Inquiry interaction method and system for multi-agent cooperative scheduling

Abstract

The application discloses a multi-agent cooperative scheduling inquiry interaction method and system, and belongs to the technical field of agent cooperation; the method specifically comprises the following steps: obtaining a to-be-analyzed scheduling data set; based on the to-be-analyzed scheduling data set, a context cooperative agent group is called to perform cooperative theme merging, so that a candidate cooperative object set and a split-source context fragment set are obtained; based on the candidate cooperative object set and the split-source context fragment set, a split-source verification cooperative agent group is called to construct a multi-source cooperative differential evidence graph, so that a key cooperative trigger item set is generated; and based on the key cooperative trigger item set, an output cooperative agent group is called to generate a multi-agent cooperative scheduling interaction result; the application improves the accuracy of multi-agent cooperative scheduling and enhances the stability of the subsequent interaction process.

Description

Technical Field

[0001] This invention belongs to the field of intelligent agent collaboration technology, specifically relating to a query interaction method and system for multi-agent collaborative scheduling. Background Technology

[0002] In the process of multi-agent collaborative scheduling and interaction, the system needs to process the current input record, historical interaction record, associated collected data, and current interaction context of the same user, and schedule different agents to generate subsequent interaction content accordingly. Because data from different sources differ in expression, recording time, and content focus, simply aggregating multi-source data into a unified context still makes it difficult to determine which content has been covered by the current input, which content is additionally prompted by historical interaction records or associated collected data, and which content deviates between different sources, thus hindering the formation of accurate collaborative scheduling criteria.

[0003] In existing technologies, multi-agent collaborative scheduling query interaction methods typically gather multi-source information through context engineering, and then assign tasks such as semantic parsing, evidence extraction, data verification, and content generation to different agents for execution. However, this approach leans more towards task decomposition and lacks processing steps for source-specific evidence collection, intra-group cross-verification, and coverage differential comparison around the same collaborative object. It is difficult to transform the uncovered, partially covered, and deviation relationships between the current input and external prompts into key collaborative scheduling triggers, leading to subsequent interactions that easily miss external prompt information, repeatedly expand on covered content, or incorrectly inherit deviated information. This results in poor accuracy of multi-agent collaborative scheduling and poor stability of continuous interaction processes. Summary of the Invention

[0004] The purpose of this invention is to provide a query interaction method and system for multi-agent collaborative scheduling. By constructing a multi-source collaborative differential evidence graph, the method performs coverage differential comparison between input coverage evidence and external prompt evidence, and generates key collaborative scheduling trigger items, thereby improving the accuracy of multi-agent collaborative scheduling and enhancing the stability of subsequent interaction processes.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0006] On one hand, the present invention provides a query interaction method for multi-agent cooperative scheduling, including:

[0007] Obtain the scheduling dataset to be parsed;

[0008] Based on the scheduling dataset to be parsed, the context cooperative agent group is invoked to perform cooperative topic merging, resulting in a candidate cooperative object set and a source context fragment set.

[0009] Based on the candidate collaborative object set and the source context fragment set, the source verification collaborative intelligent agent group is called to construct a multi-source collaborative differential evidence graph and generate a key collaborative trigger item set. The key collaborative trigger item set is a set of trigger items formed by merging the uncovered relationship, partially covered relationship and deviation relationship after comparing the coverage difference between the input coverage evidence and the external prompt evidence through the multi-source collaborative differential evidence graph.

[0010] Based on the key collaborative trigger item set, the output collaborative agent group is invoked to generate multi-agent collaborative scheduling interaction results.

[0011] Specifically, based on the scheduling dataset to be parsed, the context cooperative agent group is invoked to perform cooperative topic merging, resulting in a candidate cooperative object set and a source context fragment set, including:

[0012] Based on the scheduling dataset to be parsed, text conversion and segmentation are performed to obtain the original context fragment set;

[0013] Based on the original set of context fragments, a set of candidate context units is generated by calling a context cooperative intelligent agent group. The set of candidate context units includes the recording time and candidate cooperative topics.

[0014] Based on the candidate context unit set, intra-group cross-verification is performed to obtain the effective context unit set;

[0015] Based on the set of effective context units, effective context units pointing to the same candidate collaborative topic are merged to obtain a set of candidate collaborative objects and a set of source context fragments.

[0016] Specifically, based on the original set of context fragments, the context cooperative agent group is invoked to generate a set of candidate context units, including:

[0017] Based on the original set of context fragments, the context collaborative intelligent agent group is invoked to extract fragment elements from each original context fragment to obtain fragment element candidate results.

[0018] Based on the candidate fragment element results, candidate collaborative topic matching is performed on each original context fragment to obtain the topic candidate results;

[0019] Based on the candidate results of fragment elements and the candidate results of themes, each original context fragment is organized into a unit to obtain a set of candidate context units.

[0020] Specifically, based on the candidate collaborative object set and the source context fragment set, the source verification collaborative agent group is invoked to construct a multi-source collaborative differential evidence graph and generate a key collaborative trigger item set, including:

[0021] Based on the candidate collaborative object set and the source context fragment set, a source verification task package is constructed for each candidate collaborative object. The source verification task package includes the current input fragment, historical record fragment, collected data fragment, as well as the corresponding candidate collaborative topic, fragment elements and recording time.

[0022] Based on the source verification task package, the source verification collaborative intelligent agent group is called to construct the expression coverage retrieval direction and the external source prompt retrieval direction, and the consistency comparison between sources is performed to obtain the candidate evidence set;

[0023] Based on the candidate evidence set, the source-based verification collaborative intelligent agent group is invoked to perform intra-group cross-verification and divergence backtracking screening to obtain the effective evidence set.

[0024] Based on the effective evidence set, a multi-source collaborative differential evidence graph is constructed, and coverage differential comparison and collaborative trigger item merging are performed to obtain the key collaborative trigger item set.

[0025] Specifically, based on the source-specific verification task package, the source-specific verification collaborative intelligent agent group is invoked to construct the expression coverage retrieval direction and the external source suggestion retrieval direction, and a consistency comparison between sources is performed to obtain a candidate evidence set, including:

[0026] Based on the current input fragment in the source-based verification task package, construct the expression coverage retrieval direction to obtain input coverage evidence;

[0027] Based on the historical record fragments and collected data fragments in the source verification task package, external source hint retrieval directions are constructed to obtain external source hint evidence;

[0028] Based on the consistency comparison between input coverage evidence and external cue evidence, external cue evidence that is inconsistent with the input coverage evidence in terms of candidate collaborative themes, fragment elements, or recording time is identified as deviation evidence.

[0029] A candidate evidence set is formed based on input coverage evidence, external cue evidence, and deviation evidence.

[0030] Specifically, based on the candidate evidence set, the source-based verification collaborative intelligent agent group is invoked to perform intra-group cross-verification and divergence backtracking screening to obtain the valid evidence set, including:

[0031] Based on the candidate evidence set, the source-based verification collaborative intelligent agent group is invoked to perform multi-dimensional cross-verification of each candidate evidence, and the evidence verification result set is obtained.

[0032] Based on the evidence verification result set, identify the matching discrepancies between the same candidate evidence and candidate collaborative objects and source verification task packages, and obtain the discrepancy evidence result set.

[0033] Based on the divergent evidence result set, the corresponding source verification task package is traced back, and the candidate evidence with divergence is corrected or removed to obtain the corrected evidence result set.

[0034] Based on the candidate evidence set and the revised evidence result set, duplicate evidence is merged and low-credibility evidence is screened out to obtain the effective evidence set.

[0035] Specifically, based on the effective evidence set, a multi-source collaborative differential evidence graph is constructed, and coverage differential comparison and collaborative trigger item merging are performed to obtain a key collaborative trigger item set, including:

[0036] Based on the effective evidence set and the candidate collaborative object set, node mapping is performed on the candidate collaborative objects and their corresponding input coverage evidence, external hint evidence and deviation evidence to obtain the collaborative object node set and the evidence node set.

[0037] Based on the collaborative object node set and the evidence node set, evidence association edges, coverage comparison edges and deviation indication edges are established to construct a multi-source collaborative differential evidence graph.

[0038] Based on the multi-source collaborative differential evidence graph, input coverage evidence nodes, external prompt evidence nodes, and deviation evidence nodes under the same collaborative object node are extracted to obtain the coverage comparison evidence group.

[0039] Based on the coverage comparison evidence set, coverage difference comparison is performed on the input coverage evidence nodes, external prompt evidence nodes and deviation evidence nodes to obtain a coverage difference relationship record set.

[0040] Based on the coverage differential relationship record set, supplementary collaborative trigger items, refined collaborative trigger items, and distinguishable collaborative trigger items are generated, and deduplication and sorting are performed to obtain the key collaborative trigger item set.

[0041] Specifically, based on the coverage comparison evidence set, coverage differential comparison is performed on the input coverage evidence nodes, external prompt evidence nodes, and deviation evidence nodes to obtain a coverage differential relationship record set, including:

[0042] Based on the coverage comparison evidence group, determine the set of input coverage evidence nodes and the set of deviation evidence nodes corresponding to each external prompt evidence node under the same collaborative object node;

[0043] Based on the coverage comparison edge between the external prompt evidence node and the set of input coverage evidence nodes, the coverage degree of each external prompt evidence node relative to the input coverage evidence is determined.

[0044] Based on the coverage level and the set of deviation evidence nodes, determine the uncovered relationship, partially covered relationship and deviation relationship corresponding to each external source hint evidence node;

[0045] Based on the uncovered, partially covered, and deviation relationships corresponding to each external source evidence node, the corresponding collaborative object nodes and evidence nodes are associated and recorded to obtain a set of coverage difference relationship records.

[0046] Specifically, based on the key collaborative triggering itemset, the output collaborative agent group is invoked to generate multi-agent collaborative scheduling interaction results, including:

[0047] Based on the set of key collaborative trigger items, the output collaborative agent group is called to perform trigger item parsing and interaction content generation for each key collaborative trigger item, thereby obtaining a set of candidate interaction content.

[0048] Based on the candidate interaction content set, each candidate interaction content undergoes trigger item verification, evidence coverage verification, and duplicate content removal to obtain subsequent interaction content;

[0049] Collect user feedback content corresponding to subsequent interaction content, and associate user feedback content with corresponding key collaborative trigger items to obtain feedback association records;

[0050] Based on feedback association records, the processing status of each key collaborative trigger item is determined, and subsequent interaction content, feedback content, and processing status are collaboratively merged to obtain the multi-agent collaborative scheduling interaction results.

[0051] On the other hand, the present invention provides a query interaction system for multi-agent cooperative scheduling, comprising:

[0052] The data acquisition module is used to acquire the scheduling dataset to be parsed.

[0053] The context collaboration module, based on the scheduling dataset to be parsed, calls the context collaboration agent group to merge collaborative topics, and obtains a candidate collaboration object set and a source context fragment set;

[0054] The source verification module, based on the candidate collaborative object set and the source context fragment set, calls the source verification collaborative intelligent agent group to construct a multi-source collaborative differential evidence graph and generate a key collaborative trigger item set;

[0055] The collaborative interaction module, based on the key collaborative trigger item set, calls the output collaborative agent group to generate multi-agent collaborative scheduling and interaction results.

[0056] Compared with the prior art, the beneficial effects of the present invention include: constructing a source-specific verification task package for candidate collaborative objects, and separately constructing description coverage retrieval direction and external prompt retrieval direction, extracting input coverage evidence and external prompt evidence separately, and then determining deviation evidence through consistency comparison between sources, providing a basis for collaborative scheduling of different agents, and improving the accuracy of multi-agent collaborative scheduling and subsequent interaction content generation; constructing a multi-source collaborative differential evidence graph, and performing coverage differential comparison on input coverage evidence nodes, external prompt evidence nodes, and deviation evidence nodes, merging uncovered relationships, partially covered relationships, and deviation relationships into key collaborative scheduling trigger items, enabling the output collaborative agent group to perform collaborative scheduling and generate subsequent interaction content based on clear trigger criteria, reducing repeated processing of covered content, and enhancing the stability of multi-agent collaborative scheduling and continuous interaction processes. Attached Figure Description

[0057] Figure 1 The flowchart of the query interaction method for multi-agent collaborative scheduling provided by the present invention is shown below.

[0058] Figure 2 This is a schematic diagram of the multi-source collaborative differential evidence construction provided by the present invention;

[0059] Figure 3 This is a structural diagram of the multi-agent collaborative scheduling query interaction system provided by the present invention. Detailed Implementation

[0060] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0061] Example 1:

[0062] See Figures 1-2 This embodiment provides a query interaction method for multi-agent cooperative scheduling, including the following specific steps:

[0063] Step S1: Obtain the scheduling dataset to be parsed.

[0064] In this embodiment, a scheduling dataset to be parsed is obtained. The scheduling dataset to be parsed includes the current input, historical interaction records, associated collected data, and current interaction context of the same user.

[0065] Current input includes text entered by the user in the current round, text transcribed from speech, or content selected by the interface. Each current input must include at least the user identifier, current round, input time, input text, and input source.

[0066] Historical interaction records include the input content and system output content of the same user in historical rounds. Each historical interaction record includes at least the user ID, historical round, historical input text, historical output text, and historical record time.

[0067] The associated collected data includes data collected by devices, form data, image recognition results, sensor results, or manually entered data. Each piece of associated collected data must include at least the user identifier, collection time, collection type, collection content, data source, and source credibility.

[0068] The current interaction context includes user input from previous rounds, system output, generated interaction content, and user feedback content. Each context record must include at least the round, role, text content, and time stamp.

[0069] The collection time range of historical interaction records is determined by statistical analysis of historical samples based on the timeliness of the interaction content. The effective impact duration of different types of interaction content on the current interaction is statistically analyzed, and the duration covering 95% of effective interactions is taken as the maximum collection time range. For example, in this application embodiment, the maximum collection time range of historical interaction records is preferably 30 days.

[0070] The triggering conditions for the associated data collection are preset according to the data type and application scenario. Data that is strongly related to user interaction behavior is automatically triggered for collection when the user initiates an interaction. The data collected periodically is the complete data within the most recent collection period. For example, the preferred collection period for the sensor data collected periodically in this application embodiment is 24 hours.

[0071] The range of collection rounds for the current interaction context is determined by statistical analysis of historical samples based on the semantic coherence of the context. The average number of context rounds required for different types of interactions is statistically analyzed, and the rounds that cover 90% of the complete semantics are taken as the maximum collection round range. For example, in this application embodiment, the maximum collection round range for the current interaction context is preferably 10 rounds.

[0072] The source credibility is pre-assigned based on the reliability level of the data source. The reliability level of data collected by official devices is the highest, and the reliability level of manually entered data is based on the qualification level of the data entry personnel. For example, in this application embodiment, the source credibility of data collected by smart wearable devices is preferably 0.90, and the source credibility of data manually entered by users is preferably 0.70.

[0073] For example, the scheduling dataset to be parsed for user A is collected, where the current input is the text content entered by A in the current round, the input time is the start time of the current round recorded by the system, and the input source is a text input box on a mobile device.

[0074] The historical interaction record contains all of A's interactions with the system over the past 30 days, including 12 user input texts and 12 corresponding system output texts. Each record contains a complete user identifier, round information, and timestamp.

[0075] The associated data collection includes heart rate, sleep, and activity data from the smart bracelet worn by A over the past 24 hours, as well as health status form data filled out by A 3 days ago. The source reliability of the smart bracelet data is 0.90, and the source reliability of the form data is 0.70.

[0076] The current interaction context is the interaction content of the previous 5 rounds in A's current round, including 3 user inputs and 2 system outputs. Each record contains the corresponding round, role and time stamp.

[0077] Step S2: Based on the scheduling dataset to be parsed, call the context cooperative agent group to perform cooperative topic merging to obtain a candidate cooperative object set and a source context fragment set.

[0078] The specific steps of step S2 are as follows:

[0079] Step S201: Based on the scheduling dataset to be parsed, perform text conversion and segmentation to obtain the original context fragment set.

[0080] In this embodiment, based on the scheduling dataset to be parsed, the non-text data is converted into text. The conversion rules are formulated according to the type and semantic features of the non-text data, and the structured data and multimedia recognition results are converted into natural language text, while retaining the core information and time attributes of the original data.

[0081] All converted text data is segmented into segments based on commas, periods, question marks, semicolons, line breaks, speech pause markers, and interface option boundaries, thus dividing long texts into independent basic semantic units.

[0082] The minimum interval for marking speech pauses was determined using the statistical 3δ principle. 100,000 speech interaction samples with different speaking speeds and accents were collected. The interval duration of all valid semantic pauses was statistically analyzed. The arithmetic mean and standard deviation of all interval durations were calculated. The mean minus three times the standard deviation was used as the lower limit of the minimum interval, and the mean plus three times the standard deviation was used as the upper limit of the minimum interval. Pauses exceeding these ranges were marked as invalid pauses and were not included in segmentation.

[0083] For example, the minimum interval of the voice pause marker in the embodiments of this application is preferably 0.3 seconds to 0.7 seconds, and the optimal value is 0.5 seconds.

[0084] The interactive control separators for interface option boundaries are set based on the common interaction specifications of mainstream operating systems and applications. Separator types include system native button boundaries, dropdown menu option separators, radio button and checkbox selection boundaries, input box submit button boundaries, and pop-up operation button boundaries. By calling the UI element recognition API provided by the operating system, the coordinate range and type information of all interactive controls in the current interactive interface are obtained. The boundaries between different controls are used as interface option boundaries to segment the interactive content generated by user selections.

[0085] For example, embodiments of this application preferentially use interactive control separators defined by Android system, iOS system and Web front-end standards.

[0086] The maximum segmentation length of long texts is determined by the statistical 3δ principle. 500,000 historical interactive text samples are collected, the character length of each complete semantic unit is counted, and the arithmetic mean and standard deviation of all length values ​​are calculated. The mean plus three times the standard deviation is taken as the maximum segmentation length. For continuous texts that exceed the maximum segmentation length and have no clear segmentation boundary, forced segmentation is performed at the maximum segmentation length position.

[0087] For example, in the embodiments of this application, the maximum segmentation length of long text is preferably 100 to 200 characters, with an optimal value of 150 characters.

[0088] The priority of the segmentation boundaries, from highest to lowest, is as follows: newline character, period, question mark, semicolon, voice pause marker, interface option boundary, comma. When there are multiple segmentation boundaries at the same position, the boundary with the highest priority is used for segmentation. For multiple consecutive occurrences of the same punctuation mark, only one is retained as the segmentation boundary.

[0089] A unique segment number is generated for each segmented segment, and the corresponding user identifier, segment text, data source, recording time, interaction round, data type, and source credibility fields are associated with it. All segments are then integrated into the original context segment set.

[0090] For example, the user's voice input "I haven't been sleeping well lately, I often wake up at night, and I'm not energetic during the day" is segmented. The speech recognition results detect three speech pause markers with intervals of 0.6 seconds and 0.5 seconds, respectively, both greater than the minimum interval of 0.5 seconds. Combined with comma boundaries, the text is segmented into three segments: "I haven't been sleeping well lately," "I often wake up at night," and "I'm not energetic during the day." Corresponding segment numbers are generated and associated with all field information.

[0091] Step S202: Based on the original set of context fragments, call the context collaborative agent group to generate a set of candidate context units.

[0092] The specific steps of step S202 are as follows:

[0093] Step S2021: Based on the original context fragment set, call the context collaborative intelligent agent group to extract fragment elements from each original context fragment to obtain fragment element candidate results.

[0094] In this embodiment, based on the original context fragment set, a context collaborative intelligent agent group is invoked. Multiple agents extract elements from each original context fragment from different semantic dimensions. The extracted elements include a set of keywords, a set of time words, a set of degree words, a set of negation words, a set of change words, a set of object words, and a set of action words. The keyword set is obtained by word segmentation, part-of-speech filtering, and professional dictionary matching. The word segmentation method adopts the Chinese word segmentation algorithm commonly used in this field. The part-of-speech filtering is based on the part-of-speech range corresponding to each element type. The professional dictionary is pre-constructed according to the application scenario. The element results extracted by each agent are integrated to obtain the fragment element candidate results.

[0095] For example, the update cycle of the professional dictionary in this application embodiment is preferably 7 days, and the scope of part-of-speech filtering preferentially covers nouns, verbs, adjectives and adverbs.

[0096] Step S2022: Based on the candidate fragment element results, perform candidate collaborative topic matching on each original context fragment to obtain the topic candidate results.

[0097] In this embodiment, based on the candidate fragment element results, the topic perspective agent in the context collaborative agent group is invoked to perform semantic clustering and topic summarization on the element set of each original context fragment, generating multiple candidate collaborative topics and corresponding topic descriptions.

[0098] The pre-trained text vector model adopts the BERT-base model with a 12-layer Transformer encoder structure, a hidden layer dimension of 768, 12 attention heads, and the GELU activation function. After pre-training on general Chinese corpora, the model is fine-tuned for domain adaptation in the multi-source interaction scenario of this solution.

[0099] The training dataset contains four types of data: general dialogue corpus, intelligent customer service interaction records, health consultation text, and device log descriptions, with a total sample size of 2 million. The fine-tuning process employs a contrastive learning framework, with a batch size of 32, a learning rate of 2e-5, and 3 training epochs. The cosine similarity loss function is used. After fine-tuning, the model can accurately extract semantic features from the interactive text and adapt to the expression styles of data from different sources.

[0100] The topic matching degree between each original context fragment and each candidate collaborative topic is calculated. The topic matching degree is the cosine similarity between the fragment element vector and the topic vector. The fragment element vector is generated by encoding the fragment element set through a pre-trained text vector model, and the topic vector is generated by encoding the candidate collaborative topic description through a pre-trained text vector model.

[0101] Multiple manually labeled topic matching samples are prepared as labeled samples. The minimum screening threshold for topic matching degree is determined through the labeled samples. Candidate thresholds are iterated from 0.1 to 0.9 in increments of 0.01. For each candidate threshold, precision and recall are calculated. Precision is the ratio of the number of correct matches to the number of system matches, and recall is the ratio of the number of correct matches to the number of matches that should have been matched by manual labeling. A comprehensive value is calculated for each candidate threshold. This comprehensive value is used to comprehensively measure precision and recall. The calculation method is as follows: multiply the product of precision and recall by two and use this as the numerator; use the sum of precision and recall as the denominator; divide the numerator by the denominator to obtain the comprehensive value. The threshold with the largest comprehensive value is selected as the minimum screening threshold for topic matching degree.

[0102] For edge deployment scenarios, the pre-trained text vector model is lightened and modified. The knowledge distillation method is adopted, and the fine-tuned BERT-base model is used as the teacher model to train the student model with a 6-layer Transformer structure. After training, the student model is dynamically quantized with INT8 to convert the model parameters from 32-bit floating-point numbers to 8-bit integers.

[0103] The candidate collaborative topic with the highest matching degree and greater than or equal to the minimum screening threshold is selected as the matching topic of the segment. Segments with a matching degree lower than the minimum screening threshold are marked as segments without a clear topic. The topic matching results of all segments are integrated into topic candidate results.

[0104] For example, the minimum screening threshold for topic matching degree in this application embodiment is preferably 0.6. The original context fragment "I wake up easily at night recently and feel a little tired during the day" is subject to topic matching to generate three candidate collaborative topics, namely abnormal sleep quality, poor mental state and physical fatigue. The matching degree between the fragment and the three candidate topics is calculated to be 0.85, 0.72 and 0.68, respectively. The abnormal sleep quality with the highest matching degree is selected as the matching topic of the fragment, and the corresponding topic candidate results are generated.

[0105] Step S2023: Based on the candidate results of fragment elements and the candidate results of themes, organize each original context fragment into units to obtain a set of candidate context units.

[0106] In this embodiment, based on the candidate results of fragment elements and the candidate results of the topic, each original context fragment is organized into a unit, generating a unique candidate unit number for each fragment. At the same time, the corresponding basic fragment number, fragment text, semantic keyword set, data source, recording time, time validity, candidate collaborative topic, topic description, generating agent identifier, and candidate confidence field are associated. The candidate confidence is calculated by weighting semantic completeness, source credibility, time validity, and topic clarity. Semantic completeness is the ratio of the number of extracted elements to the number of elements to be extracted, and topic clarity is the cosine similarity between the fragment vector and the topic vector. Each weight coefficient is determined by the analytic hierarchy process based on the degree of influence of each factor on the confidence. All the organized units are integrated into a candidate context unit set.

[0107] For example, in the embodiments of this application, the semantic completeness weight is preferably 0.35, the source credibility weight is preferably 0.25, the time validity weight is preferably 0.15, and the topic clarity weight is preferably 0.25.

[0108] Step S203: Based on the candidate context unit set, perform intra-group cross-verification to obtain the effective context unit set.

[0109] In this embodiment, based on the candidate context unit set, all agents in the context collaborative agent group are invoked to cross-verify each candidate context unit.

[0110] The context-cooperative intelligent agent group consists of 6 independent agents, each corresponding to one of the 6 verification dimensions: semantic verification agent, source verification agent, time verification agent, topic verification agent, user affiliation verification agent, and duplicate verification agent. Each agent is only responsible for the verification and scoring of its corresponding dimension, and they do not interfere with each other.

[0111] The review includes checking whether the semantics of the fragment are correct, whether the data source is correct, whether the time relationship is reasonable, whether the subject attribution is accurate, whether it belongs to the same user, and whether it is duplicated with other fragments.

[0112] Each agent gives a review score between 0 and 1 based on a preset unified scoring standard.

[0113] The semantic dimension of the fragment is as follows: 1 for complete semantic correctness and no ambiguity, 0.7 for basically correct semantics but with slight ambiguity, 0.4 for partial semantic errors but identifiable core information, and 0 for complete semantic errors or unidentifiable information.

[0114] In terms of data source dimension, 1 is assigned for complete and traceable source information, 0.7 is assigned for basically complete source information but missing some auxiliary fields, 0.4 is assigned for incomplete source information but can be inferred from related information, and 0 is assigned for missing source information and cannot be inferred.

[0115] In the time relationship dimension, 1 is assigned to accurate and logically consistent time information, 0.7 is assigned to basically accurate time information with minor errors, 0.4 is assigned to contradictory time information that can be corrected, and 0 is assigned to completely chaotic and uncorrectable time information.

[0116] In terms of topic attribution, a perfect match is 1, a basic match with minor deviations is 0.7, a partial match with the correct core direction is 0.4, and a complete mismatch is 0.

[0117] For the user affiliation dimension, 1 is assigned if the user identifiers are completely identical, 0.7 is assigned if the user identifiers are related and can be confirmed as the same user, 0.4 is assigned if the user identifiers are questionable but there is no contrary evidence, and 0 is assigned if the user identifiers are obviously inconsistent.

[0118] For the duplicate segment dimension, 1 is taken if there is no duplicate content, 0.7 is taken if there is a small amount of non-core duplicate content, 0.4 is taken if there is a large amount of core duplicate content, and 0 is taken if there is no duplicate content.

[0119] Calculate the cross-review score for each unit, which is the sum of the output scores of all review agents divided by the number of agents participating in the review.

[0120] Calculate the degree of divergence for each unit, calculate the difference between each review score and the cross-review score, square each difference, sum all the squared results, divide the sum of squared results by the number of agents participating in the review, and take the square root of the result to obtain the degree of divergence.

[0121] The effective value of the context unit for each unit is calculated. The effective value of the context unit is obtained by weighting the candidate confidence score, cross-validation score and divergence correction value. The divergence correction value is the result after subtracting the divergence score. The weight coefficients corresponding to the candidate confidence score, cross-validation score and divergence correction value are determined by optimization through labeled samples.

[0122] Specifically, multiple manually labeled context unit samples are prepared. Each sample includes candidate confidence, cross-validation score, divergence degree, and corresponding validity label. The validity label is used to indicate whether the corresponding context unit is a valid unit.

[0123] Based on the manually labeled context unit samples, a set of candidate weight combinations is constructed. The effective value of the corresponding context unit is calculated for each candidate weight combination. Based on the matching results between the calculated effective value of the context unit and the effective label, the precision and recall corresponding to each candidate weight combination are determined.

[0124] Calculate the comprehensive value corresponding to each candidate weight combination. The comprehensive value is calculated by multiplying the product of precision and recall by two as the numerator, using the sum of precision and recall as the denominator, and dividing the numerator by the denominator to obtain the comprehensive value.

[0125] The candidate weight combination with the largest comprehensive value is selected as the final weight coefficient.

[0126] The candidate retention threshold is determined based on the labeled samples.

[0127] Prepare multiple manually labeled context unit samples, each labeled as a valid or invalid unit. Iterate through candidate thresholds from 0.1 to 0.9, incrementing by 0.01. For each candidate threshold, calculate precision and recall. Precision is the ratio of correctly retained units to the system-retained units, and recall is the ratio of correctly retained units to the manually labeled units that should be retained. Calculate a composite value based on precision and recall. The composite value is calculated by multiplying the product of precision and recall by two (the numerator), using the sum of precision and recall as the denominator, and dividing the numerator by the denominator. Select the candidate threshold with the highest composite value as the candidate retention threshold. For example, in this embodiment, when there are no labeled samples, the preferred candidate retention threshold is 0.60 to 0.70, with an optimal value of 0.65. Retain context units with valid values ​​greater than or equal to the candidate retention threshold, and integrate the retained units into a valid context unit set.

[0128] For example, when the candidate context unit U001 is cross-verified, the verification scores output by the six agents are 0.9, 0.8, 0.9, 0.85, 1.0 and 0.95, respectively.

[0129] Calculate the cross-review score, sum the 6 review scores to get 5.4, divide 5.4 by 6 to get the cross-review score of 0.9.

[0130] The degree of divergence was calculated. The differences between each review score and the cross-review score of 0.9 were 0, -0.1, 0, -0.05, 0.1, and 0.05, respectively. The squares of these differences were calculated to obtain 0, 0.01, 0, 0.0025, 0.01, and 0.0025, respectively. The sum of the squares was obtained to get 0.025. 0.025 was divided by 6 to get 0.004167. The square root of 0.004167 was then taken to obtain the degree of divergence of 0.0645.

[0131] The candidate confidence level of this unit is 0.88. The effective value of the context unit is calculated based on the candidate confidence level weight of 0.50, the cross-validation score weight of 0.30, and the divergence correction value weight of 0.20. The divergence correction value is 1 minus 0.0645, which gives 0.9355.

[0132] The contribution value of candidate confidence is 0.88 multiplied by 0.50, which gives 0.44; the contribution value of cross-review score is 0.9 multiplied by 0.30, which gives 0.27; and the contribution value of divergence correction value is 0.9355 multiplied by 0.20, which gives 0.1871. The sum of the three contribution values ​​gives the effective value of the context unit, which is 0.8971.

[0133] Since the valid value of the context unit is 0.8971, which is greater than the candidate retention threshold of 0.65, the candidate context unit U001 is retained and added to the set of valid context units.

[0134] Step S204: Based on the set of valid context units, merge the valid context units pointing to the same candidate collaborative topic to obtain a set of candidate collaborative objects and a set of source context fragments.

[0135] In this embodiment, based on the effective context unit set, a merge score is calculated for any two effective context units. The merge score is obtained by weighting semantic similarity, topic similarity, temporal relevance, and source complementarity. Semantic similarity is the cosine similarity of the two unit segment vectors; topic similarity is the cosine similarity of the two unit topic vectors; temporal relevance is calculated exponentially based on the time interval between the two segments. Specifically, the calculation method is to use the natural constant as the base and the negative value of the ratio of the time interval between the two segments to the time base days as the exponent to obtain the temporal relevance. Source complementarity is calculated when the two segments have different sources. The threshold is set to 1, and 0.5 is used when the sources are the same. The merging threshold is determined based on the labeled samples. Units with a merging score greater than or equal to the merging threshold are merged into the same candidate collaborative object. The disjoint-set data structure is used to achieve batch merging of multiple units. The object confidence of each candidate collaborative object is calculated. The object confidence is the sum of the valid values ​​of all context units contained in the object divided by the number of units contained in the object. All candidate collaborative objects are integrated into a candidate collaborative object set. According to the data source, the fragments in each candidate collaborative object are divided into the current input fragment, historical record fragment, collected data fragment, and context fragment, resulting in a source-specific context fragment set.

[0136] Step S3: Based on the candidate collaborative object set and the source context fragment set, call the source verification collaborative agent group to construct a multi-source collaborative differential evidence graph and generate a key collaborative trigger item set.

[0137] The specific steps of step S3 are as follows:

[0138] Step S301: Based on the candidate collaborative object set and the source context fragment set, construct a source verification task package for each candidate collaborative object.

[0139] In this embodiment, based on the candidate collaborative object set, all candidate collaborative objects are traversed in order of their numbers, and the source context fragment set corresponding to each candidate collaborative object is extracted. The source context fragment set includes the current input fragment set, the historical record fragment set, the collected data fragment set, and the context fragment set.

[0140] A unique task package number is generated for each candidate collaboration object, and associated with the corresponding candidate collaboration object number, collaboration topic description, object confidence, and four types of source fragment sets. The integrity of each source fragment set is verified, missing basic field information is filled in, and all generated and verified task packages are integrated into a source verification task package set.

[0141] The integrity verification rules for the source fragment set are determined by labeled samples. For example, in this application embodiment, the preferred integrity verification rule is that a fragment missing any basic field is marked as an invalid fragment, and a task package with more than 10% of invalid fragments is re-collected.

[0142] For example, the user's candidate collaboration object number is 001, the collaboration topic description is abnormal sleep quality, the object confidence is 0.85, and the corresponding source context fragment set contains 3 current input fragments, 5 historical record fragments, 2 collected data fragments and 4 context fragments. All fragment fields are complete, and a source verification task package with the number T001 is generated.

[0143] Step S302: Based on the source verification task package, call the source verification collaborative intelligent agent group to construct the expression coverage retrieval direction and external source prompt retrieval direction, and perform consistency comparison between sources to obtain the candidate evidence set.

[0144] The specific steps of step S302 are as follows:

[0145] Step S3021: Based on the current input fragment in the source verification task package, construct the expression coverage retrieval direction to obtain input coverage evidence.

[0146] In this embodiment, based on the current input segment set in the source verification task package, collaborative topic descriptions are extracted, word segmentation and part-of-speech filtering are performed on the current input segment set to generate a current input keyword set, a pre-trained text vector model is called to convert each current input segment into vector form to generate a current input segment vector set, and the recording time information of each current input segment is extracted to generate current input time information.

[0147] Integrating collaborative topic descriptions, current input keyword sets, current input fragment vector sets, and current input time information, the system generates a statement coverage retrieval direction. It then calls upon the semantic verification agent in the source-based verification collaborative agent group to retrieve content already expressed or covered by the current input based on the statement coverage retrieval direction. The retrieved content undergoes semantic deduplication and integrity verification to generate input coverage evidence. A unique evidence number is generated for each piece of input coverage evidence, and it is associated with the corresponding candidate collaborative object number, evidence fragment text, evidence source, evidence type, evidence keyword set, evidence time, generating agent identifier, and evidence confidence field.

[0148] In existing technologies, a unified keyword search is usually performed on all context fragments without distinguishing between user-expressed content and external prompts. This unified search leads to mixed search results, making it impossible to clearly distinguish which information has been actively provided by the user and which needs to be actively supplemented by the system. In multi-source interaction scenarios, such mixed search results increase the difficulty of subsequent evidence screening and are prone to generating redundant evidence and missing key external information.

[0149] In this embodiment, the search direction is divided into two independent dimensions: the expression coverage search direction and the external prompt search direction. The expression coverage search direction is specifically for the user's current input content and accurately locates the information that the user has clearly expressed. This division enables targeted retrieval of information from different sources, greatly improving the accuracy of search results and reducing the generation of redundant evidence.

[0150] The semantic deduplication similarity threshold is determined by labeled samples. For example, in this application embodiment, the semantic deduplication similarity threshold is preferably 0.9.

[0151] For example, the current input fragment of task package T001 is "I tend to wake up easily at night recently and feel a bit listless during the day". After word segmentation, the keyword set {recently, night, easily wake up, daytime, listless} is obtained, generating a fragment vector with a dimension of 768. The recording time is 10:30 on June 3, 2026. The generated expression covers the search direction, and the search yields two pieces of input coverage evidence, with evidence numbers E001 and E002, and evidence confidence levels of 0.92 and 0.88, respectively.

[0152] Step S3022: Based on the historical record fragments and collected data fragments in the source verification task package, construct the external source hint retrieval direction and obtain external source hint evidence.

[0153] In this embodiment, based on the historical record fragment set and the collected data fragment set in the source verification task package, word segmentation and part-of-speech filtering are performed respectively to generate a historical record keyword set and a collected data keyword set. The union of the historical record keyword set and the collected data keyword set is calculated to obtain the external source keyword set.

[0154] The difference between the external keyword set and the current input keyword set is calculated to obtain the external prompt keyword set. Collaborative topic descriptions are extracted, and a pre-trained text vector model is called to convert each historical record fragment and collected data fragment into vector form, generating a set of historical record fragment vectors and a set of collected data fragment vectors. The recording time information of each historical record fragment and collected data fragment is extracted to generate external fragment time information.

[0155] The system integrates collaborative topic descriptions, external source hint keyword sets, historical record fragment vector sets, collected data fragment vector sets, and external source fragment time information to generate external source hint retrieval directions. It then calls upon the external source verification agent in the source verification collaborative agent group to retrieve content that was hinted at in the historical records or collected data but not fully covered by the current input based on the external source hint retrieval direction. The retrieved content undergoes semantic deduplication and relevance verification to generate external source hint evidence. A unique evidence number is generated for each external source hint evidence, and it is associated with the corresponding candidate collaborative object number, evidence fragment text, evidence source, evidence type, evidence keyword set, evidence time, generating agent identifier, and evidence confidence field.

[0156] In existing technologies, external information retrieval typically uses the entire keyword set of historical records and collected data directly, without excluding keywords already included in the user's current input. This approach leads to a large number of duplicate searches, and the search results contain a lot of information that the user has already explicitly expressed, wasting computing resources. At the same time, it cannot accurately locate key hints that the user has not mentioned but that exist in the external data.

[0157] In this embodiment, a special set of external prompt keywords is obtained by calculating the difference between the set of external keywords and the set of currently input keywords. Based on this set, an external prompt retrieval direction is constructed, which can accurately retrieve content that exists in historical records and collected data but is not fully covered by the user's current input. This improvement greatly enhances the targeting of external information retrieval, reduces redundant calculations, and ensures that no key external prompt information is missed.

[0158] The keyword overlap threshold for relevance verification is determined by labeled samples. For example, in this application embodiment, the relevance verification overlap threshold is preferably 0.3.

[0159] For example, the historical keyword set of task package T001 is {sleep, poor, adjust, rest}, and the keyword set of collected data is {night, activity, frequency, increase, heart rate, speed up}. The union of these keywords yields the external keyword set {sleep, poor, adjust, rest, night, activity, frequency, increase, heart rate, speed up}. The difference between this set and the current input keyword set yields the external prompt keyword set {sleep, poor, adjust, rest, night, activity, frequency, increase, heart rate, speed up}. An external prompt retrieval direction is generated, and three external prompt evidences are retrieved, with evidence numbers E003, E004, and E005, and evidence confidence levels of 0.85, 0.90, and 0.78, respectively.

[0160] Step S3023: Based on the input coverage evidence and the external cue evidence, perform an inter-source consistency comparison, and identify the external cue evidence that is inconsistent with the input coverage evidence in terms of candidate collaborative themes, fragment elements, or recording time as deviation evidence.

[0161] In this embodiment, consistency comparison is performed based on input coverage evidence and external cue evidence from three dimensions: candidate co-occurring topics, fragment elements, and recording time. In the candidate co-occurring topic dimension, a pre-trained text vector model is called to generate topic vectors for input coverage evidence and external cue evidence. The cosine similarity between the two topic vectors is calculated to obtain the topic similarity.

[0162] In terms of fragment element dimension, the keyword sets of input coverage evidence and external cue evidence are extracted. The number of elements in the intersection of the two keyword sets is calculated and divided by the number of elements in the union of the two keyword sets to obtain the keyword overlap. In terms of recording time dimension, the number of days between the recording time of input coverage evidence and external cue evidence is calculated to determine the time consistency threshold. The time consistency threshold is determined by labeled samples. For example, in the embodiments of this application, the preferred time consistency threshold when there are no labeled samples is 7 days.

[0163] If the topic similarity is less than the preset topic consistency threshold, or the keyword overlap is less than the preset element consistency threshold, or the time interval in days is greater than the time consistency threshold, the corresponding external source hint evidence will be marked as deviation evidence, and the type field of the evidence will be updated to deviation evidence.

[0164] The topic consistency threshold is determined by maximizing the composite value of labeled samples. Multiple manually labeled topic consistency samples are prepared. Each sample contains a pair of input coverage evidence and external cue evidence, labeled as either topic consistent or topic inconsistent. Candidate thresholds are iterated from 0.1 to 0.9, incrementing by 0.01. For each candidate threshold, precision and recall are calculated. Precision is the ratio of the number of correctly identified topic consistent samples to the total number of samples correctly identified as topic consistent by the system. Recall is the ratio of the number of correctly identified topic consistent samples to the total number of manually labeled topic consistent samples. A composite value is calculated for each candidate threshold. This composite value is used to comprehensively measure precision and recall. The calculation method is as follows: multiply the product of precision and recall by two to obtain the numerator, use the sum of precision and recall as the denominator, divide the numerator by the denominator to obtain the composite value, and select the threshold with the largest composite value as the topic consistency threshold.

[0165] For example, in the embodiments of this application, the topic consistency threshold is preferably 0.6 to 0.8 when there are no labeled samples, with the optimal value being 0.7.

[0166] The feature consistency threshold is determined by maximizing the composite value of labeled samples. Multiple manually labeled feature consistency samples are prepared. Each sample contains a pair of input coverage evidence and external cue evidence, labeled as either core feature consistent or core feature inconsistent. Candidate thresholds are iterated from 0.1 to 0.9, incrementing by 0.01. For each candidate threshold, precision and recall are calculated. Precision is the ratio of the number of correctly identified features consistent to the total number of system-identified features consistent. Recall is the ratio of the number of correctly identified features consistent to the total number of manually labeled features consistent. A composite value is calculated for each candidate threshold. This composite value is used to comprehensively measure precision and recall. It is calculated by multiplying the product of precision and recall by two as the numerator, using the sum of precision and recall as the denominator, and dividing the numerator by the denominator. The threshold with the largest composite value is selected as the feature consistency threshold.

[0167] For example, in the embodiments of this application, the element consistency threshold is preferably 0.4 to 0.6 when there are no labeled samples, with the optimal value being 0.5.

[0168] In existing technologies, consistency comparison between sources usually only uses a single semantic similarity dimension. Single-dimensional comparison cannot comprehensively assess the consistency between evidence and is prone to misjudgment. For example, evidence that is semantically similar but has a long time interval between records, or evidence that is thematically similar but has completely different core elements, may be wrongly judged as consistent evidence. These wrongly judged pieces of evidence will affect the accuracy of subsequent trigger item generation.

[0169] In this embodiment, a comprehensive consistency comparison is performed from three dimensions: candidate collaborative themes, fragment elements, and recording time. These three dimensions evaluate the consistency of evidence from the perspectives of macro-themes, micro-elements, and time validity, respectively. This multi-dimensional comprehensive comparison can significantly reduce the false judgment rate, accurately identify external evidence with deviations, and provide a reliable basis for subsequent deviation relationship determination.

[0170] The comparison logic for the three dimensions is determined by labeled samples. For example, the comparison logic in this application is preferably OR logic, and any dimension that does not meet the consistency requirements is judged as deviation evidence.

[0171] For example, the topic similarity between input coverage evidence E001 and external prompt evidence E005 is 0.65, which is less than the topic consistency threshold of 0.7. The keyword overlap is 0.4, which is less than the element consistency threshold of 0.5. The time interval is 15 days, which is greater than the time consistency threshold of 7 days. Since all three dimensions do not meet the consistency requirements, E005 is marked as deviation evidence and the evidence type is updated to deviation evidence.

[0172] Step S3024: Based on input coverage evidence, external cue evidence, and deviation evidence, form a candidate evidence set.

[0173] In this embodiment, all generated input coverage evidence, external prompt evidence not marked as deviation evidence, and deviation evidence are integrated. Field integrity is checked on all evidence, missing basic information is supplemented, preliminary deduplication is performed on the evidence, duplicate evidence with identical content and the same source is removed, and all evidence that meets the requirements is integrated into a candidate evidence set.

[0174] The initial deduplication similarity threshold is determined by labeled samples. For example, in this application embodiment, the initial deduplication similarity threshold is preferably 0.95.

[0175] For example, the candidate evidence set of task package T001 contains five pieces of evidence: E001, E002, E003, E004 and E005. Among them, E001 and E002 are input coverage evidence, E003 and E004 are external prompt evidence, and E005 is deviation evidence. All evidence fields are complete and there is no duplication.

[0176] Step S303: Based on the candidate evidence set, call the source verification collaborative intelligent agent group to perform intra-group cross-verification and divergence backtracking screening to obtain the effective evidence set.

[0177] The specific steps of step S303 are as follows:

[0178] Step S3031: Based on the candidate evidence set, call the source verification collaborative intelligent agent group to perform multi-dimensional cross-verification of each candidate evidence to obtain the evidence verification result set.

[0179] In this embodiment, based on the candidate evidence set, all candidate evidence is traversed in order of evidence number. All agents in the source verification collaborative agent group are called to cross-verify each candidate evidence. The verification content includes whether the evidence corresponds to the candidate collaborative object, whether the evidence type is correct, whether the evidence is duplicated, whether the evidence deviates from other sources, and whether the evidence is suitable as the basis for triggering subsequent interactions.

[0180] Each agent scores each verification dimension according to a preset scoring standard. The scores for each dimension are then weighted and averaged to obtain an individual verification score between 0 and 1. The evidence verification score for each piece of evidence is calculated as the sum of all individual verification scores divided by the number of verifying agents. The average verification score for each piece of evidence is then calculated and equal to the total evidence verification score. The scoring standard employs a four-level quantitative scoring system, with independent judgment rules established for each of the four core dimensions: semantic matching, source credibility, time validity, and topic relevance. Under the semantic matching dimension, the semantics of the evidence and the candidate collaborating pair... A score of 1 is awarded when the evidence is completely identical, unambiguous, and complete, with both core and auxiliary information accurately matched. A score of 0.7 is awarded when there are only minor deviations in non-core information that do not affect the understanding of the core semantics. A score of 0.4 is awarded when some information is missing or biased, but the core information can be inferred from the context. A score of 0 is awarded when the semantics are completely mismatched, the content is unrecognizable, or there is serious ambiguity preventing the extraction of effective information. In terms of source credibility, a score of 1 is awarded when the evidence is automatically collected by officially certified equipment, the data transmission process is encrypted, and the original data is complete, traceable, and free from tampering risks. A score of 1 is awarded when the evidence is manually entered by the user, the information is complete, and the data is consistent with the original data. 0.7 points are awarded when one less piece of data from another source is available for cross-verification; 0.4 points are awarded when the data is authorized by a third party and has undergone basic platform verification but has minor flaws in completeness; and 0 points are awarded when the source is unknown, the data shows obvious signs of tampering, or its authenticity cannot be verified in any way. In terms of time validity, 1 point is awarded when the evidence recording time is within the preset valid time window and is completely consistent with the time logic of other related evidence; 0.7 points are awarded when the recording time exceeds the valid time window but does not exceed twice the window length and the core information has not changed significantly; and 0.7 points are awarded when the recording time exceeds twice the valid time window but is still valid. 0.4 points are awarded when the evidence still has reference value as verified by the latest related information; 0 points are awarded when the record time is completely invalid or there is an irreconcilable contradiction with the time logic of other related evidence. In terms of topic relevance, 1 point is awarded when the evidence topic is completely consistent with the candidate collaborative topic and all content revolves around the topic without any irrelevant information; 0.7 points are awarded when the topics are basically consistent and contain only a small amount of marginal irrelevant information that does not affect the core judgment; 0.4 points are awarded when the topic is partially related and the core content is somewhat related to the topic but contains a lot of irrelevant information; and 0 points are awarded when the topic is completely irrelevant and all content does not involve the topic.

[0181] Calculate the difference between the verification score of each individual entity and the average verification score, square the difference, sum all the squared values, divide by the number of verification agents, and take the square root of the result to obtain the verification divergence degree. Integrate the evidence verification score and verification divergence degree of each piece of evidence into the evidence verification result, and all evidence verification results form the evidence verification result set.

[0182] The number of agents in the source-based verification collaborative agent group is determined according to the application scenario. For example, in this application embodiment, the number of agents is preferably 4, which are verified from four dimensions: semantics, source, time, and topic.

[0183] For example, cross-validation is performed on evidence E003. The individual validation scores given by the four agents are 0.90, 0.85, 0.95 and 0.80, respectively. The calculated evidence validation score is 0.875, the average validation score is 0.875, the sum of squares of the differences is 0.0125, the validation divergence is 0.0559, and the corresponding evidence validation result is generated.

[0184] Step S3032: Based on the evidence verification result set, identify the matching discrepancies between the same candidate evidence and the candidate collaborative object and the source verification task package, and obtain the discrepancy evidence result set.

[0185] In this embodiment, a disagreement threshold is determined based on the evidence verification result set. The disagreement threshold is determined by labeled samples. For example, in this embodiment of the application, the disagreement threshold is preferably 0.20 when there are no labeled samples. The verification results of each piece of evidence are traversed. When the degree of disagreement of the evidence is greater than or equal to the disagreement threshold, it is determined that the evidence has a matching disagreement.

[0186] Extract candidate collaborative object information and source verification task package information corresponding to evidence with matching discrepancies, compare the individual verification scores and verification opinions of each agent, analyze the specific dimensions of the discrepancies, including topic matching discrepancies, evidence type discrepancies, source credibility discrepancies, and time correlation discrepancies, and integrate all evidence with matching discrepancies and their discrepancy dimension information, as well as the verification opinions of each agent, into a discrepancy evidence result set.

[0187] The criteria for determining the dimension of disagreement are determined by labeled samples. For example, in this application embodiment, the preferred criterion for determining the dimension of disagreement is that when the difference between the individual verification score and the average score of a certain dimension is greater than or equal to 0.3, it is determined that there is disagreement in that dimension.

[0188] For example, the verification disagreement degree of evidence E004 is 0.25, which is greater than the disagreement judgment threshold of 0.20. It is determined that there is a matching disagreement. Comparing the scores of each agent, it is found that the score given by the agent in the topic dimension is 0.5, which is 0.3 different from the average score of 0.8. It is determined that there is a topic matching disagreement, and the corresponding disagreement evidence result is generated.

[0189] Step S3033: Based on the divergent evidence result set, backtrack the corresponding source verification task package, and correct or remove the candidate evidence with divergence to obtain the corrected evidence result set.

[0190] In this embodiment, based on the set of divergent evidence results, all divergent evidence is traversed in order of evidence number, and the original source verification task package corresponding to each divergent evidence is traced back to extract the basis for the generation of the evidence, the original fragment content, and the detailed verification opinions of each agent.

[0191] For disagreements in different dimensions, corresponding correction methods are adopted. The topic matching disagreement is re-invoked to re-invoke the topic perspective agent and perform secondary topic matching based on the original fragment content. The secondary topic matching adopts a multi-model fusion architecture, integrating three pre-trained text models with different network structures, namely BERT-base model, RoBERTa-wwm-ext model and ALBERT-base-v2 model.

[0192] The BERT-base model employs a 12-layer Transformer encoder structure with 768 hidden layers, 12 attention heads, and the GELU activation function. The RoBERTa-wwm-ext model also employs a 12-layer Transformer encoder structure with 768 hidden layers, 12 attention heads, and a full-word masking strategy. The ALBERT-base-v2 model employs a 12-layer Transformer encoder structure with 768 hidden layers, 12 attention heads, and utilizes parameter sharing and embedding layer factorization techniques.

[0193] All three models were pre-trained on a general Chinese corpus and then fine-tuned for unified domain adaptation in the multi-source interaction scenarios of this solution. The training dataset contains four types of data: general dialogue corpus, intelligent customer service interaction records, health consultation text, and device log descriptions, with a total sample size of 2 million. The fine-tuning process uses the cross-entropy loss function, with a batch size of 32, a learning rate of 2e-5, and 3 training epochs. After fine-tuning, the three models extract text features from different semantic dimensions, forming complementary topic matching capabilities.

[0194] The multi-model fusion employs a weighted average method, with the weights of each model determined using the analytic hierarchy process (AHP). For example, in this embodiment, the preferred weights are 0.40 for the BERT-base model, 0.35 for the RoBERTa-wwm-ext model, and 0.25 for the ALBERT-base-v2 model. The fusion confidence score is the sum of the topic matching scores output by the three models multiplied by their respective weights.

[0195] The fusion confidence threshold is determined through labeled samples. 100,000 manually labeled topic matching samples are prepared. Candidate thresholds are iterated from 0.1 to 0.9, incrementing by 0.01. For each candidate threshold, precision and recall are calculated. Precision is the ratio of correct matches to system matches, and recall is the ratio of correct matches to manually labeled matches. A comprehensive value is calculated for each candidate threshold. This comprehensive value is used to comprehensively measure precision and recall. The calculation method is as follows: multiply the product of precision and recall by two to obtain the numerator, use the sum of precision and recall as the denominator, and divide the numerator by the denominator to obtain the comprehensive value. The threshold with the largest comprehensive value is selected as the fusion confidence threshold. For example, in this embodiment, the preferred fusion confidence threshold is 0.7.

[0196] Disagreements regarding evidence type necessitate a re-evaluation of the evidence type, combining information from multiple sources to determine its classification. Disagreements regarding source credibility require a reassessment of the reliability level of the data source, adjusting the source credibility value based on the data collection method and transmission process. Disagreements regarding temporal correlation necessitate a recalculation of temporal relevance, correcting the calculation results for time intervals and time validity.

[0197] For edge deployment scenarios, a lightweight modification is made to the multi-model fusion inference system. A joint knowledge distillation method is adopted, using three finely tuned large models as joint teacher models to train a unified student model with a 6-layer Transformer structure. After training, the student model is dynamically quantized with INT8 to convert the model parameters from 32-bit floating-point numbers to 8-bit integers. At the same time, operator fusion and graph optimization techniques are used to merge redundant computing nodes.

[0198] The confidence level of the revised evidence is recalculated. Evidence that cannot be revised or whose revised confidence level is still lower than the preset revised confidence threshold is removed. The revised confidence threshold is determined by maximizing the composite value of the labeled samples. Multiple manually labeled samples of the validity of the revised evidence are prepared. Each sample contains the revised evidence and its corresponding confidence level, labeled as valid or invalid evidence. The candidate thresholds are iterated from 0.1 to 0.9 in increments of 0.01. For each candidate threshold, precision and recall are calculated. Precision is the ratio of the number of correctly identified valid evidence to the total number of evidence identified as valid by the system. Recall is... The ratio of the number of correctly identified valid evidence to the total number of manually labeled valid evidence is used to calculate the comprehensive value corresponding to each candidate threshold. The comprehensive value is used to comprehensively measure precision and recall. The calculation method is as follows: multiply the product of precision and recall by two as the numerator, use the sum of precision and recall as the denominator, divide the numerator by the denominator to obtain the comprehensive value, and select the threshold with the largest comprehensive value as the corrected confidence threshold. For example, in the embodiments of this application, the corrected confidence threshold is preferably 0.4 to 0.6 when there are no labeled samples, and the optimal value is 0.5. All corrected valid evidence is integrated into a corrected evidence result set.

[0199] In existing technologies, evidence with high degree of disagreement during cross-validation is usually directly removed. Direct removal leads to the loss of a large amount of potentially valid information, especially in scenarios where there is noise in multi-source data. Many disagreements are caused by misjudgments by a single agent. Direct removal of this evidence will result in an incomplete evidence set, affecting the comprehensiveness of subsequent trigger item generation.

[0200] This embodiment does not directly discard conflicting evidence. Instead, it first backtracks to the original task package, analyzes the specific dimensions that caused the conflict, and then applies corresponding correction methods for the conflicts in different dimensions. This conflict backtracking and correction mechanism can salvage a large amount of conflicting evidence caused by misjudgment by a single agent, significantly improving the completeness of the evidence set. At the same time, through a secondary topic matching method using multi-model fusion, it can effectively improve the accuracy of topic matching and reduce the occurrence of misjudgments.

[0201] Step S3034: Based on the candidate evidence set and the revised evidence result set, merge duplicate evidence and screen out low-confidence evidence to obtain the effective evidence set.

[0202] In this embodiment, based on the candidate evidence set and the revised evidence result set, the original divergent evidence corresponding to the candidate evidence set is replaced with the evidence in the revised evidence result set to identify duplicate evidence. The criteria for determining duplicate evidence are that they have the same evidence fragment text, the same evidence source, and the same evidence time. Duplicate evidence is merged, and the evidence with the highest confidence level is retained.

[0203] The initial screening threshold for evidence is determined by maximizing the composite value of labeled samples. Historical samples are prepared, and each sample is manually labeled with correct results. Candidate thresholds are iterated from 0.50 to 0.90 in increments of 0.01. For each candidate threshold, precision and recall are calculated. Precision is the ratio of the number of correctly retained samples to the number retained by the system, and recall is the ratio of the number of correctly retained samples to the number manually labeled as the number to be retained. The composite value corresponding to each candidate threshold is calculated. The composite value is used to comprehensively measure precision and recall. It is calculated by multiplying the product of precision and recall by two as the numerator, using the sum of precision and recall as the denominator, and dividing the numerator by the denominator to obtain the composite value. The threshold with the largest composite value is selected as the initial screening threshold for evidence.

[0204] Low-confidence evidence with a confidence level below the initial screening threshold is removed. The remaining evidence undergoes field integrity verification and format standardization. All compliant evidence is then integrated into a valid evidence set.

[0205] The information inheritance rules for merging duplicate evidence are determined by labeled samples. For example, in the embodiments of this application, the preferred information inheritance rule is that the retained evidence inherits the source information and time information of all duplicate evidence.

[0206] For example, the original evidence E004 was replaced with the corrected evidence E004'. No duplicate evidence was found, and the confidence level of all evidence was greater than 0.62. The resulting valid evidence set contained five pieces of evidence: E001, E002, E003, E004', and E005.

[0207] like Figure 2 As shown, step S304: Based on the effective evidence set, construct a multi-source collaborative differential evidence graph, and perform coverage differential comparison and collaborative trigger item merging to obtain a key collaborative trigger item set.

[0208] The specific steps of step S304 are as follows:

[0209] Step S3041: Based on the valid evidence set and the candidate collaborative object set, perform node mapping on the candidate collaborative objects and their corresponding input coverage evidence, external hint evidence and deviation evidence to obtain the collaborative object node set and the evidence node set.

[0210] In this embodiment, based on the candidate collaborative object set, each candidate collaborative object is mapped to a candidate collaborative object node. The node attributes include the candidate collaborative object number, collaborative topic description, and object confidence. Based on the valid evidence set, the evidence is classified according to the evidence type. Each input coverage evidence is mapped to a supporting evidence node, each external prompt evidence is mapped to an external prompt evidence node, and each deviation evidence is mapped to a deviation evidence node.

[0211] Each evidence node's attributes include evidence number, candidate collaborating object number, evidence fragment text, evidence source, evidence type, evidence confidence level, and evidence validity value. The evidence validity value is calculated by weighting the evidence confidence level, evidence verification score, and verification disagreement degree. Each weight coefficient is determined by the analytic hierarchy process (AHP). For example, in this application embodiment, the evidence confidence level weight is preferably 0.50, the evidence verification score weight is preferably 0.30, and the verification disagreement degree weight is preferably 0.20. All candidate collaborating object nodes are integrated into a collaborating object node set, and all evidence nodes are integrated into an evidence node set.

[0212] The correspondence of node mapping is determined by labeled samples. For example, in the embodiments of this application, the node mapping preferably adopts a one-to-one correspondence method, with each candidate collaborative object and each piece of evidence being mapped to an independent node.

[0213] For example, candidate collaborative object 001 is mapped to node N001 with the attribute {001, abnormal sleep quality, 0.85}, which comes from the exemplary output of step S301.

[0214] Cross-validation of evidence E001 yielded individual validation scores of 0.92, 0.88, 0.90, and 0.90 from the four agents, respectively.

[0215] The sum of the four individual verification scores is calculated to obtain 3.60. The sum is divided by the number of agents, 4, to obtain an evidence verification score of 0.90. The differences between the individual verification scores and the average verification score are calculated to be 0.02, -0.02, 0.00, and 0.00, respectively.

[0216] Squaring each difference yields 0.0004, 0.0004, 0.0000, and 0.0000 respectively. Summing all squared values ​​yields 0.0008. Dividing the sum of squares by the number of agents (4) yields 0.0002. Taking the square root of the result gives a verification divergence of 0.0141.

[0217] To calculate the effective value of the evidence, first calculate the product of 0.5 and 0.92, which gives 0.46. Then calculate the product of 0.3 and 0.90, which gives 0.27. Next, calculate the difference between 1 and 0.0141, which gives 0.9859. Finally, calculate the product of 0.2 and 0.9859, which gives 0.1972. Add the three products together to get 0.9272.

[0218] Evidence E001 is mapped to supporting evidence node N002. Evidence E002 is cross-verified. The individual verification scores given by the four agents are 0.88, 0.85, 0.90 and 0.87 respectively. The sum of the four individual verification scores is calculated to get 3.50. The sum is divided by the number of agents 4 to get the evidence verification score of 0.875.

[0219] The differences between the individual verification scores and the average verification scores are calculated as 0.005, -0.025, 0.025, and -0.005, respectively. The squares of each difference are calculated as 0.000025, 0.000625, 0.000625, and 0.000025, respectively. The sum of all squares is calculated as 0.0013. The sum of squares is divided by the number of agents, 4, to obtain 0.000325. The square root of the result is taken to obtain the verification divergence degree as 0.0180.

[0220] To calculate the effective value of the evidence, first calculate the product of 0.5 and 0.88, which gives 0.44. Then calculate the product of 0.3 and 0.875, which gives 0.2625. Next, calculate the difference between 1 and 0.0180, which gives 0.9820. Finally, calculate the product of 0.2 and 0.9820, which gives 0.1964. Add the three products together to get 0.8989.

[0221] Evidence E002 is mapped to supporting evidence node N003. Evidence E003 is cross-verified. The individual verification scores given by the four agents are 0.90, 0.85, 0.95 and 0.80 respectively. The sum of the four individual verification scores is calculated to get 3.50. The sum is divided by the number of agents 4 to get the evidence verification score of 0.875.

[0222] The differences between the individual verification scores and the average verification score are calculated as 0.025, -0.025, 0.075, and -0.075, respectively. The squares of each difference are calculated as 0.000625, 0.000625, 0.005625, and 0.005625, respectively. The sum of all squares is calculated as 0.0125. The sum of squares is divided by the number of agents, 4, to obtain 0.003125. The square root of the result is taken to obtain the verification divergence degree as 0.0559.

[0223] To calculate the validity value of the evidence, first calculate the product of 0.5 and 0.85, which gives 0.425. Then calculate the product of 0.3 and 0.875, which gives 0.2625. Next, calculate the difference between 1 and 0.0559, which gives 0.9441. Finally, calculate the product of 0.2 and 0.9441, which gives 0.1888. Add the three products together to get 0.8763. Map evidence E003 to external hint evidence node N004. Perform cross-verification on evidence E004'. The individual verification scores given by the four agents are 0.85, 0.80, 0.85, and 0.82, respectively. Calculate the sum of the four individual verification scores, which gives 3.32. Divide the sum by the number of agents, 4, to get the evidence verification score of 0.83.

[0224] The differences between the individual verification scores and the average verification scores are calculated as 0.02, -0.03, 0.02, and -0.01, respectively. The squares of each difference are calculated as 0.0004, 0.0009, 0.0004, and 0.0001, respectively. The sum of all squares is calculated as 0.0018. The sum of squares is divided by the number of agents, 4, to obtain 0.00045. The square root of the result is then calculated to obtain the verification divergence degree as 0.0212.

[0225] To calculate the valid value of the evidence, first calculate the product of 0.5 and 0.82, which gives 0.41. Then calculate the product of 0.3 and 0.83, which gives 0.249. Next, calculate the difference between 1 and 0.0212, which gives 0.9788. Finally, calculate the product of 0.2 and 0.9788, which gives 0.1958. Add the three products together to get 0.8548. Map evidence E004' to external hint evidence node N005.

[0226] Cross-verification of evidence E005 was performed, and the individual verification scores given by the four agents were 0.75, 0.70, 0.80, and 0.75, respectively. The sum of the four individual verification scores was calculated to obtain 3.00. The sum was divided by the number of agents, 4, to obtain the evidence verification score of 0.75.

[0227] The differences between the individual verification scores and the average verification scores are calculated as 0.00, -0.05, 0.05, and 0.00, respectively. The squares of each difference are calculated as 0.0000, 0.0025, 0.0025, and 0.0000, respectively. The sum of all squares is calculated as 0.0050. The sum of squares is divided by the number of agents, 4, to obtain 0.00125. The square root of the result is then calculated to obtain the verification divergence degree as 0.0354.

[0228] To calculate the effective value of the evidence, first calculate the product of 0.5 and 0.78, which gives 0.39. Then calculate the product of 0.3 and 0.75, which gives 0.225. Next, calculate the difference between 1 and 0.0354, which gives 0.9646. Finally, calculate the product of 0.2 and 0.9646, which gives 0.1929. Add the three products together to get 0.8079.

[0229] Map evidence E005 to the off-reference evidence node N006.

[0230] Obtain the set of collaborative object nodes and the set of evidence nodes.

[0231] Step S3042: Based on the collaborative object node set and the evidence node set, establish evidence association edges, coverage comparison edges and deviation indication edges to construct a multi-source collaborative differential evidence graph.

[0232] In this embodiment, based on the set of collaborative object nodes and the set of evidence nodes, all candidate collaborative object nodes are traversed, all supporting evidence nodes associated with each candidate collaborative object node are extracted, and supporting edges are established between the candidate collaborative object nodes and the corresponding supporting evidence nodes. The weight of the supporting edge is equal to the evidence validity value of the corresponding supporting evidence.

[0233] Extract all external cue evidence nodes associated with each candidate collaborative object node. Establish cue edges between the candidate collaborative object node and the corresponding external cue evidence nodes. The weight of the cue edge is equal to the evidence validity value of the corresponding external cue evidence multiplied by the source credibility of the evidence. Traverse all supporting evidence nodes and extract all external cue evidence nodes belonging to the same candidate collaborative object as the supporting evidence node. Establish coverage comparison edges between each supporting evidence node and all corresponding external cue evidence nodes. Coverage comparison edges are used to calculate the coverage of the current input with the external cue content.

[0234] Traverse all deviation evidence nodes, extract supporting evidence nodes and external hint evidence nodes associated with each deviation evidence node, establish deviation edges between each deviation evidence node and its corresponding supporting evidence node and external hint evidence node, and use deviation edges to represent the degree of inconsistency between different sources. Integrate all nodes and edges to construct a multi-source collaborative differential evidence graph.

[0235] The multi-source collaborative differential evidence graph is stored using a node table, an edge table, and an adjacency index table. The node table includes fields for node number, node type, candidate collaborative object number, evidence number, evidence source, evidence type, evidence confidence level, evidence validity value, record time, and node status. Node types include collaborative object nodes, input coverage evidence nodes, external source cue evidence nodes, and deviation evidence nodes. The edge table includes fields for edge number, starting node number, ending node number, edge type, edge weight, candidate collaborative object number, generation time, and edge status. Edge types include evidence association edges, coverage comparison edges, and deviation indication edges. The adjacency index table includes node number, adjacent node number, edge number, edge type, and candidate collaborative object number.

[0236] When querying the multi-source collaborative differential evidence graph, the corresponding collaborative object node is located in the node table based on the candidate collaborative object number. Then, the input coverage evidence nodes, external source hint evidence nodes, and deviation evidence nodes under that collaborative object node are read based on the adjacency index table. During coverage differential alignment, the coverage alignment edge is queried in the adjacency index table based on the external source hint evidence node number to obtain the corresponding set of input coverage evidence nodes. During deviation relation queries, the deviation indicator edge is queried in the adjacency index table based on the external source hint evidence node number to obtain the corresponding set of deviation evidence nodes. A joint index is established using the candidate collaborative object number and edge type, enabling direct location of evidence nodes and relation edges under the same collaborative object, avoiding traversal of all evidence nodes.

[0237] For example, a support edge with a weight of 0.904 is established between nodes N001 and N002, a hint edge with a weight of 0.85 × 0.9 = 0.765 is established between node N001 and external hint evidence node N004, a coverage comparison edge is established between nodes N002 and N004, and a deviation edge is established between node N001 and deviation evidence node N006. All nodes and edges are integrated to construct a multi-source collaborative differential evidence graph.

[0238] Step S3043: Based on the multi-source collaborative differential evidence map, extract the input coverage evidence nodes, external prompt evidence nodes, and deviation evidence nodes under the same collaborative object node to obtain the coverage comparison evidence group.

[0239] In this embodiment, based on the multi-source collaborative differential evidence graph, all candidate collaborative object nodes are traversed in order of candidate collaborative object number. All supporting evidence nodes associated with each candidate collaborative object node through supporting edges are extracted, all external cue evidence nodes associated with each candidate collaborative object node through cue edges are extracted, and all deviation evidence nodes associated with each candidate collaborative object node through deviation edges are extracted.

[0240] The three types of evidence nodes under the same candidate collaborative object node are grouped, and each group corresponds to a coverage comparison evidence group. A unique group number is generated for each coverage comparison evidence group, which is associated with the corresponding candidate collaborative object number. All coverage comparison evidence groups are integrated into a coverage comparison evidence group set.

[0241] The method for generating the coverage comparison evidence set is determined based on the graph structure. For example, the preferred generation method in this application embodiment is a depth-first traversal method, which traverses all evidence nodes directly connected to the candidate collaborative object node.

[0242] For example, traversing node N001, extracting 2 supporting evidence nodes, 2 external hint evidence nodes, and 1 deviation evidence node, generating a coverage comparison evidence group with the number G001, and associating it with candidate collaborative object number 001.

[0243] Step S3044: Based on the coverage comparison evidence group, perform coverage difference comparison on the input coverage evidence nodes, external prompt evidence nodes and deviation evidence nodes to obtain the coverage difference relationship record set.

[0244] The specific steps of step S3044 are as follows:

[0245] Step S30441: Based on the coverage comparison evidence group, determine the set of input coverage evidence nodes and the set of deviation evidence nodes corresponding to each external prompt evidence node under the same collaborative object node.

[0246] In this embodiment, based on the coverage comparison evidence group, all coverage comparison evidence groups are traversed in order of group number. All external prompt evidence nodes in each coverage comparison evidence group are traversed. All supporting evidence nodes associated with the external prompt evidence node are extracted through the coverage comparison edge to form the input coverage evidence node set corresponding to the external prompt evidence node.

[0247] By extracting all the deviation evidence nodes that are associated with the external source hint evidence node through the deviation edge, a set of deviation evidence nodes corresponding to the external source hint evidence node is formed. For each external source hint evidence node, the corresponding set of input overlay evidence nodes and set of deviation evidence nodes are recorded.

[0248] The maximum capacity of the input coverage evidence node set is determined based on computing resources. For example, in this application embodiment, the maximum capacity is preferably 10. When the capacity is exceeded, only the 10 nodes with the highest valid evidence values ​​are retained.

[0249] For example, the external prompt evidence node N004 in the coverage comparison evidence group G001 is traversed, and two supporting evidence nodes N002 and N003 are extracted through the coverage comparison edge to form the input coverage evidence node set. No relevant deviation evidence nodes are extracted through the deviation edge, and the deviation evidence node set is empty.

[0250] Step S30442: Based on the coverage comparison edge between the external prompt evidence node and the set of input coverage evidence nodes, determine the coverage degree of each external prompt evidence node relative to the input coverage evidence.

[0251] In this embodiment, based on the coverage comparison edge between the external prompt evidence node and the set of input coverage evidence nodes, the set of input coverage evidence nodes corresponding to each external prompt evidence node is traversed, and a pre-trained text vector model is invoked to generate the fragment vectors of the external prompt evidence node and each input coverage evidence node.

[0252] The pre-trained text vector model uses the BERT-base-Chinese model with a 12-layer Transformer encoder structure, a hidden layer dimension of 768, 12 attention heads, a feedforward neural network dimension of 3072, the GELU activation function, and a maximum sequence length of 128.

[0253] To address the semantic features of the multi-source interactive text in this solution, the pre-trained model is fine-tuned for domain adaptation. The training dataset includes five types of data: general dialogue corpus, intelligent customer service interaction records, health consultation text, device log descriptions, and user feedback text, with a total sample size of 2.5 million. The fine-tuning process adopts a contrastive learning framework, constructing positive and negative sample pairs. Positive sample pairs are different expressions of text with the same semantics, while negative sample pairs are texts with different semantics. The batch size is set to 32, the learning rate is set to 2e-5, the number of training rounds is set to 4, and the contrastive loss function is used.

[0254] The fragment vector generation uses the mean pooling method to segment the input text, generate a token sequence, add and label it, truncate sequences longer than 126, and pad sequences shorter than 126. The processed token sequence is input into a pre-trained text vector model to obtain the output of the last hidden layer. The mean of the hidden layer vectors of all tokens is calculated to obtain a fragment vector with a dimension of 768.

[0255] Calculate the cosine similarity between two fragment vectors to obtain the similarity between the external cue evidence node and the input overlay evidence node. The cosine similarity is calculated by dividing the dot product of the two vectors by the product of the magnitudes of the two vectors. The value ranges from -1 to 1, and the larger the value, the higher the semantic similarity.

[0256] Traverse all nodes in the input coverage evidence node set, calculate all similarity values, and take the maximum value among all similarities as the coverage of the external prompt evidence node. When the input coverage evidence node set is empty, the coverage of the external prompt evidence node is 0.

[0257] For edge deployment scenarios, the pre-trained text vector model is lightweighted and a task-oriented knowledge distillation method is adopted. The fine-tuned BERT-base-Chinese model is used as the teacher model to train a student model with a 6-layer Transformer structure. The hidden layer dimension is kept at 768. During training, soft labels and hidden layer features are passed at the same time. The loss function is composed of a weighted sum of soft label loss and feature matching loss. After training, the student model is statically quantized with INT8 to convert the model parameters and activation values ​​from 32-bit floating-point numbers to 8-bit integers. At the same time, operator fusion technology is adopted to merge adjacent convolutional layers, normalization layers and activation layers into a single operator.

[0258] For example, to calculate the similarity between node N004 and node N002, a pre-trained text vector model is called to generate a fragment vector of node N004. The pre-trained text vector model is also called to generate a fragment vector of node N002. The dot product of the two vectors is calculated, resulting in 0.36. The magnitude of the vector of node N004 is calculated, resulting in 0.75. The magnitude of the vector of node N002 is calculated, resulting in 0.80. The product of the two magnitudes is calculated, resulting in 0.60. The dot product is divided by the product of the magnitudes, resulting in a similarity of 0.6. The similarity between node N004 and node N003 is then calculated.

[0259] The pre-trained text vector model is invoked to generate the fragment vector of node N003. The dot product of the vectors of node N004 and node N003 is calculated, resulting in 0.30. The magnitude of the vector of node N003 is calculated, resulting in 0.75. The product of the two magnitudes is calculated, resulting in 0.5625. The dot product is divided by the magnitude product, resulting in a similarity of 0.5. The maximum value between 0.6 and 0.5 is taken, resulting in 0.6. 0.6 is used as the coverage of node N004.

[0260] Step S30443: Based on the coverage level and the set of deviation evidence nodes, determine the uncovered relationship, partially covered relationship and deviation relationship corresponding to each external source hint evidence node.

[0261] In this embodiment, the deviation degree of each external prompt evidence node is calculated based on the coverage degree and the set of deviation evidence nodes. The deviation degree is calculated by weighting the semantic non-coverage degree, contradiction trigger value, trend difference value and time difference value. The semantic non-coverage degree is 1 minus the coverage degree.

[0262] The contradiction trigger value is used to determine whether there is negation, reversal, reversal, or obvious inconsistency in the expression. It compares the semantic content of external cue evidence and input coverage evidence. If the above situations exist, it is set to 1; otherwise, it is set to 0. The trend difference value is used to determine whether the changing trends of the two segments are opposite. If the current input indicates relief, normality, or absence, while the external cue indicates increase, rise, or abnormality, the trend difference value is set to 1; otherwise, it is set to 0. The time difference value is calculated based on the time interval between the two segments. Specifically, the ratio of the time interval between the two segments to the time base number of days is used as the input of the exponential decay term. The natural constant is used as the base, and the negative value of the ratio is used as the exponent to calculate the time correlation value. The time difference value is obtained by subtracting the time correlation value from 1. The time base number of days is determined through historical samples.

[0263] Each weight coefficient is determined through optimization using labeled samples. Multiple sets of manually labeled difference judgment samples are prepared. Each set includes input coverage evidence, external cue evidence, deviation evidence, and corresponding relationship labels. The relationship labels include uncovered relationships, partially covered relationships, and deviation relationships. Based on these labeled samples, different weight combinations are iterated through. The coverage difference results calculated for each weight combination are compared with the manually labeled results, and the optimal weight combination is determined based on the classification accuracy.

[0264] Determining the uncovered threshold, the fully covered threshold, and the deviation threshold specifically includes: based on the labeled samples, traversing the candidate thresholds, matching the uncovered relationships, partially covered relationships, and deviation relationships generated under different candidate thresholds with the manually labeled results, calculating the precision and recall corresponding to each candidate threshold, determining the comprehensive value based on the precision and recall, and selecting the candidate threshold with the largest comprehensive value as the corresponding relationship judgment threshold.

[0265] When the coverage is less than the uncovered threshold, it is determined to be an uncovered relationship. When the coverage is greater than or equal to the uncovered threshold and less than the full coverage threshold, it is determined to be a partial coverage relationship. When the coverage is greater than or equal to the full coverage threshold, it is determined to be a covered relationship. When the deviation is greater than or equal to the deviation threshold, it is determined to be a deviation relationship. If the same external source hint evidence satisfies both the deviation relationship and the partial coverage relationship, it is marked as a deviation relationship first, while the partial coverage relationship is retained as an auxiliary marker.

[0266] In existing technologies, coverage difference relationships are usually determined solely based on coverage, categorizing relationships into three types: uncovered, partially covered, and covered. This single-dimensional approach cannot identify situations where semantic coverage is partial but contradictory or shows opposing trends. For example, a user might say "sleep quality has improved," while external data shows "nighttime activity has significantly increased in the last three days." Both statements show partial semantic coverage but opposite trends. Existing technologies would classify this as a partial coverage relationship, generating a refined trigger item, when in fact it should be classified as a deviation relationship, generating a partitioned trigger item. This misjudgment leads to incorrect trigger item types, affecting the accuracy of subsequent interactions.

[0267] This embodiment introduces the concept of deviation, which is calculated from four dimensions: semantic non-coverage, contradiction, trend difference, and time difference. It also stipulates that when the same evidence satisfies both deviation relationship and partial coverage relationship, it is marked as deviation relationship. This joint judgment mechanism can accurately identify situations with contradiction or opposite trends, avoid misjudgment of trigger item type, and greatly improve the accuracy of coverage difference relationship judgment.

[0268] The method for determining the contradiction trigger value is determined according to the application scenario. For example, the preferred method in this application embodiment is the keyword matching method, which presets a negative word library and a reverse expression word library. If the match is successful, the trigger value is 1.

[0269] Step S30444: Based on the uncovered relationship, partially covered relationship and deviation relationship corresponding to each external source prompt evidence node, associate and record the corresponding collaborative object node and evidence node to obtain the coverage difference relationship record set.

[0270] In this embodiment, based on the coverage differential relationship corresponding to each external prompt evidence node, a unique relationship record number is generated for each relationship, and the corresponding candidate collaborative object number, external prompt evidence number, supporting evidence number, coverage, deviation and relationship type are associated. The relationship record is then checked for completeness, missing association information is filled in, and all generated and verified relationship records are integrated into a coverage differential relationship record set.

[0271] The storage method for the differential relationship record set is determined based on the amount of data. For example, in this application embodiment, the preferred storage method is relational database storage, where each record corresponds to a row of data in the database.

[0272] For example, a relationship record with the number R001 is generated for the partial coverage relationship corresponding to node N004, and associated with candidate collaborative object number 001, external source hint evidence number E004, supporting evidence number E001, coverage 0.6, deviation 0.2048 and relationship type partial coverage. Similarly, other relationship records are generated to obtain a set of coverage difference relationship records.

[0273] Step S3045: Based on the coverage differential relationship record set, generate supplementary collaborative trigger items, refined collaborative trigger items, and distinguishable collaborative trigger items, and sort them by deduplication to obtain the key collaborative trigger item set.

[0274] In this embodiment, based on the cover differential relationship record set, all cover differential relationship records are traversed in order of relationship record number. Uncovered relationships generate supplementary collaborative trigger items, partially covered relationships generate refined collaborative trigger items, and deviation relationships generate distinguishable collaborative trigger items.

[0275] For each collaborative trigger, a unique trigger number is generated, which is associated with the corresponding candidate collaborative object number, trigger type, trigger source, corresponding supporting evidence, corresponding external source hint evidence, coverage difference relationship, coverage, deviation, trigger score, and trigger description. The trigger score is calculated by weighting the hint edge weight, 1 minus coverage, deviation, evidence validity value, and source diversity. The source diversity is 1 when the external source hint evidence comes from both historical interaction records and associated collected data, and 0.5 when it comes from only one source.

[0276] Each weight coefficient is determined by the analytic hierarchy process. For example, in this application embodiment, the edge weight is preferably 0.25, the coverage weight (1 minus coverage) is preferably 0.25, the deviation weight is preferably 0.25, the evidence validity weight is preferably 0.15, and the source diversity weight is preferably 0.10. The trigger threshold is determined by the labeled samples. For example, in this application embodiment, the trigger threshold is preferably 0.60 when there are no labeled samples. Collaborative triggering items with trigger scores greater than or equal to the trigger threshold are retained.

[0277] The retained collaborative trigger items are deduplicated. If two key collaborative trigger items correspond to the same candidate collaborative object, the same external cue evidence, and the same trigger item type, the one with the higher trigger score is retained. The deduplicated collaborative trigger items are sorted, with distinguishing trigger items having the highest priority, supplementary trigger items having the second highest priority, and refined trigger items having the third highest priority. Within the same type, they are sorted from high to low trigger scores. All sorted collaborative trigger items are integrated into a key collaborative trigger item set.

[0278] In existing technologies, the generation of collaborative triggers typically employs a simple threshold determination method, generating triggers solely based on whether the coverage falls below a certain threshold. This method fails to consider factors such as the credibility of evidence, the diversity of sources, and the degree of deviation, resulting in an inability to distinguish the importance of triggers. Furthermore, the ordering of triggers is usually based solely on coverage from low to high, failing to prioritize the most important triggers. In scenarios with a large number of triggers, this ordering method can lead to important deviation-related triggers being submerged among a large number of ordinary triggers.

[0279] This embodiment designs a multi-dimensional trigger score calculation method, which comprehensively considers five factors: prompt edge weight, coverage, deviation, evidence validity value, and source diversity. At the same time, it adopts a hierarchical sorting mechanism, setting the distinguishing trigger items as the highest priority, followed by supplementary trigger items, and then detailed trigger items. This design can accurately quantify the importance of each trigger item, ensuring that the most important trigger items can be processed first, and greatly improving the targeting and effectiveness of subsequent interactions.

[0280] The maximum capacity of the key collaborative trigger item set is determined according to the interaction scenario. For example, the maximum capacity of this application embodiment is preferably 20. When the capacity is exceeded, only the 20 trigger items with the highest trigger scores are retained.

[0281] exist Figure 2In the diagram, the three input unit groups arranged vertically on the left represent a set of multi-source evidence sources under the same candidate collaborative object. The diamond-shaped node at the left end of each input unit group represents different source evidence entry points extracted from the valid evidence set, including input-covered evidence sources, external-source hint evidence sources, and off-source evidence sources. The horizontal connecting line to the right of the diamond node represents the process of sending evidence from each source to the evidence node mapping process corresponding to the same candidate collaborative object. Each dashed box represents a sub-source evidence node group, and the circular node inside the dashed box represents the evidence node formed after node mapping. Multiple circular nodes correspond to input-covered evidence nodes, external-source hint evidence nodes, and off-source evidence nodes, respectively. The dashed box indicates that the above evidence nodes belong to the same sub-source verification range and can be uniformly extracted and compared in subsequent steps. The connecting line where all the circular nodes converge towards the center indicates that the sub-source evidence under the same candidate collaborative object is associated with the collaborative object node to form evidence association relationships. Figure 2 The large circular node in the center represents a collaborative object node, used to carry the unified association between input coverage evidence, external cue evidence, and deviation evidence under the same candidate collaborative topic. The three connecting lines extending to the right of the large circular node in the center represent the extraction of different types of coverage comparison evidence groups under the same collaborative object node. The three rounded rectangles arranged vertically on the right represent coverage comparison evidence groups formed by input coverage evidence nodes, external cue evidence nodes, and deviation evidence nodes. Each rounded rectangle represents a set of evidence combinations to be compared by coverage difference. The multiple arc-shaped connecting lines to the right of the rounded rectangles represent the coverage comparison edges and deviation indication edges within the evidence group. The coverage comparison edges are used to indicate the coverage relationship between the input coverage evidence node and the external cue evidence node, and the deviation indication edges are used to indicate the deviation indication relationship between the deviation evidence node and the external cue evidence node. The multiple arc-shaped connecting lines converge to the black confluence node on the right, indicating that a coverage difference relationship is formed after performing coverage difference comparison on each coverage comparison evidence group. Record sets; the connecting lines extending from the black confluence node to the three dashed circular areas on the right indicate that the record sets of coverage difference relationships have been grouped into different types of key collaborative triggers; the radial hexagonal node clusters in the upper right dashed circular area represent supplementary collaborative triggers, with the central hexagon representing the trigger body and the surrounding hexagons representing supplementary content derived from external cue evidence corresponding to the uncovered relationship; the cross-connected hexagonal node clusters in the middle right dashed circular area represent distinguishing collaborative triggers, with the cross-connection indicating a deviation between the input coverage evidence and the external cue evidence, requiring differentiation in subsequent interactions; the hierarchical hexagonal node clusters in the lower right dashed circular area represent refined collaborative triggers, with the hierarchical structure indicating that the external cue content has been partially covered but still needs further refinement and confirmation; the dashed circular areas on the right represent the merging results of different trigger types, and the connecting lines between the hexagonal nodes represent the evidence support relationship and trigger basis relationship within the same trigger.

[0282] Step S4: Based on the key collaborative trigger item set, call the output collaborative agent group to generate multi-agent collaborative scheduling interaction results.

[0283] The specific steps of step S4 are as follows:

[0284] Step S401: Based on the key collaborative trigger item set, call the output collaborative agent group to perform trigger item parsing and interaction content generation for each key collaborative trigger item, and obtain a candidate interaction content set.

[0285] In this embodiment, based on the key collaborative trigger item set, all key collaborative trigger items are traversed from high to low according to the trigger score, and the output collaborative agent group is called. Multiple agents in the output collaborative agent group jointly generate candidate interaction content around the same key collaborative trigger item.

[0286] The trigger reason explanation angle agent generates an explanation of why the trigger item was triggered; the content generation angle agent generates candidate interactive text; the correspondence verification angle agent determines whether the candidate interactive text corresponds to the key collaborative trigger item; and the expression adaptation angle agent determines whether the candidate interactive text is suitable for the current context.

[0287] Generate a unique candidate content number for each candidate interaction content, associate it with the corresponding key collaborative trigger item number, candidate interaction text, trigger item type, generation basis, and candidate content confidence field, and integrate all generated candidate interaction content into a candidate interaction content set.

[0288] The number of agents in the output collaborative agent group is determined according to the application scenario. For example, in this application embodiment, the number of agents is preferably 4, corresponding to the four angles mentioned above.

[0289] For example, the key collaborative trigger item set is traversed, and the supplementary trigger item T002 with the highest trigger score is selected. The output collaborative agent group is called, and the trigger reason explanation agent generates the description "According to your smart bracelet data, your nighttime activity has increased significantly in the last three days". The content generation agent generates the candidate interactive text "Your nighttime activity has increased significantly in the last three days. Do you have difficulty falling asleep or frequent awakenings at night?" The correspondence verification agent confirms that the content corresponds to the trigger item. The expression adaptation agent confirms that the content is suitable for the current context and generates candidate interactive content with the candidate content number C001. The confidence level of the candidate content is 0.90.

[0290] Step S402: Based on the candidate interaction content set, perform trigger item corresponding verification, evidence coverage verification, and duplicate content screening on each candidate interaction content to obtain subsequent interaction content.

[0291] In this embodiment, based on the candidate interaction content set, a matching score is calculated for each candidate interaction content. The matching score is obtained by weighting the trigger item matching degree, evidence coverage, expression clarity, and duplication penalty. The weight coefficients corresponding to each indicator are determined through optimization using labeled samples. Specifically, multiple manually labeled interaction content evaluation samples are prepared. Each sample includes candidate interaction content, corresponding key collaborative trigger items, related evidence, and the effectiveness result of manual labeling. Different weight combinations are iterated, and the optimal weight combination is determined based on the degree of matching between the matching score calculated for each weight combination and the effectiveness result of manual labeling. This yields the weight coefficients corresponding to the trigger item matching degree, evidence coverage, expression clarity, and duplication penalty.

[0292] Trigger matching degree is the cosine similarity between the candidate interaction content vector and the key collaborative trigger vector; evidence coverage is the number of trigger evidences covered by the candidate interaction content divided by the total number of evidences corresponding to the trigger; clarity of expression is calculated by weighting the purpose clarification value, the length reasonable value, and the proportion of ambiguous words; and duplication penalty is the maximum similarity between the candidate interaction content and the historical interaction content.

[0293] The system calls upon multiple agents in the output collaborative agent group to cross-verify the candidate interaction content. The verification criteria include whether it corresponds to a key collaborative trigger, whether it covers the evidence corresponding to the trigger, whether it is repetitive, whether it is clearly expressed, and whether it is suitable for the current context. The system calculates the verification score and validity value for each candidate interaction content.

[0294] The candidate interactive content with the highest valid value that reaches the valid threshold is selected as the subsequent interactive content. If none of the candidate interactive content reaches the threshold, the candidate interactive content is regenerated. If the threshold is still not reached after a preset number of regenerations, the content with the highest valid value is selected as the low-confidence interactive content.

[0295] The effective threshold for interactive content is determined by labeled samples. For example, in this application embodiment, the effective threshold for interactive content is preferably 0.65 when there are no labeled samples, and the maximum number of times it is continuously regenerated is preferably 3 times.

[0296] Step S403: Collect user feedback content corresponding to subsequent interaction content, and associate the user feedback content with the corresponding key collaborative trigger items to obtain feedback association records.

[0297] In this embodiment, the system receives feedback from the user regarding subsequent interactions and extracts the user identifier, feedback text, feedback time, corresponding subsequent interaction content number, and corresponding key collaborative trigger item number fields from the feedback content.

[0298] Semantic fragment extraction is performed on the feedback text to obtain feedback fragments. These fragments are then associated with the corresponding key collaborative triggers and subsequent interactive content. A unique feedback record number is generated for each association, and all association information is integrated into a feedback association record.

[0299] The feedback association record includes the feedback record number, user identifier, subsequent interaction content number, key collaborative trigger item number, feedback fragment text, feedback time, feedback clarity, feedback resolution, trigger item processing status, associated evidence number, and iteration usage marker field. Feedback clarity indicates whether the feedback content clearly responds to subsequent interaction content. Feedback resolution indicates whether the feedback content resolves the corresponding key collaborative trigger item. Iteration usage marker indicates whether the feedback record has entered the subsequent rule update, threshold update, and model sample update process.

[0300] The method for extracting semantic fragments from feedback text is consistent with the basic semantic fragment extraction method described above. For example, the segmentation criteria for semantic fragment extraction in this application embodiment are preferably commas, periods, question marks, and semicolons.

[0301] For example, receiving feedback from user A, "Yes, I've been waking up two or three times every night lately," we extract the user identifier as A, the feedback text as the aforementioned content, the feedback time as 10:35 AM on June 3, 2026, the corresponding subsequent interaction content number as C001, and the corresponding key collaborative trigger item number as T002. Semantic fragment extraction is performed on the feedback text to obtain the feedback fragment "Yes, I've been waking up two or three times every night lately," generating a feedback association record with feedback record number F001. This feedback record has a feedback clarity of 1, a feedback resolution of 0.98, a trigger item processing status of "processed," and is marked as available for iteration.

[0302] Step S404: Based on the feedback association record, determine the processing status of each key collaborative trigger item, and collaboratively merge the subsequent interaction content, feedback content and processing status to obtain the multi-agent collaborative scheduling interaction result.

[0303] In this embodiment, the feedback resolution score is calculated based on the feedback association records. The feedback resolution score is obtained by weighting the similarity between the feedback and the triggering item, the coverage of feedback evidence, and the clarity of feedback. The weight coefficients for each indicator are determined through optimization using labeled samples. Specifically, multiple manually labeled feedback resolution score samples are constructed. Each sample includes a key collaborative triggering item, corresponding feedback content, associated evidence, and a manually labeled resolution status tag. Based on these labeled samples, different weight combinations are iterated. The feedback resolution score results calculated for each weight combination are matched with the manually labeled resolution status tags. The optimal weight combination is determined based on the matching accuracy, resulting in the weight coefficients corresponding to the similarity between the feedback and the triggering item, the coverage of feedback evidence, and the clarity of feedback.

[0304] Feedback and trigger similarity is the cosine similarity between the feedback text vector and the key collaborative trigger vector. Feedback evidence coverage is the number of evidence covered by the feedback divided by the total number of evidence corresponding to the trigger. Feedback clarity is determined based on the clarity of the feedback content: 1 for explicit explanations, explicit denials, or explicit supplements; 0.5 for ambiguous expressions; and 0 for irrelevant triggers.

[0305] When the feedback resolution reaches the feedback resolution threshold, the corresponding key collaborative trigger item is marked as processed; otherwise, it is marked as pending further processing. The user identifier, unified collaborative context summary, candidate collaborative object set, key collaborative trigger item set, subsequent interaction content, user feedback content, processed trigger items, pending further processing trigger items, and collaborative output description are integrated to generate multi-agent collaborative scheduling interaction results.

[0306] The feedback resolution threshold is determined by labeled samples. For example, in the embodiments of this application, the feedback resolution threshold is preferably 0.65 when there are no labeled samples.

[0307] Based on the feedback association records, feedback records that meet the iterative usage conditions are iteratively aggregated. These conditions include a feedback clarity greater than or equal to 0.8, a feedback resolution greater than or equal to the feedback resolution threshold, a clear correspondence between the feedback content and key collaborative triggers, and traceable evidence sources for the feedback records. Feedback records meeting these conditions are written into a feedback sample pool, which includes positive feedback samples, supplementary feedback samples, and error-correcting feedback samples. Positive feedback samples confirm the match between key collaborative triggers and subsequent interaction content; supplementary feedback samples expand the keywords and thematic expressions corresponding to external source prompt evidence; and error-correcting feedback samples correct erroneous correspondences between deviation evidence, trigger types, and candidate interaction content.

[0308] Based on the feedback sample pool, the agent rules are periodically updated. For key collaborative triggers that are repeatedly confirmed by the user, the priority weight of the corresponding trigger type is increased. For key collaborative triggers that are repeatedly rejected by the user or marked as irrelevant, the priority weight of the corresponding trigger type is decreased, and the corresponding evidence combination is written into the exclusion rules. For keywords, object words, and status words newly added in user feedback, they are added to the fragment element dictionary and the external prompt dictionary. The rule update cycle is determined according to the number of feedback samples. When the number of new samples added to the feedback sample pool reaches a preset sample number threshold, an update is triggered. For example, in this embodiment, the preset sample number threshold is preferably 100.

[0309] Based on the feedback sample pool, the topic matching threshold, evidence screening threshold, trigger threshold, and feedback resolution threshold are updated. The threshold update adopts a sliding window statistical method. In multiple feedback samples, the candidate thresholds from 0.1 to 0.9 are traversed in increments of 0.01, and the precision, recall, and comprehensive value are calculated respectively. The candidate threshold with the largest comprehensive value is selected as the updated threshold. To avoid threshold abrupt changes, an exponential smoothing method is used to update the threshold. The updated threshold is equal to the smoothing coefficient multiplied by the previous round threshold, plus the difference between the smoothing coefficient and the current round candidate threshold. For example, in this application embodiment, the smoothing coefficient is preferably 0.8.

[0310] Based on the feedback sample pool, the generative pre-trained model is incrementally fine-tuned. The incremental fine-tuning samples consist of feedback text, key collaborative triggers, subsequent interaction content, processing status, and manual confirmation labels. Positive feedback samples are used as positive samples, and candidate interaction content rejected by the user in the error correction feedback samples are used as negative samples to construct paired training samples. Incremental fine-tuning adopts a mini-batch training method, with a learning rate lower than the initial training learning rate. For example, in this embodiment, the incremental fine-tuning learning rate is preferably one-tenth of the initial training learning rate. After incremental fine-tuning is completed, the accuracy of the model output is verified using a validation sample set. When the accuracy does not decrease and the proportion of duplicate content does not increase, the online model parameters are replaced.

[0311] Example 2:

[0312] See Figure 3 This embodiment provides a multi-agent collaborative scheduling query interaction system, including: a data acquisition module, a context collaboration module, a source verification module, and a collaborative interaction module.

[0313] The data acquisition module is used to acquire the scheduling dataset to be parsed.

[0314] The context collaboration module, based on the scheduling dataset to be parsed, calls the context collaboration agent group to merge collaboration topics, and obtains a candidate collaboration object set and a source context fragment set.

[0315] The source-specific verification module, based on the candidate collaborative object set and the source-specific context fragment set, calls the source-specific verification collaborative intelligent agent group to construct a multi-source collaborative differential evidence graph and generate a key collaborative trigger item set.

[0316] The collaborative interaction module, based on the key collaborative trigger item set, calls the output collaborative agent group to generate multi-agent collaborative scheduling and interaction results.

[0317] The specific functions of each module described above are explained in the relevant content of the method in Embodiment 1, and will not be repeated here.

[0318] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0319] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A query interaction method for multi-agent cooperative scheduling, characterized in that, include: Obtain the scheduling dataset to be parsed; Based on the scheduling dataset to be parsed, the context cooperative agent group is invoked to perform cooperative topic merging, resulting in a candidate cooperative object set and a source context fragment set. Based on the candidate collaborative object set and the source context fragment set, the source verification collaborative intelligent agent group is called to construct a multi-source collaborative differential evidence graph and generate a key collaborative trigger item set. The key collaborative trigger item set is a set of trigger items formed by merging the uncovered relationship, partially covered relationship and deviation relationship after comparing the coverage difference between the input coverage evidence and the external prompt evidence through the multi-source collaborative differential evidence graph. Based on the key collaborative trigger item set, the output collaborative agent group is invoked to generate multi-agent collaborative scheduling interaction results.

2. The query interaction method for multi-agent cooperative scheduling according to claim 1, characterized in that, The process involves invoking a context-cooperative intelligent agent group to perform cooperative topic merging based on the scheduling dataset to be parsed, resulting in a candidate cooperative object set and a source-specific context fragment set, including: Based on the scheduling dataset to be parsed, text conversion and segmentation are performed to obtain the original context fragment set; Based on the original set of context fragments, a set of candidate context units is generated by calling a context cooperative intelligent agent group. The set of candidate context units includes the recording time and candidate cooperative topics. Based on the candidate context unit set, intra-group cross-verification is performed to obtain the effective context unit set; Based on the set of effective context units, effective context units pointing to the same candidate collaborative topic are merged to obtain a set of candidate collaborative objects and a set of source context fragments.

3. The query interaction method for multi-agent cooperative scheduling according to claim 2, characterized in that, The step of generating a candidate context unit set by invoking a context cooperative agent group based on the original context fragment set includes: Based on the original set of context fragments, the context collaborative intelligent agent group is invoked to extract fragment elements from each original context fragment to obtain fragment element candidate results, which include semantic elements, source elements and time elements. Based on the candidate fragment element results, candidate collaborative topic matching is performed on each original context fragment to obtain the topic candidate results; Based on the candidate results of fragment elements and the candidate results of themes, each original context fragment is organized into a unit to obtain a set of candidate context units.

4. The query interaction method for multi-agent cooperative scheduling according to claim 1, characterized in that, The process involves constructing a multi-source collaborative differential evidence graph based on a candidate collaborative object set and a source-specific context fragment set, and generating a key collaborative trigger item set, including: Based on the candidate collaborative object set and the source context fragment set, a source verification task package is constructed for each candidate collaborative object. The source verification task package includes the current input fragment, historical record fragment, collected data fragment, as well as the corresponding candidate collaborative topic, fragment elements and recording time. Based on the source verification task package, the source verification collaborative intelligent agent group is called to construct the expression coverage retrieval direction and the external source prompt retrieval direction, and the consistency comparison between sources is performed to obtain the candidate evidence set; Based on the candidate evidence set, the source-based verification collaborative intelligent agent group is invoked to perform intra-group cross-verification and divergence backtracking screening to obtain the effective evidence set. Based on the effective evidence set, a multi-source collaborative differential evidence graph is constructed, and coverage differential comparison and collaborative trigger item merging are performed to obtain the key collaborative trigger item set.

5. The query interaction method for multi-agent cooperative scheduling according to claim 4, characterized in that, The process involves using a source-based verification task package to invoke a source-based verification collaborative intelligent agent group to construct a description coverage retrieval direction and an external source suggestion retrieval direction, and then performing a consistency comparison between sources to obtain a candidate evidence set, including: Based on the current input fragment in the source-based verification task package, construct the expression coverage retrieval direction to obtain input coverage evidence; Based on the historical record fragments and collected data fragments in the source verification task package, external source hint retrieval directions are constructed to obtain external source hint evidence; Based on the consistency comparison between input coverage evidence and external cue evidence, external cue evidence that is inconsistent with the input coverage evidence in terms of candidate collaborative themes, fragment elements, or recording time is identified as deviation evidence. A candidate evidence set is formed based on input coverage evidence, external cue evidence, and deviation evidence.

6. The query interaction method for multi-agent cooperative scheduling according to claim 4, characterized in that, The process involves invoking a collaborative intelligent agent group for source verification based on the candidate evidence set to perform intra-group cross-verification and divergence backtracking screening, resulting in a valid evidence set, including: Based on the candidate evidence set, the source-based verification collaborative intelligent agent group is invoked to perform multi-dimensional cross-verification of each candidate evidence, and the evidence verification result set is obtained. Based on the evidence verification result set, identify the matching discrepancies between the same candidate evidence and candidate collaborative objects and source verification task packages, and obtain the discrepancy evidence result set. Based on the divergent evidence result set, the corresponding source verification task package is traced back, and the candidate evidence with divergence is corrected or removed to obtain the corrected evidence result set. Based on the candidate evidence set and the revised evidence result set, duplicate evidence is merged and low-credibility evidence is screened out to obtain the effective evidence set.

7. The query interaction method for multi-agent cooperative scheduling according to claim 4, characterized in that, Based on the effective evidence set, a multi-source collaborative differential evidence graph is constructed, and coverage differential comparison and collaborative trigger item merging are performed to obtain a key collaborative trigger item set, including: Based on the effective evidence set and the candidate collaborative object set, node mapping is performed on the candidate collaborative objects and their corresponding input coverage evidence, external hint evidence and deviation evidence to obtain the collaborative object node set and the evidence node set. Based on the collaborative object node set and the evidence node set, evidence association edges, coverage comparison edges and deviation indication edges are established to construct a multi-source collaborative differential evidence graph. Based on the multi-source collaborative differential evidence graph, input coverage evidence nodes, external prompt evidence nodes, and deviation evidence nodes under the same collaborative object node are extracted to obtain the coverage comparison evidence group. Based on the coverage comparison evidence set, the input coverage evidence nodes, external prompt evidence nodes and deviation evidence nodes are compared by coverage difference to obtain a coverage difference relationship record set. Based on the coverage differential relationship record set, supplementary collaborative trigger items, refined collaborative trigger items, and distinguishable collaborative trigger items are generated, and deduplication and sorting are performed to obtain the key collaborative trigger item set.

8. The query interaction method for multi-agent cooperative scheduling according to claim 7, characterized in that, The coverage comparison evidence set involves performing coverage differential comparison on the input coverage evidence nodes, external prompt evidence nodes, and deviation evidence nodes to obtain a coverage differential relationship record set, including: Based on the coverage comparison evidence group, determine the set of input coverage evidence nodes and the set of deviation evidence nodes corresponding to each external prompt evidence node under the same collaborative object node; Based on the coverage comparison edge between the external prompt evidence node and the set of input coverage evidence nodes, the coverage degree of each external prompt evidence node relative to the input coverage evidence is determined. Based on the coverage level and the set of deviation evidence nodes, determine the uncovered relationship, partially covered relationship and deviation relationship corresponding to each external source hint evidence node; Based on the uncovered, partially covered, and deviation relationships corresponding to each external source evidence node, the corresponding collaborative object nodes and evidence nodes are associated and recorded to obtain a set of coverage difference relationship records.

9. The query interaction method for multi-agent cooperative scheduling according to claim 1, characterized in that, The step of calling the output cooperative agent group to generate multi-agent cooperative scheduling interaction results based on the key cooperative trigger item set includes: Based on the set of key collaborative trigger items, the output collaborative agent group is called to perform trigger item parsing and interaction content generation for each key collaborative trigger item, thereby obtaining a set of candidate interaction content. Based on the candidate interaction content set, each candidate interaction content undergoes trigger item verification, evidence coverage verification, and duplicate content removal to obtain subsequent interaction content; Collect user feedback content corresponding to subsequent interaction content, and associate user feedback content with corresponding key collaborative trigger items to obtain feedback association records; Based on feedback association records, the processing status of each key collaborative trigger item is determined, and subsequent interaction content, feedback content, and processing status are collaboratively merged to obtain the multi-agent collaborative scheduling interaction results.

10. A multi-agent cooperative scheduling query interaction system, used to implement the multi-agent cooperative scheduling query interaction method according to any one of claims 1 to 9, characterized in that, include: The data acquisition module is used to acquire the scheduling dataset to be parsed. The context collaboration module, based on the scheduling dataset to be parsed, calls the context collaboration agent group to merge collaborative topics, and obtains a candidate collaboration object set and a source context fragment set; The source verification module, based on the candidate collaborative object set and the source context fragment set, calls the source verification collaborative intelligent agent group to construct a multi-source collaborative differential evidence graph and generate a key collaborative trigger item set; The collaborative interaction module, based on the key collaborative trigger item set, calls the output collaborative agent group to generate multi-agent collaborative scheduling and interaction results.

Images