating revised hypothesis data based on the argumentation data and the attack data includes: Determine the type of conflict between the argument data and the attack data; In the case of a conflict of fact, the data corresponding to the source that meets the preset conditions of timeliness and verifiability will be determined as the conflict resolution result. In the case where the conflict type is an interpretation conflict, the corresponding conflict content is marked as pending verification, and the conflict content marked as pending verification is transferred to the verification process after the falsification query operation. In the case of a logical conflict, if there is supplementary evidence, the supplementary evidence is obtained; if there is no supplementary evidence, the process is reversed and the hypothesis data to be processed is modified. In the case where the conflict type is a methodological conflict, assertions that exceed the capability boundary are restricted according to the capability boundary of the skill unit; The hypothetical data to be processed is revised based on the processing results corresponding to each of the conflict types, and the revised hypothetical data is generated.

7. The method according to claim 1, characterized in that, After executing the falsification query operation corresponding to the falsification condition based on the invocation plan, the process further includes: If the falsification query result does not meet the falsification condition, at least two independent evidence sources are invoked according to the invocation plan to perform the corresponding evidence acquisition operations, and cross-validation is performed on each of the acquired evidence data. Each claim data generated based on the aforementioned evidence data is associated and bound with its corresponding source identifier; wherein, the source identifier includes tool name, query identifier or document access information; Filter out claim data that are not bound to the source identifier, and generate an evidence synthesis report based on each of the bound claim data.

8. The method according to claim 1, characterized in that, Also includes: When a rerun instruction is received for a specified evidence dimension, the specified evidence dimension is marked as a rerun state. If there is an associated evidence dimension that has the specified evidence dimension as a prerequisite, the associated evidence dimension is also marked as a rerun state, and the evidence acquisition operation corresponding to the evidence dimension marked as a rerun state is re-executed.

9. A data processing system based on multi-agent systems, characterized in that, Used in orchestration equipment, including: The acquisition module is used to acquire hypothesis data to be processed. The processing module is used to input the hypothesis data to be processed into the first sub-agent and the second sub-agent respectively, and to obtain the argument data output by the first sub-agent and the attack data output by the second sub-agent; wherein the first sub-agent and the second sub-agent are configured with mutually isolated running contexts, and the running contexts do not contain the historical inference data of the orchestration device; The adjudication module is used to generate revised hypothesis data based on the argument data and the attack data; the revised hypothesis data includes falsification conditions. The planning module is used to determine multiple evidence dimensions and a corresponding invocation plan for each evidence dimension based on the revised hypothesis data; wherein, the evidence dimension represents the evidence acquisition category used to verify the revised hypothesis data; The verification module is used to execute the falsification query operation corresponding to the falsification condition based on the call plan; wherein, if the falsification query result satisfies the falsification condition, an interruption instruction is generated to terminate the query operation corresponding to the remaining evidence dimensions, and negative result data is output.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the multi-agent-based data processing method as described in any one of claims 1 to 8.

Images