High-precision reasoning retrieval method and system based on multi-target cooperation

Abstract

The application discloses a high-precision reasoning retrieval method and system based on multi-target cooperation. The method comprises the following steps: analyzing a retrieval request and obtaining relevant information from a knowledge graph, defining evidence units as protocol element construction sets, selecting optimal protocol element combinations to form initial evidence chains under the condition of considering cost and safety, generating a dynamic adjustment factor according to the research stage characteristics of the retrieval request, adjusting the membership threshold, performing multi-target cooperative optimization, establishing a correction parameter library, matching the best correction parameter group from the library according to real-time characteristics to optimize the evidence chain in real time, obtaining the final optimized evidence chain, obtaining target knowledge data and configuring reasoning weights, and generating reasoning reply information. Through the implementation of the application, the problems of redundant evidence chain construction, rigid evidence weight distribution, difficulty in balancing reasoning accuracy and calculation efficiency, lack of privacy security and efficiency cooperative optimization mechanism in the knowledge retrieval method for research-type intelligent agents in the prior art are solved.

Description

Technical Field

[0001] This invention relates to the field of knowledge reasoning and intelligent retrieval technology, and more specifically to a high-precision reasoning retrieval method and system based on multi-objective collaboration. Background Technology

[0002] With the rapid development of big data and artificial intelligence technologies, the demand for in-depth research in knowledge-intensive industries such as technology, finance, healthcare, and manufacturing is growing. Autonomous research agents, exemplified by DeepResearch, operate on a core paradigm of problem decomposition, multi-source retrieval, fact-checking, and report generation. This requires the underlying knowledge retrieval system to not only quickly return relevant information but also possess the ability to deeply screen, organically integrate, and logically deduce multi-source heterogeneous evidence.

[0003] Existing graph network multimodal knowledge retrieval systems often exhibit fragmented and redundant search results when facing in-depth research scenarios, lacking the ability to effectively screen and organically integrate multi-source evidence. The ranking and evidence integration layers of traditional retrieval systems typically rely solely on simple semantic similarity or fixed weights to pile up results, failing to select the optimal subset of evidence from massive candidate evidence that satisfies logical completeness, controllable computational costs, and a balanced diversity of results. This leads to severe deficiencies in the accuracy, interpretability, and traceability of reasoning conclusions, making it difficult to support the stringent requirements of research agents for fact-checking and multi-hop reasoning. Meanwhile, the evidence requirements of research agents differ significantly at different research stages. In the overview scanning stage, agents prioritize response speed and recall, allowing for a moderate sacrifice in the depth of analysis of individual pieces of evidence; in the key point drilling stage, agents require high-precision, high-credibility in-depth analysis of core entities; and in the logical deduction stage, agents emphasize the logical consistency and causal completeness of the overall evidence chain. However, existing technologies generally employ static ranking strategies, failing to adaptively adjust the weights of evidence precision, speed, and data dependency based on the dynamic characteristics of the research stage. This results in rigid allocation of retrieval resources, interference between complex research queries and simple information queries, and a difficulty in simultaneously achieving both overall system retrieval accuracy and inference efficiency. Furthermore, in multi-institutional collaborations or cross-domain knowledge service scenarios, evidence chain construction must balance the strength of data privacy protection with computational efficiency. Existing technologies lack mechanisms to formally incorporate privacy and security constraints into the evidence screening process, leaving highly sensitive data such as medical records and financial transaction records at risk of privacy breaches during retrieval and inference, or causing the loss of relevance of key evidence due to overly conservative access controls.

[0004] Therefore, it is necessary to design a new method to address the problems of redundant evidence chain construction, rigid evidence weight allocation, difficulty in balancing reasoning accuracy and computational efficiency, and lack of a mechanism for coordinating privacy and efficiency optimization in existing knowledge retrieval methods for research agents. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a high-precision reasoning retrieval method and system based on multi-target collaboration.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a high-precision reasoning retrieval method based on multi-target collaboration, comprising: Obtain a retrieval request, retrieve retrieval information from the knowledge graph based on the retrieval request, define each evidence unit in the retrieval information as a protocol element, and construct a protocol element set; A fuzzy matroid model is constructed based on the set of protocol elements; Based on the fuzzy matroid model, the robust minimum weight basis is selected in the worst-case scenario by using the robust minimum weight basis solution algorithm to obtain the protocol element combination that satisfies the preset security threshold and association threshold and has the minimum total cost, thus forming the initial evidence chain; The research stage characteristics of the search request are analyzed to generate a dynamic adjustment factor. Based on the dynamic adjustment factor, the membership threshold of the fuzzy independence constraint in the fuzzy matroid model is adjusted to redetermine the fuzzy independent set family. Based on the adjusted fuzzy matroid model, the preliminary evidence chain is optimized in a multi-objective collaborative manner to construct a correction parameter library containing multiple subspace correction parameters. Based on the real-time characteristics of the search request, the target correction parameter group is matched from the correction parameter library to correct the initial evidence chain in real time, thereby obtaining the optimized evidence chain. Based on the optimized evidence chain, target knowledge data is obtained and inference weights are configured to generate inference response information.

[0007] This invention also provides a high-precision reasoning retrieval system based on multi-target collaboration, comprising: The acquisition unit is used to acquire a retrieval request, retrieve retrieval information from the knowledge graph based on the retrieval request, define each evidence unit in the retrieval information as a protocol element, and construct a protocol element set. Construction unit, used to construct a fuzzy matroid model based on the set of protocol elements; The initial selection unit is used to select, in the worst-case scenario, a combination of protocol elements that satisfies the preset security threshold and the association threshold and has the lowest total cost, based on the fuzzy matte model and the robust minimum weight basis solution algorithm, to obtain the robust minimum weight basis and form the initial selection evidence chain; The correction parameter library establishment unit is used to analyze the research stage characteristics of the retrieval request to generate dynamic adjustment factors, adjust the membership degree threshold of the fuzzy independence constraint in the fuzzy matte model based on the dynamic adjustment factors, redetermine the fuzzy independent set family, and perform multi-objective collaborative optimization on the initial evidence chain based on the adjusted fuzzy matte model to construct a correction parameter library containing multiple subspace correction parameters. The correction unit is used to match the target correction parameter group from the correction parameter library according to the real-time characteristics of the retrieval request, and to correct the initial evidence chain in real time to obtain the optimized evidence chain. The generation unit is used to obtain target knowledge data based on the optimized evidence chain, configure reasoning weights, and generate reasoning response information.

[0008] The advantages of this invention compared to existing technologies are as follows: This invention acquires retrieval requests and extracts relevant evidence units from a knowledge graph, defining them as protocol elements and constructing a protocol element set. Based on this set, a fuzzy matte model is established, and a robust minimum basis algorithm is used to select the protocol element combination with the lowest total cost while satisfying preset security and association thresholds, forming a preliminary evidence chain, effectively avoiding the problem of redundant evidence chain construction. Subsequently, the research stage characteristics of the retrieval request are analyzed to generate a dynamic adjustment factor. Based on this factor, the membership degree threshold of the fuzzy independence constraint in the fuzzy matte model is adjusted, the fuzzy independent set family is redefined, and the preliminary evidence chain is optimized through multi-objective collaborative optimization, constructing a correction parameter library. This solves the problems of rigid evidence weight allocation and difficulty in balancing inference accuracy and computational efficiency. Finally, the preliminary evidence chain is corrected in real time based on the real-time features of the retrieval request by matching the optimal correction parameter set, resulting in an optimized evidence chain. Based on this, target knowledge data is obtained to configure inference weights and generate response information, achieving collaborative optimization of privacy security and efficiency, and improving the performance of the research agent in complex query scenarios. This series of steps ensures a logically complete chain of evidence with minimal computational overhead, while adaptively adjusting evidence weights to meet the needs of different research stages, thereby enhancing the system's response speed and the accuracy of its conclusions.

[0009] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 A flowchart illustrating the high-precision reasoning retrieval method based on multi-target collaboration provided in an embodiment of the present invention; Figure 2 A schematic block diagram of a high-precision reasoning retrieval system based on multi-target collaboration provided in an embodiment of the present invention; Figure 3 A schematic block diagram of a computer device provided for an embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0014] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0015] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0016] Please see Figure 1 Figure 1 This is a flowchart illustrating a high-precision reasoning retrieval method based on multi-objective collaboration provided in an embodiment of the present invention. This method is applied in a server. Through a series of meticulous steps, it effectively solves the main problems existing in research-oriented intelligent agent knowledge retrieval. First, it introduces a fuzzy matroid model and a robust minimum basis algorithm to construct an initial evidence chain, ensuring that security and relevance requirements are met in the worst-case scenario while minimizing total cost. Second, it optimizes the model by dynamically adjusting factors to adapt to the needs of different research stages, achieving a balance between computational accuracy and response speed. Next, it uses a multi-objective collaborative optimization strategy to construct a modified parameter library, matching the optimal parameter combination based on real-time features to further optimize the evidence chain. Finally, it obtains and configures weights based on the optimized evidence chain to generate the final response information, thereby improving the efficiency of evidence chain construction, effectively balancing reasoning accuracy and computational efficiency, and strengthening the collaborative optimization mechanism of privacy protection and efficiency. This method significantly reduces the redundancy of evidence chain construction and improves the flexibility and accuracy of evidence weight allocation.

[0017] Figure 1This is a flowchart illustrating the high-precision reasoning and retrieval method based on multi-target collaboration provided in an embodiment of the present invention. Figure 1 As shown, the method includes the following steps S110 to S160.

[0018] S110. Obtain a retrieval request, retrieve retrieval information from the knowledge graph based on the retrieval request, define each evidence unit in the retrieval information as a protocol element, and construct a protocol element set.

[0019] The system first receives a search request input by a user or research agent. The search request can be a natural language query or a multimodal query, such as "analyze the latest advances in targeted therapy for lung cancer and the evidence linking it to imaging." The system uses a pre-built feature extraction module to segment, denoise, and vectorize the search request, obtaining a combination of a set of key terms and semantic vectors as the request features.

[0020] The pre-stored knowledge graph is retrieved based on the requested features to obtain target graph nodes that match the requested features and their graph association information, thus obtaining retrieval information.

[0021] In this embodiment, the knowledge graph refers to an expert-level industry knowledge network composed of millions of multimodal entity nodes and tens of millions of entity-relation triples, and uses a hybrid architecture of graph database and columnar storage for persistent storage. The nodes in the knowledge graph include text entity nodes, image feature nodes, time-series signal nodes, and graph relation nodes. Semantic associations are established between nodes through weighted directed edges, which store relation types, confidence weights, and evidence source pointers. The knowledge graph supports precise indexing based on entity IDs, approximate nearest neighbor indexing based on semantic vectors, and graph traversal indexing based on relation types, ensuring millisecond-level node location capabilities at the billion-level scale, and possessing evidence reverse location capabilities, meaning that any inferred conclusion can be quickly traced back to the original data fragment through the original data index associated with the node.

[0022] After acquiring the target graph node, the system further extracts the graph association information of the target graph node in the knowledge graph as retrieval information. The graph association information includes the attribute data of the target graph node (such as entity name, creation time, data source, quality score), tag information, node feature vector, and the attributes, tags, relationship types, and confidence weights of the corresponding edges of the adjacent nodes that have one-hop or multi-hop associations with the target graph node, forming a structured multimodal evidence network.

[0023] In this embodiment, the protocol element set refers to a finite set consisting of all evidence units in the retrieved information, serving as the ground set for the subsequent fuzzy matroid model. Each protocol element is structured as a six-tuple, including: a unique identifier for the element, a modality type identifier, a feature vector of the evidence unit in the unified semantic space, a physical storage index pointing to the original data fragment in the knowledge base, and a topological location description of the evidence unit in the knowledge graph (including the node ID, associated edge type, and hop count). The modality type identifier includes text, image, graph, and time series.

[0024] Upon receiving a retrieval request input by a user or research agent, the system first extracts the request features corresponding to the retrieval request. These request features are data representations that characterize the core semantic intent of the retrieval request, derived by converting the retrieval request described in natural language into a numerical representation in a vector space. Specifically, a pre-defined feature extraction module performs word segmentation, denoising, and vectorization processing on the retrieval request to obtain a combination of a set of key terms and semantic vectors as the request features.

[0025] The pre-stored knowledge graph is retrieved based on the requested features to obtain target graph nodes that match the requested features. A target graph node refers to a graph entity in the knowledge graph whose node feature vector meets a preset similarity threshold with the requested features. After obtaining the target graph node, the graph association information of the target graph node in the knowledge graph is further obtained as the corresponding retrieval information. The graph association information specifically includes the attribute data, tag information, and node feature vector of the target graph node, as well as the attributes, tags, and relationship types of adjacent nodes that have direct connections or logical associations with the target graph node, forming a structured evidence network.

[0026] Each evidence unit (such as triples, text paragraphs, image slices, and time-series signal fragments) in the retrieved information is defined as a protocol element, and a protocol element set is constructed. Each protocol element includes its topological position in the knowledge graph, its original data index, its modality type identifier, and its initial semantic vector.

[0027] S120. Construct a fuzzy matroid model based on the set of protocol elements.

[0028] In this embodiment, the fuzzy matroid model refers to a mathematical structure used to optimize the process of constructing the chain of evidence. This model treats protocol elements (such as triples and text segments in a knowledge graph) as base set elements and considers differential privacy protection strength, computational efficiency, and logical redundancy as fuzzy independence constraints to ensure that the selected evidence combination satisfies both security and validity requirements while minimizing total cost.

[0029] In one embodiment, step S120 described above may include steps S121 to S123.

[0030] S121. Using protocol elements as the base set, and differential privacy protection strength, computational efficiency, and logical redundancy as fuzzy independence constraints, a membership function is used to characterize the degree to which each candidate subset satisfies the fuzzy independence constraints, and a fuzzy independent set family is determined based on a preset membership threshold.

[0031] In this embodiment, fuzzy independence constraints refer to a series of conditions that need to be met during the construction of the chain of evidence, including but not limited to differential privacy protection strength, computational efficiency, and logical redundancy. These conditions are not absolute binary states (satisfied or not satisfied), but can be quantitatively evaluated through membership functions.

[0032] The membership function is used to measure the degree to which each candidate subset meets the aforementioned fuzzy independence constraints. For example, a protocol element may have high differential privacy protection strength, but its computational cost is also high. In this case, the membership function can give a value between 0 and 1, indicating the extent to which it satisfies the fuzzy independence constraints. The membership function uses a multi-attribute weighted sigmoid function, which includes factors such as differential privacy protection strength, computational efficiency index, and logical redundancy, and uses cosine similarity to calculate logical redundancy. In addition, it includes preset weights and steepness coefficients to adjust the importance of each factor and the sensitivity of its influence.

[0033] The determination of the fuzzy independent set family adopts an incremental, layer-by-layer greedy construction method: initialization For each protocol element e in the base set E, if but For any existing set Iterate through each element e in E\I, if , ,and Satisfy the preset security threshold and association threshold, and verify each element i in I. ,but Repeat until no new set is added.

[0034] The membership function uses a multi-attribute weighted sigmoid function: ,in, For differential privacy protection ( (This is the upper bound of the safety strength interval for element e). Preset safety threshold; For calculating efficiency indicators; These are the upper bounds of the computation cost range and the upper bounds of the communication overhead range for protocol element e, respectively. For logical redundancy, cos(.,.) is the cosine similarity. For the preset weights, satisfy , Steepness coefficient, As a bias term, a preset membership threshold is set. , only The candidate subset is included in the fuzzy independent set family. When hour, ;when hour, .

[0035] A fuzzy independent set family is a combination of protocol elements that meet specific conditions, selected based on the membership function and a preset membership threshold. Only when all members of a subset reach or exceed the membership threshold is it considered part of a fuzzy independent set family.

[0036] In this way, fuzzy matroid models can effectively evaluate and select the most suitable combination of protocol elements for constructing a chain of evidence, thereby ensuring the security, effectiveness, and economy of the chain of evidence.

[0037] Specifically, the algorithm for constructing the fuzzy matroid model is as follows: Step 1: Initialize the fuzzy independent set family I f ={ }; Step 2: For each protocol element e in the base set E, if its single-element membership degree... Then {e} I f ; Step 3 (Incremental Verification): For any existing set Iterate through each element e in E\I: a. Verification It satisfies the membership threshold: ; b. Verification Meets the preset safety threshold: ; c. Verification Meets the preset association threshold: ; d. For each element i in I, verify the membership degree of the newly added tuple {i, e}: ; e. If steps a and b are all successful, then... Join I f ; Repeat step 3 until no new sets are added.

[0038] The incremental verification is based on the principle of genetic transmission: because All its subsets have satisfied the membership threshold; for The newly added subset must contain element e. The single-element subset {e} has been verified in step 2; all binary subsets {i,e} containing e are enumerated and verified one by one in step 3d; for subsets containing e with a size greater than 2... ,because and It possesses the subset monotonically non-decreasing property (the privacy budget and computational cost of the subset do not exceed those of the parent set), and Under controlled binary interaction conditions, its average value is bounded, therefore The membership degree of the tuple must be no less than that of the verified tuple, thus automatically satisfying the threshold. This incremental verification reduces the complexity from The time complexity is reduced to O(|I|), enabling the construction of fuzzy independent set families in polynomial time.

[0039] S122. Define the computational cost interval, communication overhead interval, security strength interval, and semantic relevance interval for each protocol element to obtain the interval weight matrix.

[0040] In this embodiment, the interval weight matrix is ​​a multi-dimensional data structure used to characterize the performance of each protocol element within its respective interval. Specifically: Computational cost range: quantifies the computational resources (such as CPU cycles, memory usage) required to perform a certain operation.

[0041] Communication overhead range: measures the network latency and bandwidth consumption incurred during data transmission.

[0042] Security strength range: Taking the differential privacy parameter ε as an example, it represents the level of protection for sensitive information.

[0043] Semantic relevance range: reflects the similarity or relevance between protocol elements and the feature vector of the retrieval request.

[0044] The interval weight matrix takes into account the above four aspects, providing basic data support for subsequent optimization algorithms.

[0045] Specifically, S123. The fuzzy matroid model is formed by the fuzzy independent set family and the interval weight matrix.

[0046] At this stage, the final fuzzy matroid model is formed using the previously defined fuzzy independent set family and interval weight matrix. This model not only contains information about which protocol elements should be selected into the evidence chain, but also provides methods for adjusting the selection strategy according to different objectives (such as improving accuracy, reducing latency, etc.).

[0047] Based on the aforementioned set of protocol elements, a fuzzy matroid model is constructed. Within the theoretical framework of fuzzy matroids, each protocol element is considered as an element in the fuzzy matroid, and the differential privacy protection strength and computational efficiency requirements of protocol execution during the evidence chain construction process are used as fuzzy independence constraints of the fuzzy matroid. The formal definition of the fuzzy independence constraint is: if two protocol elements both contain privacy attributes of the same sensitive entity and their differential privacy parameter ε values, when superimposed, exceed a preset security threshold, then the two are fuzzily correlated in the fuzzy matroid and cannot be selected simultaneously; the superposition adopts the sequential combination theorem. Or parallel combination theorem The selection is determined by whether the datasets accessed by the protocol elements intersect. The two are fuzzily correlated in the fuzzy matte and cannot be selected at the same time. If the cumulative computational efficiency of the combination of protocol elements is lower than the preset efficiency threshold, the fuzzy independence constraint is violated.

[0048] Based on the fuzzy matroid model, multidimensional weights are defined for each protocol element, and a comprehensive evaluation system is established to obtain the element weight matrix. The multidimensional weights include: Computational cost: quantified by CPU cycles, memory usage, and I / O overhead; Communication overhead: characterized by the amount of data transmitted across nodes and network latency; Security strength: quantified by the differential privacy parameter ε value, the smaller the ε value, the higher the privacy protection strength; Semantic relevance: quantified by the cosine similarity or Euclidean distance between protocol elements and the feature vector of the retrieval request.

[0049] This approach enables the system to flexibly respond to the needs of different research stages while ensuring the logical integrity and security of the evidence chain, achieving accurate and efficient reasoning and retrieval.

[0050] S130. Based on the fuzzy matroid model, the robust minimum basis algorithm is used to select the protocol element combination that satisfies the preset security threshold and association threshold and has the minimum total cost in the worst-case scenario, thereby obtaining the robust minimum basis and forming the initial evidence chain.

[0051] In this embodiment, the initial evidence chain refers to a subset selected from all candidate protocol elements using a robust minimum basis algorithm. This subset not only meets the preset security strength and semantic relevance threshold requirements, but also ensures that the sum of total computational cost and communication overhead is minimized in the worst case.

[0052] Based on the element weight matrix, a minimum weight basis algorithm is designed and implemented. The goal of the algorithm is to select a subset from the protocol element set such that, while satisfying that the security strength is not lower than a preset security threshold and the semantic relevance is not lower than a preset relevance threshold, the sum of the total computational cost and communication overhead is minimized.

[0053] Specifically, firstly, a total order relation is established based on the interval weight matrix using a pessimistic-optimistic comprehensive sorting method: for the protocol element e i Define pessimistic comprehensive cost ,according to Sort in ascending order, when Optimistic overall cost Ascending order is used as the second key to obtain a candidate element sequence. A greedy strategy is used to sequentially select protocol elements that meet the constraints from the candidate element sequence to construct an initial protocol element combination. Subsequently, element replacement and combination optimization are performed iteratively: attempts are made to replace selected high-cost elements with unselected low-cost protocol elements. If the replacement still meets the safety threshold and association threshold and reduces the total cost, the replacement is accepted. Iteration continues until the total cost cannot be reduced in multiple consecutive iterations, at which point a locally optimal element combination is obtained.

[0054] Finally, the local optimal combination of elements is verified to satisfy the heritability, local maxima, and worst-case cost minimization of fuzzy matroids: the heritability verification ensures that there are no privacy conflicts or efficiency redundancies among the selected protocol elements; the local maxima verification ensures that no more protocol elements can be added without violating the constraints; the worst-case cost minimization verification ensures that the sum of the total computational cost and communication overhead reaches local optimum in the worst-case scenario.

[0055] Specifically, the interval weight matrix uses a pessimistic-optimistic comprehensive sorting method to establish a total order relation: for protocol element e i Interval weights, defining pessimistic comprehensive cost ,in These are the upper bounds of the computation cost interval and the upper bound of the communication overhead interval, respectively; when At that time, with an optimistic overall cost Sort by using it as the second key.

[0056] In one embodiment, step S130 described above may include steps S131 to S134.

[0057] S131. Based on the interval weight matrix, a pessimistic-optimistic comprehensive sorting method is used to establish a total order relation and determine the candidate element sequence.

[0058] In this embodiment, the selected set includes all protocol elements, each of which has its corresponding computational cost range, communication overhead range, security strength range, and semantic relevance range.

[0059] Candidate element sequence: These protocol elements are selected based on pessimistic comprehensive cost. Sort in ascending order, when At that time, based on the optimistic comprehensive cost Using ascending order as the second key, we obtain a sequence of candidate elements, which allows us to prioritize elements that still provide high privacy protection in the worst case.

[0060] S132. Based on the candidate element sequence, a greedy strategy is used to sequentially select protocol elements that satisfy the upper bound of the security strength interval not greater than a preset security threshold and the lower bound of the semantic relevance interval not less than a preset relevance threshold. The total computation cost and communication overhead of the currently selected element combination in the worst scenario are calculated to obtain the initial protocol element combination.

[0061] In this embodiment, a greedy strategy is employed: The system first selects protocol elements that meet the highest security standards while maintaining the lowest cost. Specifically, for each candidate element, it checks whether it meets preset security strength and semantic relevance thresholds.

[0062] Initial Protocol Element Combination: Once a suitable element is found, it is added to the initial combination, and the total computational cost and communication overhead of the combination in the worst-case scenario are calculated. This step aims to construct a preliminary, but optimized, chain of evidence.

[0063] S133. Based on the initial protocol element combination, element replacement and combination optimization are performed through iteration. Under the premise that the upper bound of the security strength interval is not greater than the preset security threshold and the lower bound of the semantic relevance interval is not less than the preset relevance threshold, the total computation cost and communication overhead in the worst scenario are continuously reduced. When the cost cannot be reduced after multiple iterations, the iteration is stopped, and a local optimal element combination is obtained.

[0064] In this embodiment, the iterative approach is as follows: attempt to replace the selected high-cost element with an unselected low-cost protocol element. If the new combination after replacement still meets the security and relevance requirements, and the total computational cost and communication overhead are reduced, then the replacement is accepted.

[0065] Specifically, when the total computational cost and communication overhead in the worst-case scenario cannot be reduced in 50 consecutive iterations, or the total cost reduction in a single iteration is less than 1%, the iteration is stopped to obtain a locally optimal combination of elements. Local optimal combination of elements: Continue iterating until the convergence condition is met (such as no improvement after 50 consecutive iterations or the total cost reduction is lower than a certain set value). The combination obtained at this time is the local optimum.

[0066] S134. Based on the local optimal element combination, verify whether it satisfies the heritability of the fuzzy independent set family and verify its local maxima. After the verification is passed, the robust minimum weight basis is obtained.

[0067] Based on the locally optimal element combination, verify whether it satisfies the heritability of the fuzzy independent set family, and verify its local maxima: for each element e in the selected combination B, verify... (Hereditary); Verify its local maxima: for Verify each element e' in the data. At least one of the following thresholds—security threshold, association threshold, or membership threshold—is violated; upon successful verification, the robust minimum weight basis is obtained. Robust minimum basis: The final chain of evidence not only satisfies all the above conditions, but also minimizes costs in the worst case, thus becoming the initial chain of evidence.

[0068] Through this series of steps, the method in this embodiment can effectively filter out high-quality chains of evidence, ensuring both information security and accuracy while maintaining computational efficiency, making it suitable for the needs of different research stages. This method is particularly well-suited for handling complex, multi-source, heterogeneous data environments, improving the performance of research agents in deep research tasks. S140. Analyze the research stage characteristics of the retrieval request to generate a dynamic adjustment factor. Based on the dynamic adjustment factor, adjust the membership threshold of the fuzzy independence constraint in the fuzzy matte model, redetermine the fuzzy independent set family, and perform multi-objective collaborative optimization on the preliminary evidence chain based on the adjusted fuzzy matte model to construct a correction parameter library containing multiple subspace correction parameters.

[0069] In this embodiment, the modified parameter library refers to a database constructed through a series of steps (S141-S149), which contains a set of parameters optimized for different research stages and needs. These parameters are used to adjust the combination of protocol elements in the initial chain of evidence to meet the accuracy, speed, and data dependency requirements in specific scenarios.

[0070] Specifically, the characteristics of the research stage to which the search request belongs are analyzed, and corresponding dynamic adjustment factors are set. Specifically: The research stage characteristics of the retrieval request are extracted using a pre-defined feature extraction model (based on keyword density, sentence complexity, and query depth level), and a feature vector is constructed. The aggressiveness value of the search request is analyzed, which is calculated by comprehensively considering indicators such as terminology specialization, query nesting depth, and logical connector density. The aggressiveness value is compared with a preset aggressiveness threshold to determine the adjustment direction: if the aggressiveness value is greater than the preset aggressiveness threshold, it indicates that the user or intelligent agent is in a deep exploration state, and the adjustment direction is set to prioritize response speed; if the aggressiveness value is less than or equal to the preset aggressiveness threshold, it indicates that the user or intelligent agent is in a cautious verification state, and the adjustment direction is set to prioritize calculation accuracy. The integrity of the data within the retrieval task is analyzed by calculating information entropy, the data quality score is calculated, and the adjustment weight of the data dependency target is determined. Combining the adjustment direction and adjustment weight, a corresponding dynamic adjustment factor is generated through a preset vector generation model.

[0071] Based on the aforementioned dynamic adjustment factor, the coordinates of each evidence element in the initial evidence chain are dynamically weighted and adjusted in the precision-speed-dependency multidimensional space. Constraints are then used to limit the range of coordinate adjustment, resulting in a dynamic reference point set. Drawing inspiration from the reference point distribution concept in multi-objective optimization, the evidence space is divided into multiple subspaces based on the distribution of each reference point in the dynamic reference point set. Each subspace corresponds to a combination of optimization objectives (e.g., high precision-low latency, high recall-strong diversity, strong privacy-high credibility).

[0072] Within each subspace, target weights for computational accuracy, response speed, and data dependency are set. A mapping relationship between correction parameters and optimization objectives is established by considering the difference between the initial evidence chain and the preset ideal evidence chain. Based on this mapping relationship and target weights, the Pareto optimal correction parameters for each subspace are solved using a multi-objective optimization algorithm (such as NSGA-III). Inflection points in the Pareto front are extracted as representative correction parameters, and the solved correction parameters are correlated to construct a correction parameter library.

[0073] In one embodiment, step S140 described above may include steps S141 to S149.

[0074] S141. Analyze the research stage characteristics of the retrieval request, extract corresponding features through a preset feature extraction model, construct a feature vector, and analyze the radicality value of the retrieval request; wherein, the research stage characteristics include the query depth, terminology density, and sentence complexity attributes of the retrieval request in the overview scanning, key point drilling, or logical deduction stages; the feature extraction model is a vector encoding network constructed based on query nesting depth, terminology specialization, and logical connector density, used to output the radicality value.

[0075] First, the system needs to analyze the research stage characteristics of the search requests. This includes attributes such as query depth, term density, and sentence complexity, specifically divided into three stages: overview scanning, drill-down of key points, or logical deduction.

[0076] Feature extraction model: This is a vector encoding network built based on query nesting depth, terminology specialization, and logical connector density, capable of outputting a value indicating the aggressiveness of the retrieval request. For example, in the "general overview" stage, the query is usually broad, with low terminology density and simple sentence structure; while in the "logical deduction" stage, the query is more in-depth, with specialized terminology and more logical connectors.

[0077] Specifically, the feature extraction model is a vector encoding network based on bidirectional LSTM and self-attention mechanism, including: an input layer that uses a domain pre-trained language model to embed words in the retrieval request, and an output dimension of n. 768; Bidirectional LSTM coding layer, containing two layers of bidirectional LSTM with 256 hidden units each, Dropout of 0.3, activation function of tanh, and output dimension n. 512; Self-attention pooling layer, using learnable attention vectors The LSTM output is weighted and summed, with attention weights applied. ,in h i The output of the LSTM at the i-th time step has a dimension of 512; the external feature concatenation layer concatenates the query nesting depth, terminology specialization, and logical connective density (three scalars normalized to Min-Max) with the attention pooling output, resulting in an output dimension of 515; the fully connected layer contains two fully connected network layers: the first layer maps the 515-dimensional code to 128-dimensional code using Linear+ReLU+BatchNorm, and the second layer maps it to 32-dimensional code using Linear+ReLU; the output layer outputs the aggression value r using Linear+Sigmoid. in [0,1].

[0078] Training configuration: Using a dataset labeled with research phases, the loss function is binary cross-entropy, the optimizer is Adam, and the learning rate is 10. -4 Batch size 64, early stop patience value 10.

[0079] Aggressiveness value: This value reflects the depth and accuracy requirements of the user's or agent's current query and serves as the basis for subsequent adjustments.

[0080] S142. Compare the aggression level value with the preset aggression threshold to determine the adjustment direction. If the aggression level value is greater than the preset aggression threshold, the adjustment direction is set to prioritize the response speed target. If the aggression level value is less than or equal to the preset aggression threshold, the adjustment direction is set to prioritize the calculation accuracy target.

[0081] The calculated degree of aggression is compared with a preset aggression threshold to determine the direction of adjustment.

[0082] If the aggressiveness value is greater than the preset threshold, it indicates that the user is in a deep exploration state. At this time, the adjustment direction is set to prioritize response speed.

[0083] If the aggressiveness value is less than or equal to the preset threshold, it indicates that the user is more concerned with prudent verification, and the adjustment direction is set to prioritize calculation accuracy.

[0084] S143. Analyze the integrity of the data within the retrieval request by calculating information entropy, calculate the data quality score, and determine the adjustment weight of the data dependency target.

[0085] By calculating and analyzing the completeness of the data within the retrieval request using information entropy, a data quality score is obtained.

[0086] Data quality score: This reflects whether the data involved in the search request is complete and whether there are any missing or redundant data. A high score means comprehensive data coverage, while a low score may indicate data gaps.

[0087] Adjusted weights for data dependency objectives: Based on the data quality score, determine the adjusted weights for data dependency objectives. For example, if there are significant missing data points, it may be necessary to increase the data dependency weights to ensure the reliability of the results.

[0088] S144. Combine the adjustment direction and adjustment weight to generate a dynamic adjustment factor vector; dynamically adjust the membership threshold of the fuzzy independence constraint according to the dynamic adjustment factor vector. When the adjustment direction prioritizes response speed, lower the membership threshold to allow more protocol element combinations to be selected. When the adjustment direction prioritizes computational accuracy, increase the membership threshold to filter higher quality protocol element combinations.

[0089] By combining the above adjustment directions and data dependency adjustment weights, a dynamic adjustment factor vector is generated.

[0090] Dynamically adjustable factor vector: This represents the different emphases on accuracy, speed, and data dependency in the context of the current query.

[0091] Membership threshold adjustment: The membership threshold of the fuzzy independence constraint is adjusted based on the dynamic adjustment factor vector. When the adjustment direction is speed priority, the membership threshold is lowered to allow more protocol elements to be selected; conversely, if the direction is accuracy priority, the membership threshold is raised to select higher quality protocol elements.

[0092] S145. Based on the adjusted fuzzy matte model, redetermine the fuzzy independent set family, and dynamically weight adjust the coordinates of each evidence element in the initial evidence chain in the precision-speed-dependency multidimensional space according to the redetermined fuzzy independent set family to obtain a dynamic reference point set.

[0093] Based on the adjusted fuzzy matte model, the fuzzy independent set family is redefined, and the coordinates of each evidence element in the preliminary evidence chain in the precision-speed-dependency multidimensional space are dynamically weighted and adjusted accordingly.

[0094] Fuzzy independent set family: refers to all protocol element combinations that meet the newly set membership threshold conditions.

[0095] Dynamic reference point set: The new set of reference points formed after weighted adjustment, representing the optimized distribution of the evidence chain.

[0096] Specifically, the candidate element set is first selected based on the newly determined fuzzy independent set family. For each evidentiary element in E', based on its original coordinates in the preliminary chain of evidence... Perform dynamic weighting: ,in These are the three-dimensional weights corresponding to the dynamically adjusted factor vector. S146. Cluster the distribution of reference points in the dynamic reference point set to obtain multiple clusters, each cluster corresponding to a subspace.

[0097] Cluster analysis is performed on the dynamic reference point set to form multiple clusters, each corresponding to a subspace.

[0098] Clustering: Each cluster represents a group of protocol elements with similar characteristics, which can be regarded as different combinations of optimization objectives (such as high precision-low latency, balanced, high speed-low precision).

[0099] S147. Analyze the data distribution characteristics in each cluster, determine the boundary requirements for computational accuracy, response speed and data dependency in the corresponding subspace, and construct constraints.

[0100] Analyze the data distribution characteristics of each cluster to determine the specific boundary requirements for computational accuracy, response speed, and data dependency in its corresponding subspace.

[0101] Boundary requirements: Set a clear target weight range for each subspace. For example, for a high-precision subspace, the weight of computational accuracy is higher, while the weight of response speed is lower.

[0102] S148. Based on the constraints, the optimal solution set of the corresponding subspace is obtained through multi-objective optimization.

[0103] Based on the above boundary requirements, the optimal solution set for the corresponding subspace is obtained by using a multi-objective optimization algorithm (such as NSGA-III).

[0104] Optimal solution set: The Pareto front in each subspace represents the optimal combination of protocol elements in that subspace.

[0105] S149. Select inflection points in different directions from the optimal solution set, identify the correction parameters of the corresponding subspace, the correction parameters include the membership threshold configuration and target weight configuration of the subspace, and construct the correction parameter library by associating the solved correction parameters.

[0106] Inflection points in different directions are selected from the set of optimal solutions in each subspace, and the correction parameters of the corresponding subspace are identified.

[0107] Correction parameters: These include the membership threshold configuration and target weight configuration corresponding to the subspace, which together determine how to optimize the evidence chain under specific conditions.

[0108] Correction Parameter Library: Finally, the correction parameters of all subspaces are integrated into a library for subsequent real-time matching, ensuring that the evidence chain can be efficiently adjusted under different query backgrounds, thereby improving retrieval accuracy and reasoning efficiency.

[0109] Specifically, the multi-objective optimization employs the NSGA-III algorithm, setting the population size to 60. Reference points are generated using the Dasand-Dennis structured generation method, taking four uniform points along each dimension of the 3D objective space to generate 15 reference points. The crossover operator uses simulated binary crossover with a distribution index of [missing information]. Crossover probability The mutation operator employs multinomial mutation, with a distribution index of... Probability of mutation The maximum number of iterations is 200, and convergence is determined by the relative change rate of the Hypervolume index being less than 10 for 20 consecutive iterations. -4 For the optimal solution set, the inflection point index is used. Identify inflection points in different directions and select... The K smallest solutions are used as representative correction parameters, where Let k be the objective function value. The coordinates are the ideal point coordinates. The correction parameters include the membership threshold configuration and target weight configuration corresponding to the subspace, and the correction parameter library is constructed by associating the solved correction parameters.

[0110] Through this series of steps, the system can adaptively adjust its evidence chain construction strategy according to the user's actual needs, significantly improving the quality and efficiency of retrieval and reasoning, especially when dealing with complex, multi-source, heterogeneous data. This approach not only enhances the system's flexibility and adaptability but also provides strong support for automated deep research.

[0111] S150. Match the target correction parameter group from the correction parameter library according to the real-time characteristics of the retrieval request, and correct the initial evidence chain in real time to obtain the optimized evidence chain.

[0112] In this embodiment, the target correction parameter set refers to a set of the most suitable adjustment parameters selected from the correction parameter library based on the specific needs and context of the current retrieval request. These parameters are used to dynamically adjust the combination of protocol elements in the initial evidence chain to ensure that the accuracy, speed, and data dependency requirements under specific scenarios are met.

[0113] Specifically, based on the real-time strategy pattern of the retrieval request, an additive multi-attribute utility model is used to calculate the utility value: in For calculation accuracy, For response time, Data coverage; For single-attribute exponential utility functions: , Risk attitude coefficient and When the real-time strategy mode is fast response mode, the weight configuration is w=[0.2,0.6,0.2]; when it is high precision mode, the weight configuration is w=[0.6,0.2,0.2]; when it is balanced mode, the weight configuration is w=[1 / 3,1 / 3,1 / 3]; select the parameter combination with the highest utility value as the target correction parameter group.

[0114] In one embodiment, step S150 described above may include: Based on the real-time characteristics of the retrieval request, a retrieval feature vector is constructed; wherein, the real-time characteristics include an instantaneous attribute vector of the current retrieval request's data size, modality type, response time limit, and system resource occupancy status.

[0115] The correction parameter information corresponding to different subspaces is analyzed in the correction parameter library to generate corresponding subspace fingerprint vectors. Calculate the similarity between the retrieved feature vector and the subspace fingerprint vector, and filter out correction parameters with similarity greater than a preset similarity threshold as a set of candidate parameters; Based on the real-time strategy mode of the retrieval request, the utility value of different parameter combinations in the candidate parameter set is calculated, and the parameter combination with the highest utility value is selected as the target correction parameter group.

[0116] Based on the target correction parameter set, extract the corresponding accuracy-speed-dependency target weight, membership threshold, and evidence element screening threshold; Based on the target weight, the coordinates of each evidence element in the initial evidence chain are re-weighted and mapped in the precision-speed-dependency multidimensional space to obtain the re-mapped set of evidence coordinates. Based on the membership threshold, the fuzzy independence verification is performed again on each combination of evidence elements in the remapped evidence coordinate set, and evidence elements with a membership degree lower than the membership threshold or that cause logical redundancy are removed to obtain a filtered evidence element set. Based on the evidence element screening threshold, the inference coefficient of each evidence element in the filtered evidence element set is recalculated, the core evidence elements with inference coefficients greater than the screening threshold are retained, and the comprehensive inference weight of each evidence element is configured accordingly. The core evidence elements are sorted and reorganized according to the comprehensive reasoning weights to obtain the optimized evidence chain.

[0117] In this embodiment, the system first needs to construct a retrieval feature vector based on the real-time characteristics of the retrieval request.

[0118] Real-time features: These are instantaneous attribute vectors that include the data size of the current retrieval request (e.g., how many records are involved), modality type (text, image, time series, etc.), response time (expected return time), and system resource usage status (e.g., CPU utilization, memory usage).

[0119] Feature vector retrieval: The above real-time features are quantized and encoded into vector form for subsequent comparison with subspace fingerprint vectors in the correction parameter library.

[0120] The correction parameter information corresponding to different subspaces is analyzed in the correction parameter library, and the corresponding subspace fingerprint vector is generated.

[0121] Subspace fingerprint vector: Each subspace has its unique configuration of correction parameters, which reflect the specific weighting of computational accuracy, response speed, and data dependency in that subspace. The fingerprint vector is a numerical representation of these parameters, facilitating rapid comparison.

[0122] Calculate the similarity between the retrieved feature vector and the fingerprint vector of each subspace, and select the correction parameters with similarity greater than a preset threshold (such as 0.85) as the candidate parameter set.

[0123] Similarity calculation: Cosine similarity or other appropriate metrics can be used to assess the degree of similarity between two vectors.

[0124] Candidate parameter set: consists of subspace correction parameters that closely match the characteristics of the current retrieval request.

[0125] Based on the real-time strategy pattern of the retrieval request (e.g., speed priority or accuracy priority), the utility value of different parameter combinations in the candidate parameter set is calculated, and the parameter combination with the highest utility value is selected as the target correction parameter group.

[0126] Utility value calculation: Taking into account factors such as accuracy, speed, and dependency, the expected performance score of each parameter combination is given in the current query context.

[0127] Target adjustment parameter set: The final set of parameters, which includes the most suitable precision-speed-dependency weight configuration, membership threshold, and evidence element screening threshold for the current retrieval task.

[0128] Based on the target correction parameter set, the coordinates of each evidence element in the initial evidence chain are re-weighted and mapped in the multi-dimensional space of precision-speed-dependency, resulting in the re-mapped set of evidence coordinates.

[0129] Reweighting: Adjusting the position of each evidence element according to the new weighting scheme to reflect its relative importance under the new conditions.

[0130] Fuzzy independence verification: Based on the membership threshold, check whether the combination of each evidence element in the remapped evidence coordinate set meets the logical consistency and privacy security constraints, and remove evidence elements that do not meet the conditions.

[0131] Based on the evidence element screening threshold, the inference coefficient of each evidence element in the filtered evidence element set is recalculated, the core evidence elements with inference coefficients greater than the screening threshold are retained, and the comprehensive inference weight of each evidence element is configured accordingly.

[0132] Inference coefficient calculation: Usually derived from the attention mechanism in the inference analysis neural network, it reflects the relevance and contribution of the evidence elements to the conclusion derivation.

[0133] Comprehensive reasoning weight configuration: Combine the weight correction coefficients in the target correction parameter group to determine the importance ranking of each core evidence element in the entire evidence chain.

[0134] The core evidence elements are sorted and reorganized according to the comprehensive reasoning weights to form an optimized chain of evidence.

[0135] Optimized evidence chain: After the above series of processing steps, a more refined, efficient and logically rigorous evidence chain is formed, which can better support the subsequent knowledge retrieval and reasoning process.

[0136] Specifically, the inference coefficients are calculated using a graph attention network, which includes: a graph attention layer employing the GATv2 multi-head attention mechanism with 8 heads, each head having an output dimension of 64, resulting in a total output dimension of 512, and attention coefficients. ,in For learnable weight matrix, As a learnable attention vector, x iThe initial feature vector for protocol element i is given by LeakyReLU with a negative slope of 0.2. The Transformer encoding layer consists of two layers with a hidden dimension of 512, 8 heads, a feedforward dimension of 2048, and a Dropout value of 0.1. The input sequence begins with the [CLS] marker and concatenates the features of each evidence element. The [CLS] output is then used as the global representation. The output layer concatenates the evidence element features with the global representation. The inference coefficients are output via Linear+Sigmoid. ; S160. Obtain target knowledge data based on the optimized evidence chain and configure reasoning weights to generate reasoning response information.

[0137] In one embodiment, step S160 described above may include steps S161 to S163.

[0138] S161. Obtain target knowledge data matching the target knowledge node from the knowledge base according to the target knowledge node corresponding to the optimized evidence chain.

[0139] Based on the target knowledge nodes corresponding to the optimized evidence chain, target knowledge data matching these nodes is obtained from the knowledge base.

[0140] Target knowledge nodes: These refer to the evidence points that have been screened and verified and are considered the most critical.

[0141] Target knowledge data: Contains all necessary information related to these nodes, such as text descriptions, charts, data analysis results, etc.

[0142] S162. Based on the inference coefficients corresponding to each evidence element in the optimized evidence chain, and combined with the weight correction coefficients in the target correction parameter group, the comprehensive inference weight of each target knowledge data is calculated through weighted fusion.

[0143] Based on the inference coefficients corresponding to each evidence element in the optimized evidence chain, and combined with the weight correction coefficients in the target correction parameter group, the comprehensive inference weight of each target knowledge data is calculated through weighted fusion.

[0144] Inference coefficient: measures the magnitude of the role of each element of evidence in deriving the conclusion.

[0145] Weighting adjustment coefficient: Adjusts the importance ratio of each dimension (accuracy, speed, dependency) according to the real-time strategy mode.

[0146] Comprehensive reasoning weight: The final score, which combines the reasoning coefficient and the weighting correction coefficient, determines the order and display priority of each knowledge data in the final answer.

[0147] Specifically, the weighted fusion formula is as follows: , where r e The inference coefficients are... These are the weight correction coefficients in the target correction parameter set. , .

[0148] S163. According to the preset response generation rules, the target knowledge data with completed reasoning weight configuration is organized into a structured text containing conclusions, evidence support and logical deduction process, and the reasoning response information is generated.

[0149] Based on the preset response generation rules, the target knowledge data with completed reasoning weight configuration is organized into structured text containing conclusions, supporting evidence, and logical deduction processes to generate reasoning response information.

[0150] Response generation rules: Define how to integrate various types of knowledge data into an easy-to-understand and traceable format, which may include specifications for headings, paragraph divisions, citations, etc.

[0151] Reasoning response information: The final output document not only provides clear conclusions, but also includes detailed supporting evidence and logical reasoning paths, ensuring that the entire analysis process is transparent and traceable.

[0152] Through this series of steps, the system can flexibly adjust its evidence chain construction strategy when faced with complex and ever-changing research queries, significantly improving retrieval efficiency and reasoning accuracy, thereby more effectively supporting automated in-depth research tasks.

[0153] Based on the real-time characteristics of the retrieval request, a target set of corrected parameters is matched in the corrected parameter library. Specifically: a retrieval feature vector is constructed; fingerprint vectors for each subspace are generated in the corrected parameter library; the similarity between the retrieval feature vector and the subspace fingerprint vector is calculated, and corrected parameters with similarity greater than a preset similarity threshold are selected as a set of candidate parameters; based on the real-time strategy mode of the retrieval request, the utility value of different parameter combinations in the candidate parameter set is calculated, and the parameter combination with the highest utility value is selected as the target set of corrected parameters.

[0154] The initial evidence chain is revised in real time to obtain an optimized evidence chain. Target knowledge data is retrieved from the knowledge base based on the target knowledge nodes corresponding to the optimized evidence chain. The comprehensive reasoning weights for each target knowledge data are configured according to the inference coefficients of each evidence element in the evidence chain (calculated using the attention mechanism in the inference analysis neural network) and the weight correction coefficients in the target correction parameter group.

[0155] Based on preset response generation rules, the target knowledge data with configured reasoning weights is organized into structured text containing conclusions, supporting evidence, and logical derivation processes to generate reasoning response information. During the generation process, data with high comprehensive reasoning weights and security strength meeting preset security thresholds are prioritized as core arguments, while data with low comprehensive reasoning weights are used as supplementary explanations or background information to ensure that the response information is clear in structure, highly credible, and has strong traceability.

[0156] For example, taking the medical research agent performing the retrieval task of "analyzing the latest advances in targeted therapy for lung cancer and the evidence of its correlation with imaging" as an example, the execution process of the present invention will be explained in detail.

[0157] In step S110, the system receives the search request "Analyze the latest advances in targeted therapy for lung cancer and the evidence linking it to imaging." The feature extraction module extracts the request features using a pre-trained language model in the medical field (such as PubMedBERT), obtaining a keyword set {lung cancer, targeted therapy, imaging, latest advances} and the corresponding 768-dimensional semantic vector. The system searches the knowledge graph to obtain the target graph node "targeted therapy for non-small cell lung cancer" and its graph association information, including adjacent nodes such as "EGFR mutation," "osimertinib," and "CT imaging features," as well as the relationship types of the connecting edges such as "indication," "imaging manifestations," and "resistance mechanism."

[0158] The system defines each piece of evidence in the retrieved information as a protocol element and constructs a set of protocol elements: e1: Text modality, a paragraph from a tertiary hospital's 2025 clinical guidelines for targeted therapy of lung cancer, containing the entity "EGFRT790M mutation"; e2: Image modality, ROI region of lesion in lung CT image, including the imaging feature "ground-glass nodule with lobulation sign"; e3: Spectral modality, drug-target interaction ternary group (osimertinib-EGFR T790M-inhibition). e4: Temporal modality, the gene sequencing temporal signal of a patient during the treatment process in 2024-2025.

[0159] In step S120, the system constructs a fuzzy matroid model based on the protocol element set. The protocol elements are used as the base set elements, and the differential privacy protection strength, computational efficiency, and logical redundancy of the evidence chain are used as fuzzy independence constraints. A membership function is used to characterize the degree to which each candidate subset satisfies the fuzzy independence constraints: for a candidate subset S, its membership value... Considering the privacy budget superposition bias of elements within the quantum set, the probability of computational resource conflicts, and semantic redundancy, the value range is [0, 1]. A preset membership threshold is set. =0.80, only The candidate subset is included in the fuzzy independent set family \mathcal{I}_f.

[0160] Meanwhile, an interval weight matrix is ​​defined for each protocol element to characterize the range of parameter fluctuations under uncertain scenarios.

[0161] Table 1. Range of parameter fluctuations under uncertainty scenarios

[0162] The fuzzy independent set family I f The fuzzy matroid model is formed by the interval weight matrix.

[0163] In step S130, the system selects a combination of protocol elements in the worst-case scenario based on the fuzzy matroid model using a robust minimum basis algorithm. A preset security threshold ε is set. max =0.05, the preset association threshold is 0.80.

[0164] First, based on the interval weight matrix, a total order relation is established using a pessimistic-optimistic comprehensive sorting method, and the candidate element sequence is determined: e2 (0.010), e4 (0.012), e1 (0.015), e3 (0.030). A greedy strategy is used to select elements sequentially: e2 is selected because the upper bound of the safety strength interval is 0.010 ≤ 0.05, and the lower bound of the semantic relevance interval is 0.80 ≥ 0.80, satisfying the constraints; in the worst-case scenario, the computational cost is 350, the communication overhead is 250, and the cumulative cost is 600; e4 is selected because the upper bound of the safety strength interval is 0.012 ≤ 0.05, and the lower bound of the semantic relevance interval is 0.85 ≥ 0.80, satisfying the constraints; in the worst-case scenario, the cost is 600 + 600 = 1200; e1 is selected because the upper bound of the safety strength interval is 0.015. The lower bound of the semantic relevance interval is 0.90 ≥ 0.80, satisfying the constraints; the worst-case cost is 1200 + 120 = 1320; e3 is attempted: the upper bound of the security strength interval is 0.030 ≤ 0.05, and the lower bound of the semantic relevance interval is 0.85 ≥ 0.80, satisfying the single constraint; however, the upper bound of the cumulative security strength interval is 0.010 + 0.012 + 0.015 + 0.030 = 0.067 > 0.05, therefore e3 is determined to conflict with the selected combination under the fuzzy independence constraint, and is skipped. The initial protocol element combination {e2, e4, e1} is obtained. Entering iterative optimization: Trying to replace e1 with e3, the worst-case cost becomes 350 + 600 + 180 = 1130 << 1320, but the upper bound of the cumulative safety strength interval (0.010 + 0.012 + 0.030 = 0.052 > 0.05) does not meet the safety constraint; trying to replace e2 with e3, the worst-case cost becomes 120 + 600 + 180 = 900 << 1320, but the upper bound of the cumulative safety strength interval (0.015 + 0.012 + 0.030 = 0.057 > 0.05) also does not meet the constraint. No better replacement is found, so iteration stops. Verify the locally optimal element combination {e1,e2,e4}: Calculate its membership value μ({e1,e2,e4}) = 0.85 ≥ 0.80, satisfying the fuzzy independence constraint; verify the heritability: μ({e2,e4}) = 0.88 ≥ 0.80, μ({e1,e4}) = 0.86 ≥ 0.80, μ({e1,e2}) = 0.83 ≥ 0.80, all satisfying the constraint. Obtain the robust minimum weight basis {e1,e2,e4} as the initial evidence chain.

[0165] In step S140, the system analyzes the research stage characteristics of the retrieval request to generate a dynamic adjustment factor. These research stage characteristics include query depth, terminology density, and sentence complexity attributes at the overview scanning, key point drilling, or logical deduction stages of the retrieval request. Corresponding features are extracted using the feature extraction model—a vector encoding network built based on query nesting depth, terminology specialization, and logical connector density—to construct a feature vector. The retrieval request contains in-depth exploration terms such as "latest progress" and "related evidence," and the query nesting depth is two levels, with an aggressiveness value calculated as 0.75. The preset aggressiveness threshold is 0.60; since 0.75 > 0.60, the adjustment direction is set to prioritize response speed.

[0166] The system calculates the completeness of data within the retrieval request using information entropy: e1, e2, and e4 originate from three different data sources, indicating complete data coverage and an information entropy score of 0.82. The adjustment weight for the data dependency target is determined to be 0.30. Combining the adjustment direction [speed priority] with the adjustment weight of 0.30, a dynamic adjustment factor vector [0.8, 0.2, 0.3] (corresponding to speed, accuracy, and dependency dimensions, respectively) is generated. This dynamic adjustment factor vector is then normalized to a sum of 1 using Min-Max before use. Let the original vector be... Normalized weights The normalized weights are [0.615, 0.154, 0.231].

[0167] Among them, the weight of the velocity dimension Precision dimension weight Dependency dimension weight Before being used for coordinate weighting, the three-dimensional weights are normalized to a sum of 1 using Min-Max to obtain normalized weights [0.615, 0.154, 0.231].

[0168] The membership threshold of the fuzzy independence constraint is dynamically adjusted based on the dynamic adjustment factor vector: since the adjustment direction prioritizes response speed, the system lowers the membership threshold to expand the fuzzy independent set family, setting the membership threshold from... 0=0.80 lowered to 1 = 0.65. Based on the adjusted membership threshold, the system redetermines the fuzzy independent set family I. f This allows for the inclusion of more protocol element combinations in subsequent optimizations.

[0169] In step S150, the system redetermines the fuzzy independent set family I based on the adjusted fuzzy matroid model. fThe coordinates of each evidence element in the initial evidence chain are dynamically weighted and adjusted in the precision-speed-dependency multidimensional space based on the newly determined fuzzy independent set family. Assuming the original coordinates of e1 are (0.9, 0.6, 0.7), after adjustment with the target weight [0.2, 0.6, 0.2], they become (0.18, 0.48, 0.21); e2 is adjusted to (0.16, 0.40, 0.24); and e4 is adjusted to (0.18, 0.72, 0.27). This yields a dynamic reference point set.

[0170] K-means clustering (k=3) is performed on the dynamic reference point set to obtain 3 clusters, corresponding to 3 subspaces: Subspace A: high precision-low speed (suitable for careful verification); Subspace B: balanced (suitable for routine analysis); Subspace C: high speed-low precision (suitable for rapid scanning).

[0171] Standardized Euclidean distance is used as the distance metric: ,in Let k be the standard deviation of the k-th dimension across all reference points; initialization uses the K-means++ algorithm to ensure that the initial distance between centroids is maximized; the convergence condition is the change in centroid coordinates. Or it can reach the maximum number of iterations, 100.

[0172] This search request, prioritizing speed, falls into subspace C. The data distribution characteristics of subspace C are analyzed to determine its boundary requirements for computational accuracy, response speed, and data dependency, thus constructing constraints. The optimal solution set for subspace C is solved using the NSGA-III algorithm, extracting the inflection point on the Pareto front and identifying correction parameters: accuracy weight 0.2, speed weight 0.6, dependency weight 0.2, and membership threshold configuration 0.65. The solved correction parameters are then correlated to construct the correction parameter library.

[0173] In step S160, the system matches a target set of correction parameters from the correction parameter library based on the real-time characteristics of the retrieval request. The real-time characteristics include an instantaneous attribute vector of the current retrieval request's data size (large), modality type (multimodal), response time (milliseconds), and system resource usage status (CPU usage 60%). A retrieval feature vector vq is constructed as follows: [Large-scale, multimodal, low latency, medium load].

[0174] The correction parameter information corresponding to different subspaces is analyzed in the correction parameter library to generate a subspace fingerprint vector. The subspace fingerprint vector f k Generated directly from the correction parameters: ,in, This is the membership threshold; The target weight; Filter thresholds for evidence elements; This is the upper bound of the average privacy budget within the cluster; The average pessimistic overall cost within the cluster; Encode the primary modality type within the cluster (one-hot vector, e.g., [1,0,0,0] represents text-dominant). The retrieved feature vector... Using the same dimension encoding, cosine similarity and Compare and filter subspaces with similarity greater than a preset similarity threshold (e.g., 0.85).

[0175] The similarity between the retrieved feature vector and the fingerprint vector of subspace C is calculated to be 0.91, which is greater than the preset similarity threshold of 0.85. Therefore, the modified parameters of subspace C are included in the candidate parameter set. Based on the real-time strategy mode (fast response mode) of the retrieval request, the utility values ​​of different parameter combinations in the candidate parameter set are calculated. The parameter combination with the highest utility value is selected as the target modified parameter group, using an additive multi-attribute utility model. ,in In step S170, the system performs real-time correction on the initial evidence chain according to the target correction parameter group to obtain the optimized evidence chain.

[0176] The correction parameters for subspace C are: target weight (0.2, 0.6, 0.2), membership threshold 0.65, and evidence element screening threshold delta=0.50.

[0177] Extraction parameters: target weights (0.2, 0.6, 0.2), membership threshold 0.65, and evidence element selection threshold. =0.50; Coordinate remapping: Based on the target weight, the coordinates of each evidence element in the initial evidence chain {e1, e2, e4} in the precision-speed-dependency multidimensional space are re-weighted and mapped to obtain the remapped set of evidence coordinates. Fuzzy independence verification: Based on a membership threshold of 0.65, fuzzy independence verification is performed again on each combination of evidence elements in the remapped evidence coordinate set. The membership value of {e1, e2, e4} is verified to be 0.85. 0.65, which satisfies the constraints; there is no logically redundant evidence, resulting in the filtered set of evidence elements {e1, e2, e4}; Filtering based on a threshold: =0.50, and the inference coefficients for the filtered evidence elements are recalculated: e2 is 0.88, e4 is 0.82, and e1 is 0.45. These inference coefficients are calculated using a graph attention network: the evidence element features are input into a GATv2 graph attention layer with 8 heads and an output dimension of 512, and then the inference coefficients are calculated. Aggregate neighborhood information; then extract global features through a 2-layer Transformer encoder, take the [CLS] label output and concatenate it with the evidence element features, and obtain the inference coefficients through a fully connected layer and a Sigmoid function; Core evidence elements e2 and e4 with inference coefficients greater than 0.50 are retained, and corresponding comprehensive inference weights are assigned: e2 is 0.95 and e4 is 0.85; e1 is downgraded to background supplement because its inference coefficient is 0.45 < 0.50. Reordering and Reorganization: The core evidence elements are ordered and reorganized according to the comprehensive reasoning weight to obtain the optimized evidence chain B*=[e2, e4, e1].

[0178] In step S180, the system acquires target knowledge data and configures inference weights based on the optimized evidence chain, generating inference response information. Based on the target knowledge node corresponding to the optimized evidence chain, the system retrieves target knowledge data matching the target knowledge node from the knowledge base. Based on the inference coefficients corresponding to each evidence element in the optimized evidence chain, combined with the weight correction coefficients in the target correction parameter group, the system calculates the comprehensive inference weights of each target knowledge data through weighted fusion: e2 is 0.95, e4 is 0.85, and e1 is 0.30.

[0179] Based on the pre-defined response generation rules, the target knowledge data with configured reasoning weights is organized into structured text containing conclusions, supporting evidence, and logical derivation processes, generating the following reasoning response: "Conclusion: This case is positive for the EGFR T790M mutation, and imaging shows typical ground-glass nodules with lobulation. Based on dynamic monitoring data of gene sequencing from 2024 to 2025, osimertinib is recommended as a first-line treatment. Supporting evidence: 1. CT image ROI analysis (weight 0.95); 2. Gene sequencing time-series trend (weight 0.85); 3. Clinical guideline background (weight 0.30)." Suggested preset parameter settings: Safety threshold max： Healthcare: 0.05; Finance: 0.01; Manufacturing: 0.10. Association threshold: 0.80 (general), 0.85 (high precision mode); Aggressive threshold: 0.60; Initial membership threshold 0:0.80 (general), can be adjusted down to 0.65 in speed priority mode, and up to 0.90 in precision priority mode; Similarity threshold (corrected parameter matching): 0.85; The stopping condition for robust minimum weight basis iteration is: no improvement for 50 consecutive iterations or the total cost reduction in the worst-case scenario.

[0180] By modeling the evidence chain construction problem as a fuzzy matroid minimum weight basis problem, the evidence cost is minimized under differential privacy and semantic relevance constraints. This effectively avoids the problems of redundant evidence stacking and low-quality evidence mixing in traditional retrieval, and minimizes the computational cost of the evidence chain while ensuring logical completeness.

[0181] By introducing a multi-objective collaborative optimization and dynamic adjustment factor mechanism, the system can adaptively adjust the weight distribution of evidence in the multi-dimensional space of accuracy, speed, and dependence according to the characteristics of the research stage to which the retrieval request belongs (overview scanning, key point drilling, logical deduction). This enables the evidence chain construction strategy to dynamically adapt to the needs of different research scenarios, significantly improving the conclusion accuracy and response speed of the research agent in multi-hop reasoning and fact-checking scenarios.

[0182] By constructing a correction parameter library and a real-time matching mechanism, the system achieves online correction and dynamic optimization of the evidence chain, ensuring high retrieval accuracy and inference quality even in cold starts or new scenarios. The final generated inference response information is organized through a dual weighting mechanism (inference weights reflect the relevance between evidence and inference categories, while correction parameters reflect the dynamic needs of the research stage). The output conclusions possess professional characteristics such as clear hierarchy, sufficient evidence, logical consistency, and traceability, effectively supporting the needs of complex tasks such as automated in-depth research.

[0183] The aforementioned high-precision reasoning retrieval method based on multi-objective collaboration obtains retrieval requests and extracts relevant evidence units from the knowledge graph, defining them as protocol elements and constructing a protocol element set. Based on this set, a fuzzy matte model is established, and a robust minimum basis algorithm is used to select the protocol element combination with the lowest total cost while satisfying preset security and association thresholds, forming a preliminary evidence chain, effectively avoiding the problem of redundant evidence chain construction. Subsequently, the research stage characteristics of the retrieval request are analyzed to generate a dynamic adjustment factor. Based on this factor, the membership degree threshold of the fuzzy independence constraint in the fuzzy matte model is adjusted, the fuzzy independent set family is redefined, and the preliminary evidence chain is optimized through multi-objective collaboration, constructing a correction parameter library. This solves the problems of rigid evidence weight allocation and difficulty in balancing reasoning accuracy and computational efficiency. Finally, the optimal correction parameter set is matched according to the real-time characteristics of the retrieval request to perform real-time correction on the preliminary evidence chain, obtaining an optimized evidence chain. Based on this, target knowledge data is obtained to configure reasoning weights and generate response information, achieving synergistic optimization of privacy and efficiency, and improving the performance of the research agent in complex query scenarios. This series of steps ensures a logically complete chain of evidence with minimal computational overhead, while adaptively adjusting evidence weights to meet the needs of different research stages, thereby enhancing the system's response speed and the accuracy of its conclusions.

[0184] Figure 3This is a schematic block diagram of a high-precision reasoning retrieval system 300 based on multi-target collaboration provided in an embodiment of the present invention. Figure 3 As shown, corresponding to the above-described high-precision reasoning retrieval method based on multi-objective collaboration, the present invention also provides a high-precision reasoning retrieval system 300 based on multi-objective collaboration. This high-precision reasoning retrieval system 300 based on multi-objective collaboration includes a unit for executing the above-described high-precision reasoning retrieval method based on multi-objective collaboration, and the system can be configured in a server. Specifically, please refer to... Figure 3 The high-precision reasoning retrieval system 300 based on multi-target collaboration includes an acquisition unit 301, a construction unit 302, a preliminary selection unit 303, a correction parameter library establishment unit 304, a correction unit 305, and a generation unit 306.

[0185] Acquisition unit 301 is used to acquire a retrieval request, retrieve retrieval information from the knowledge graph based on the retrieval request, define each evidence unit in the retrieval information as a protocol element, and construct a protocol element set; Construction unit 302 is used to construct a fuzzy matte model based on the protocol element set; Preliminary selection unit 303 is used to select, based on the fuzzy matte model, the protocol element combination that satisfies the preset security threshold and association threshold and has the minimum total cost in the worst-case scenario using a robust minimum basis solving algorithm, to obtain a robust minimum basis and form a preliminary evidence chain; Correction parameter library establishment unit 304 is used to analyze the research stage characteristics of the retrieval request to generate A dynamic adjustment factor is used to adjust the membership threshold of the fuzzy independence constraint in the fuzzy matte model, redetermine the fuzzy independent set family, and perform multi-objective collaborative optimization on the initial evidence chain based on the adjusted fuzzy matte model to construct a correction parameter library containing multiple subspace correction parameters; a correction unit 305 is used to match the target correction parameter group from the correction parameter library according to the real-time features of the retrieval request, and perform real-time correction on the initial evidence chain to obtain an optimized evidence chain; a generation unit 306 is used to obtain target knowledge data and configure inference weights according to the optimized evidence chain to generate inference response information.

[0186] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the above-mentioned high-precision reasoning retrieval system 300 based on multi-target collaboration and each unit can be referred to the corresponding description in the foregoing method embodiments. For the sake of convenience and brevity, it will not be repeated here.

[0187] The aforementioned high-precision reasoning retrieval system 300 based on multi-target collaboration can be implemented as a computer program, which can be used in, for example... Figure 3 It runs on the computer device shown.

[0188] Please see Figure 3 , Figure 3 This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 500 can be a server, wherein the server can be a standalone server or a server cluster composed of multiple servers.

[0189] See Figure 3 The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0190] The non-volatile storage medium 503 may store an operating system 5031 and a computer program 5032. The computer program 5032 includes program instructions that, when executed, cause the processor 502 to perform a high-precision reasoning retrieval method based on multi-target collaboration.

[0191] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0192] The internal memory 504 provides an environment for the execution of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute a high-precision reasoning retrieval method based on multi-target collaboration.

[0193] This network interface 505 is used for network communication with other devices. Those skilled in the art will understand that... Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 500 to which the present application is applied. The specific computer device 500 may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0194] The processor 502 is used to run the computer program 5032 stored in the memory to implement all the steps of the high-precision reasoning retrieval method based on multi-target collaboration.

[0195] It should be understood that in the embodiments of this application, the processor 502 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0196] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A high-precision reasoning and retrieval method based on multi-objective collaboration, characterized in that, include: Obtain a retrieval request, retrieve retrieval information from the knowledge graph based on the retrieval request, define each evidence unit in the retrieval information as a protocol element, and construct a protocol element set; A fuzzy matroid model is constructed based on the set of protocol elements; Based on the fuzzy matroid model, the robust minimum weight basis is selected in the worst-case scenario by using the robust minimum weight basis solution algorithm to obtain the protocol element combination that satisfies the preset security threshold and association threshold and has the minimum total cost, thus forming the initial evidence chain; The research stage characteristics of the search request are analyzed to generate a dynamic adjustment factor. Based on the dynamic adjustment factor, the membership threshold of the fuzzy independence constraint in the fuzzy matroid model is adjusted to redetermine the fuzzy independent set family. Based on the adjusted fuzzy matroid model, the preliminary evidence chain is optimized in a multi-objective collaborative manner to construct a correction parameter library containing multiple subspace correction parameters. Based on the real-time characteristics of the search request, the target correction parameter group is matched from the correction parameter library to correct the initial evidence chain in real time, thereby obtaining the optimized evidence chain. Based on the optimized evidence chain, target knowledge data is obtained and inference weights are configured to generate inference response information.

2. The high-precision reasoning retrieval method based on multi-objective collaboration according to claim 1, characterized in that, The construction of the fuzzy matroid model based on the protocol element set includes: Using protocol elements as the base set, and differential privacy protection strength, computational efficiency, and logical redundancy as fuzzy independence constraints, a membership function is used to characterize the degree to which each candidate subset satisfies the fuzzy independence constraints, and a fuzzy independent set family is determined based on a preset membership threshold. Define the computational cost interval, communication overhead interval, security strength interval, and semantic relevance interval for each protocol element to obtain the interval weight matrix; The fuzzy matroid model is formed by the fuzzy independent set family and the interval weight matrix.

3. The high-precision reasoning retrieval method based on multi-objective collaboration according to claim 2, characterized in that, The step involves selecting a combination of protocol elements that satisfies a preset security threshold and a correlation threshold while minimizing total cost in the worst-case scenario based on the fuzzy matte model using a robust minimum basis solution algorithm, thereby obtaining a robust minimum basis and forming a preliminary evidence chain, including: Based on the interval weight matrix, a pessimistic-optimistic comprehensive sorting method is used to establish a total order relation and determine the candidate element sequence; Based on the candidate element sequence, a greedy strategy is used to sequentially select protocol elements that satisfy the upper bound of the security strength interval not greater than the preset security threshold and the lower bound of the semantic relevance interval not less than the preset relevance threshold. The total computation cost and communication overhead of the currently selected element combination in the worst scenario are calculated to obtain the initial protocol element combination. Based on the initial protocol element combination, element replacement and combination optimization are performed iteratively. Under the premise that the upper bound of the security strength interval is not greater than the preset security threshold and the lower bound of the semantic relevance interval is not less than the preset relevance threshold, the total computation cost and communication overhead in the worst scenario are continuously reduced. When the total computation cost and communication overhead in the worst scenario cannot be reduced in 50 consecutive iterations, or the total cost reduction in a single iteration is less than 1%, the iteration is stopped, and a locally optimal element combination is obtained. Based on the locally optimal combination of elements, verify whether it satisfies the heritability of the fuzzy independent set family, and verify its local maxima. If the verification is successful, the robust minimum weight basis is obtained.

4. The high-precision reasoning and retrieval method based on multi-objective collaboration according to claim 3, characterized in that, The analysis of the research stage characteristics of the retrieval request generates a dynamic adjustment factor, and the membership threshold of the fuzzy independence constraint in the fuzzy matroid model is adjusted based on the dynamic adjustment factor, including: The research stage characteristics of the retrieval request are analyzed, and corresponding features are extracted through a preset feature extraction model to construct a feature vector and analyze the radicality value of the retrieval request. The research stage characteristics include the query depth, terminology density, and sentence complexity attributes of the retrieval request in the overview scanning, key point drilling, or logical deduction stages. The feature extraction model is a vector encoding network constructed based on query nesting depth, terminology specialization, and logical connector density, used to output the radicality value. The aggressiveness value is compared with a preset aggressiveness threshold to determine the adjustment direction. If the aggressiveness value is greater than the preset aggressiveness threshold, the adjustment direction is set to prioritize the response speed target. If the aggressiveness value is less than or equal to the preset aggressiveness threshold, the adjustment direction is set to prioritize the calculation accuracy target. The integrity of the data within the retrieval request is analyzed by calculating information entropy, a data quality score is calculated, and the adjustment weight of the data dependency target is determined. By combining the adjustment direction and adjustment weight, a dynamic adjustment factor vector is generated. The membership threshold of the fuzzy independence constraint is dynamically adjusted according to the dynamic adjustment factor vector. When the adjustment direction prioritizes response speed, the membership threshold is lowered to allow more protocol element combinations to be selected. When the adjustment direction prioritizes computational accuracy, the membership threshold is increased to filter higher-quality protocol element combinations.

5. The high-precision reasoning and retrieval method based on multi-objective collaboration according to claim 4, characterized in that, The preliminary evidence chain is optimized using a multi-objective collaborative optimization based on the adjusted fuzzy matte model, and a correction parameter library containing multiple subspace correction parameters is constructed, including: Based on the adjusted fuzzy matte model, the fuzzy independent set family is redefined, and the coordinates of each evidence element in the initial evidence chain in the precision-speed-dependency multidimensional space are dynamically weighted and adjusted according to the redefined fuzzy independent set family to obtain a dynamic reference point set. The distribution of reference points in the dynamic reference point set is clustered to obtain multiple clusters, each cluster corresponding to a subspace; Analyze the data distribution characteristics in each cluster, determine the boundary requirements for computational accuracy, response speed and data dependency in the corresponding subspace, and construct constraints. Based on the aforementioned constraints, the optimal solution set for the corresponding subspace is obtained through multi-objective optimization. Inflection points in different directions are selected from the optimal solution set, and correction parameters for the corresponding subspaces are identified. The correction parameters include the membership threshold configuration and target weight configuration for the subspace. The correction parameters are correlated to construct the correction parameter library.

6. The high-precision reasoning and retrieval method based on multi-objective collaboration according to claim 5, characterized in that, The step of matching the target correction parameter group from the correction parameter library based on the real-time features of the retrieval request includes: Based on the real-time characteristics of the retrieval request, a retrieval feature vector is constructed; wherein, the real-time characteristics include an instant attribute vector of the current retrieval request's data size, modality type, response time limit, and system resource occupancy status; The correction parameter information corresponding to different subspaces is analyzed in the correction parameter library to generate corresponding subspace fingerprint vectors. Calculate the similarity between the retrieved feature vector and the subspace fingerprint vector, and filter out correction parameters with similarity greater than a preset similarity threshold as a set of candidate parameters; Based on the real-time strategy mode of the retrieval request, the utility value of different parameter combinations in the candidate parameter set is calculated, and the parameter combination with the highest utility value is selected as the target correction parameter group.

7. The high-precision reasoning retrieval method based on multi-objective collaboration according to claim 6, characterized in that, The step of real-time correction of the initial evidence chain to obtain an optimized evidence chain includes: Based on the target correction parameter set, extract the corresponding accuracy-speed-dependency target weight, membership threshold, and evidence element screening threshold; Based on the target weight, the coordinates of each evidence element in the initial evidence chain are re-weighted and mapped in the precision-speed-dependency multidimensional space to obtain the re-mapped set of evidence coordinates. Based on the membership threshold, the fuzzy independence verification is performed again on each combination of evidence elements in the remapped evidence coordinate set, and evidence elements with a membership degree lower than the membership threshold or that cause logical redundancy are removed to obtain a filtered evidence element set. Based on the evidence element screening threshold, the inference coefficient of each evidence element in the filtered evidence element set is recalculated, the core evidence elements with inference coefficients greater than the screening threshold are retained, and the comprehensive inference weight of each evidence element is configured accordingly. The core evidence elements are sorted and reorganized according to the comprehensive reasoning weights to obtain the optimized evidence chain.

8. The high-precision reasoning retrieval method based on multi-objective collaboration according to claim 7, characterized in that, The step of obtaining target knowledge data based on the optimized evidence chain, configuring inference weights, and generating inference response information includes: Based on the target knowledge node corresponding to the optimized evidence chain, target knowledge data matching the target knowledge node is obtained from the knowledge base; Based on the inference coefficients corresponding to each evidence element in the optimized evidence chain, and combined with the weight correction coefficients in the target correction parameter group, the comprehensive inference weight of each target knowledge data is calculated by weighted fusion. According to the preset response generation rules, the target knowledge data with completed reasoning weight configuration is organized into structured text containing conclusions, supporting evidence, and logical deduction processes to generate the reasoning response information.

9. A high-precision reasoning and retrieval system based on multi-objective collaboration, characterized in that, include: The acquisition unit is used to acquire a retrieval request, retrieve retrieval information from the knowledge graph based on the retrieval request, define each evidence unit in the retrieval information as a protocol element, and construct a protocol element set. Construction unit, used to construct a fuzzy matroid model based on the set of protocol elements; The initial selection unit is used to select, in the worst-case scenario, a combination of protocol elements that satisfies the preset security threshold and the association threshold and has the lowest total cost, based on the fuzzy matte model and the robust minimum weight basis solution algorithm, to obtain the robust minimum weight basis and form the initial selection evidence chain; The correction parameter library establishment unit is used to analyze the research stage characteristics of the retrieval request to generate dynamic adjustment factors, adjust the membership degree threshold of the fuzzy independence constraint in the fuzzy matte model based on the dynamic adjustment factors, redetermine the fuzzy independent set family, and perform multi-objective collaborative optimization on the initial evidence chain based on the adjusted fuzzy matte model to construct a correction parameter library containing multiple subspace correction parameters. The correction unit is used to match the target correction parameter group from the correction parameter library according to the real-time characteristics of the retrieval request, and to correct the initial evidence chain in real time to obtain the optimized evidence chain. The generation unit is used to obtain target knowledge data based on the optimized evidence chain, configure reasoning weights, and generate reasoning response information.

Images