Multi-module collaborative intelligent agent, device and medium for enhancing credibility of intelligence analysis

Abstract

The application discloses a multi-module cooperative intelligent agent for enhancing the credibility of intelligence analysis, equipment and medium, and relates to the technical field of artificial intelligence, comprising: a weighted retrieval mechanism based on a multi-dimensional quality evaluation index is constructed to score and sort the literature candidate set in response to the target task instruction to obtain an external knowledge source; the target task instruction is matched with a reasoning database to obtain a target reasoning paradigm and a target multi-stage prompt structure, and a large language model is driven to conduct structured reasoning on the external knowledge source according to the target reasoning paradigm and the target multi-stage prompt structure to obtain a reasoning conclusion; a thinking chain prompt engineering technology is used to conduct consistency verification on the reasoning conclusion and the external knowledge source, and when the consistency verification fails, a dynamic error correction mechanism is triggered to form a feedback closed loop. The application solves the technical problems of uncontrollable reasoning process and large noise of knowledge source in the prior art, which leads to low conclusion credibility, and achieves the technical effect of improving the credibility of intelligence analysis.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to multi-module collaborative intelligent agents, devices, and media for enhancing the credibility of intelligence analysis. Background Technology

[0002] In intelligence analysis systems built on large language models, agents typically rely on open networks or general databases to acquire background knowledge and automatically generate inference results based on input content. However, when performing complex intelligence reasoning tasks, these systems suffer from opaque reasoning paths and uncontrollable execution processes, resulting in a lack of structured logical support in the generation process. Furthermore, because literature retrieval mechanisms are mostly based on semantic similarity and do not screen and evaluate indicators such as the authority of knowledge sources, citation volume, and publication time, the search results are prone to containing low-quality, highly subjective, or factually unfounded information. These problems collectively lead to weak supporting evidence, broken logical chains, and poor factual consistency in the agent's reasoning conclusions, making it difficult to meet the demand for highly credible conclusions. Summary of the Invention

[0003] This application provides a multi-module collaborative intelligent agent, device, and medium to enhance the credibility of intelligence analysis, addressing the technical problems of uncontrollable reasoning processes and high noise in knowledge sources leading to low credibility of conclusions in existing technologies.

[0004] In view of the above problems, this application provides a multi-module collaborative intelligent agent, device and medium to enhance the credibility of intelligence analysis.

[0005] A first aspect of this application provides a multi-module collaborative intelligent agent to enhance the credibility of intelligence analysis, the agent comprising:

[0006] The external tool support module is used to score and rank the candidate literature set in response to the target task instructions based on a weighted retrieval mechanism constructed based on multi-dimensional quality evaluation indicators, thereby obtaining external knowledge sources. The reasoning module is used to match the target task instructions with the reasoning database to obtain the target reasoning paradigm and the target multi-stage prompting structure, and drive the large language model to perform structured reasoning on the external knowledge sources based on the target reasoning paradigm and the target multi-stage prompting structure to obtain reasoning conclusions. The reasoning database is used to store the task type set and reasoning paradigm set obtained after reasoning paradigm extraction based on the scientific thinking chain construction mechanism, as well as the multi-stage prompting structure set obtained after prompting structure mapping and transformation of the reasoning paradigm set through a chain-style reasoning scheduling mechanism. The credibility supervision module is used to perform consistency verification between the reasoning conclusions and external knowledge sources using thinking chain prompting engineering technology. When the consistency verification fails, a dynamic error correction mechanism is triggered to form a feedback loop.

[0007] A second aspect of this application provides an electronic device, including: a memory for storing executable instructions; and a processor for executing the executable instructions stored in the memory to implement the multi-module collaborative intelligent agent for enhancing the credibility of intelligence analysis provided in this application.

[0008] A third aspect of the embodiments of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the multi-module collaborative intelligent agent provided in this application to enhance the credibility of intelligence analysis.

[0009] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0010] This application employs a weighted retrieval mechanism based on multidimensional quality evaluation indicators to score and rank a candidate set of documents in response to target task instructions, thereby obtaining external knowledge sources. It then matches the target task instructions with a reasoning database to obtain target reasoning paradigms and target multi-stage prompting structures. This drives a large language model to perform structured reasoning on the external knowledge sources based on the target reasoning paradigms and target multi-stage prompting structures, resulting in a reasoning conclusion. The reasoning database stores a set of task types and a set of reasoning paradigms obtained after extracting reasoning paradigms using a scientific thinking chain construction mechanism, as well as a set of multi-stage prompting structures obtained after mapping and transforming the reasoning paradigm set using a chain-based reasoning scheduling mechanism. Finally, it uses thinking chain prompting engineering technology to verify the consistency between the reasoning conclusions and external knowledge sources. If the consistency verification fails, a dynamic error correction mechanism is triggered, forming a feedback loop. This invention addresses the technical problems of uncontrollable reasoning processes and high noise levels in knowledge sources leading to low reliability of conclusions in existing technologies. By constructing a multi-module collaborative intelligent agent architecture that includes an external tool support module, a reasoning module, and a reliability supervision module, a closed-loop reasoning control process of retrieval-reasoning-supervision is formed. This achieves the technical effect of improving the structuring of the reasoning process, enhancing the authority of knowledge sources and the verification capability of conclusions, thereby improving the reliability of intelligence analysis. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 A schematic diagram of a multi-module collaborative intelligent agent structure for enhancing the credibility of intelligence analysis provided in this application embodiment;

[0013] Figure 2This is a schematic diagram of the structure of an exemplary electronic device of this application.

[0014] Explanation of reference numerals in the attached drawings: Bus 300, Receiver 301, Processor 302, Transmitter 303, Memory 304, Bus Interface 305, External Tool Support Module 11, Inference Module 12, Credibility Supervision Module 13. Detailed Implementation

[0015] This application addresses the technical problems of low reliability of conclusions due to uncontrollable reasoning processes and high noise in knowledge sources in existing technologies by providing a multi-module collaborative intelligent agent, device, and medium to enhance the credibility of intelligence analysis. It constructs a multi-module collaborative intelligent agent architecture that includes an external tool support module, a reasoning module, and a credibility supervision module, forming a closed-loop reasoning control process of retrieval-reasoning-supervision. This improves the structuring of the reasoning process, enhances the authority of knowledge sources and the verification capability of conclusions, thereby improving the technical effect of enhancing the credibility of intelligence analysis.

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0017] It should be noted that any variation of the terms "comprising" and "having" is intended to cover non-exclusive inclusion, for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to such processes, methods, products, or devices.

[0018] Example 1, as Figure 1 As shown, this application provides a multi-module collaborative intelligent agent to enhance the credibility of intelligence analysis, the intelligent agent comprising:

[0019] External tool support module 11 is used to score and rank the candidate literature set in response to the target task instructions by a weighted retrieval mechanism based on multidimensional quality evaluation indicators, thereby obtaining external knowledge sources.

[0020] In this embodiment of the application, the external tool support module 11 first receives the target task instruction during execution. The instruction contains basic information about the task, such as task type, analysis target, and semantic keywords, which is used to clarify the subject scope and semantic orientation of the literature retrieval.

[0021] Next, the external tool support module 11 retrieves semantically relevant preliminary data from the connected literature database to form a candidate set of literature. This candidate set of literature refers to a collection of literature that is semantically related to the content of the task instructions, and is initially screened out through semantic matching.

[0022] Next, the external tool support module 11 initiates a weighted retrieval mechanism based on multidimensional quality evaluation indicators. This mechanism filters high-quality literature by comprehensively scoring and ranking each document in the candidate set. The multidimensional quality evaluation indicators include four dimensions: topic matching, citation count, source authority, and time novelty. These measures the semantic fit between the document and the task objective, academic influence, the authority level of the publication channel, and the coverage of the latest viewpoints, respectively. The weighted retrieval mechanism internally includes a multidimensional document quality scoring function. This function takes the normalized scores of the four dimensions as input and performs a weighted calculation based on preset weights to obtain a comprehensive score for each candidate document. After scoring, the external tool support module 11 sorts the candidate document set from highest to lowest score and identifies the top K documents as external knowledge sources.

[0023] Furthermore, in the intelligent agent provided in the application embodiment, the external tool support module 11 further includes:

[0024] The function extraction unit is used to extract the multidimensional document quality scoring function from the weighted retrieval mechanism constructed based on multidimensional quality evaluation indicators, wherein the multidimensional document quality scoring function is:

[0025] S(d)=α1×R match (d)+α2×R cite (d)+α3×R source (d)+α4×R time (d);

[0026] Where S(d) represents the overall score of candidate document d, and R match (d), R cite (d), R source (d), R time (d) represents the normalized scores of candidate document d in four dimensions: topic matching degree, citation volume, source authority, and time novelty. α1, α2, α3, and α4 are the weights of topic matching degree, citation volume, source authority, and time novelty when performing quality scoring. The scoring unit is used to score the candidate document set based on the multidimensional document quality scoring function, and sort them in descending order of score. The top K documents are added to the external knowledge source, where K is a positive integer.

[0027] In this embodiment, the function extraction unit extracts a multidimensional document quality scoring function for evaluating the quality of candidate documents. In this multidimensional document quality scoring function, S(d) represents the comprehensive score of candidate document d, and R... match (d), R cite (d), R source (d), R time (d) represents the normalized scores of candidate document d on four dimensions: topic matching degree, number of citations, source authority, and time novelty. α1, α2, α3, and α4 are the weights of topic matching degree, number of citations, source authority, and time novelty when scoring quality, respectively, and are set by default to α1 = 0.3, α2 = 0.3, α3 = 0.2, and α4 = 0.2.

[0028] Among them, in calculating the topic matching degree R match In step (d), the scoring unit matches the semantic keywords (e.g., image recognition, medical applications, etc.) in the task instructions with the title, abstract, and keyword fields of the candidate documents. The text is represented as a vector using a semantic embedding model (such as BERT), and the cosine similarity between the task keyword vector and the document vector is calculated, with the result normalized to the 0-1 range. For example, if the task objective is agricultural image recognition, and a document's title is "A Method for Crop Disease Image Analysis Based on Neural Networks," and the semantic matching yields a cosine similarity of 0.86, then the document's R... match (d) = 0.86.

[0029] In calculating the number of citations R cite (d) In this case, the number of citations of the candidate document is used as the scoring criterion for this dimension. The citation frequency of the document is retrieved from databases such as CNKI and normalized to the highest citation value in the current candidate document set. For example, if a document is cited 220 times and the highest citation value in the current candidate set is 275 times, then its citation score is R. cite (d) = 220 / 275 = 0.8.

[0030] Regarding the authority of the source R source (d) The scoring unit assigns a score to the professionalism of the literature source by referring to a pre-defined authoritative information table of publication sources. This information table is usually constructed based on dimensions such as whether the journal is indexed by SCI and the journal's impact factor. For example, if a document is published in a core journal with wide influence in the fields of image recognition and artificial intelligence, such as the *Journal of Image and Graphics* or the *Chinese Journal of Computers*, its source authority score can be set to above 0.90; while if the source of the document is a regional general journal or an unindexed journal, its score will be below 0.3. Finally, all scores are normalized to obtain the corresponding source authority of the document.

[0031] Temporal novelty Rtime (d) Based on the publication year of the literature, linear normalization is performed according to a preset time window. Assuming the window is set from 2010 to 2025, the score of literature published in 2023 is (2023-2010) / (2025-2010)≈0.867. Newer literature scores higher, reflecting its ability to cover current research trends.

[0032] The scoring unit multiplies the four normalized sub-scores mentioned above with their corresponding weight coefficients one by one and sums them up to obtain the comprehensive score S(d) for each candidate document. After scoring, the scoring unit sorts the candidate document set from high to low scores and selects the top K documents to form external knowledge sources, where K is a positive integer, such as 10 or 20.

[0033] The reasoning module 12 is used to match the target task instructions with the reasoning database to obtain the target reasoning paradigm and the target multi-stage prompting structure, and drive the large language model to perform structured reasoning on the external knowledge source according to the target reasoning paradigm and the target multi-stage prompting structure to obtain a reasoning conclusion. The reasoning database is used to store the task type set and reasoning paradigm set obtained after the reasoning paradigm is extracted according to the scientific thinking chain construction mechanism, and the multi-stage prompting structure set obtained after the prompting structure is mapped and transformed by the chain reasoning scheduling mechanism.

[0034] In this embodiment, after receiving the target task instruction, the inference module 12 first converts the instruction into a vector representation using an embedded semantic encoding method, extracts its semantic features using semantic embedding models such as BERT, and matches it one-to-one with the task type set stored in the inference database. Cosine similarity calculation is used to identify the task type with the highest similarity as the matching result. The task type set refers to a list of standard task types constructed through expert annotation and intelligence task classification, such as hotspot identification, development trend identification, technical bottleneck identification, and potential main line identification. Each task type corresponds to a specific inference logic requirement. For example, if the task instruction involves the current development trend of artificial intelligence image recognition, its type is determined to be development trend identification.

[0035] For example, the task type can be hotspot identification, which uses a constructed multi-module collaborative agent to identify hotspots on a reference standard dataset as shown in Table 1, which has been pre-constructed by those skilled in the art.

[0036] Table 1 Reference Standard Dataset

[0037]

[0038] After task type matching is completed, reasoning module 12 retrieves the target reasoning paradigm corresponding to the task type from the reasoning paradigm set in the reasoning database. The reasoning paradigm set refers to a set of structured reasoning templates extracted from a large number of typical intelligence task materials based on the scientific thinking chain construction mechanism. The scientific thinking chain construction mechanism is a structural modeling method used to simulate the cognitive process followed by intelligence experts in actual tasks. Its core lies in uniformly modeling typical tasks into a structural framework composed of four elements: "reasoning objective - variable structure - causal path - judgment output." Specifically, the reasoning objective defines the core problem to be solved by the task; the variable structure clarifies the core variables and attribute elements to be analyzed in the task; the causal path describes the logical dependencies and deductive chains between variables; and the judgment output specifies the final form of the reasoning conclusion and the verification conditions. Under the guidance of this construction mechanism, each type of task forms an operable reasoning paradigm, which is stored in the reasoning database in a standardized manner.

[0039] Taking trend identification tasks as an example, the corresponding goal reasoning paradigm is a structured reasoning process extracted from the aforementioned four-element modeling framework, specifically including four reasoning nodes: background identification, hotspot extraction, technological bottleneck identification, and future trend prediction. Each reasoning node, as an independent cognitive operation unit, logically progresses in the order of "reasoning goal - variable structure - causal path - judgment output." For example, the reasoning goal of the technological bottleneck identification node is to identify the limitations of the method; the variable structure includes the number of model parameters, data dependencies, and computing resources; the causal path is large number of parameters → deployment difficulties; and the judgment output is that the current model is difficult to adapt to edge environments.

[0040] After the target reasoning paradigm is determined, the reasoning module 12 calls the target multi-stage hint structure corresponding to the target reasoning paradigm from the reasoning database. This hint structure has been identified and converted through the chain reasoning scheduling mechanism and stored in the reasoning database in advance.

[0041] Subsequently, the reasoning module 12 assembles the aforementioned target reasoning paradigm, target multi-stage prompting structure, and external knowledge sources provided by the external tool support module 11 into a standard input format that can be processed by the large language model. Each set of input content consists of "stage prompting structure + knowledge material fragment + output control format," where the knowledge material is selected from the title, abstract, method description, and research conclusion paragraphs of external literature, and the output control format is used to limit the structure, number, and items of the results generated by the language model. For example, in the technical bottleneck identification stage, the model input content is "Please identify the limitations of current mainstream image recognition methods," accompanied by a literature abstract describing the difficulties in deploying convolutional neural networks, and instructs the output of 3 questions and their supporting paragraphs.

[0042] The reasoning module 12 calls the large language model to perform structured reasoning in stages according to the order of the reasoning nodes. In each stage, the large language model completes variable extraction, logical judgment, and intermediate conclusion generation based on the prompt structure, and the results are stored in the cache for use in the next stage. For example, in the background recognition stage, the model generates content such as image recognition, which has recently focused on edge deployment and scene generalization. This information will be used to support subsequent causal analysis in bottleneck recognition.

[0043] After all stages are completed, the reasoning module 12 calls the integration logic to organize the intermediate results in a structured manner, generate the reasoning conclusion in the order of "reasoning goal - variable structure - causal path - judgment output" in the original reasoning paradigm, and bind the conclusive judgment to the supporting fragments in the knowledge material one by one.

[0044] Furthermore, in the intelligent agent provided in the application embodiment, the inference module 12 further includes:

[0045] The system comprises the following components: a priori classification unit for classifying typical intelligence task sets to obtain M task types and M task material sets, where M is a positive integer; a reasoning paradigm extraction unit for extracting reasoning paradigms from the M task material sets based on a scientific thinking chain construction mechanism to obtain M initial reasoning paradigm sets; an iterative memory extraction unit for iteratively extracting memories from the M initial reasoning paradigm sets to determine M reasoning paradigms; and a summarization unit for summarizing the M task types and M reasoning paradigms to obtain the task type set and reasoning paradigm set, where each task type corresponds to one reasoning paradigm.

[0046] In this embodiment, the prior classification unit first receives a set of typical intelligence tasks, which consists of intelligence question texts either manually annotated or automatically captured. For prior classification, the TF-IDF vectorization method is used to extract keywords and calculate the inverse document frequency (IVF) of each intelligence task text in the set, converting the task text into a vector representation. Subsequently, the K-means clustering algorithm is applied to automatically divide the tasks into M clusters based on the distribution of these vectors in the high-dimensional semantic space. Each cluster represents a task type, and the corresponding output consists of M task types and the task text materials contained in each type, forming M task material sets. Taking typical tasks such as identifying technical bottlenecks, summarizing hot topics, or analyzing development trends as examples, different texts are automatically grouped into the same task category due to similar semantic distributions.

[0047] Subsequently, the reasoning paradigm extraction unit processes M task material sets, extracting the corresponding initial reasoning paradigm from each set based on the scientific thinking chain construction mechanism. The reasoning target extraction stage uses keyword matching and syntactic subject-verb-object structure recognition methods to identify the sentence structures in the task description that indicate the analytical intent, such as identifying the future development direction of image recognition algorithms. During variable structure extraction, non-entity recognition (NER) technology is used to extract recurring technical elements, research objects, or performance indicators in the material, such as model size, data quality, and computing speed, forming a list of variables that need to be analyzed in the task. Causal path extraction is based on causal relationship recognition methods, utilizing a pre-defined causal connector matching strategy (such as because, leading to, further, result, cause, influence, prompt, due to, thus, etc.), combined with dependency parsing, to extract logical dependencies between variables in the material, such as logical chains like "insufficient data → decreased accuracy → user churn." The judgment output extraction stage uses logical reasoning sentence structure recognition methods to locate the final judgmental language from the paragraph conclusions, such as "existing models cannot meet edge deployment requirements," clarifying the expression and verification direction of the conclusions formed by the task. The above four structural elements together form a standardized reasoning paradigm. Each task material set outputs a batch of well-structured reasoning templates through the above process, ultimately resulting in M ​​initial reasoning paradigm sets.

[0048] The iterative memory retrieval unit then performs iterative memory retrieval on the M initial reasoning paradigm sets. In this process, firstly, an initial reasoning paradigm is randomly selected without replacement from each initial reasoning paradigm set, and keywords are extracted based on a preset set of reasoning node keywords to identify the corresponding node information, constructing the first round of memory representation. Subsequently, in each iteration, new initial reasoning paradigms are randomly selected from each set, the keyword extraction process is repeated, and the representation is compared and merged with the previous round of memory representation, gradually updating the keyword set content and frequency structure. After all initial reasoning paradigms have been processed, a filtering operation is performed on the final memory structure to determine the M most representative reasoning paradigms, each corresponding to one of the M task types.

[0049] Finally, the aggregation unit matches and integrates the above M task types and their corresponding M reasoning paradigms one by one, forming a task type set and a reasoning paradigm set respectively, and establishes a structured mapping for each task type corresponding to a reasoning paradigm.

[0050] Furthermore, in the intelligent agent provided in the application embodiments, the iterative memory retrieval unit further includes:

[0051] A random extraction subunit is used to randomly extract M first initial reasoning paradigms without replacement from the M initial reasoning paradigm sets respectively; a keyword extraction subunit is used to extract keywords from the M first initial reasoning paradigms according to a preset reasoning node keyword set, determine M first reasoning node keyword sets, and add the preset reasoning node keyword set and the M first reasoning node keyword sets to M first iterative memory units, wherein each first reasoning node keyword in the first reasoning node keyword set has a node identifier; an update subunit is used to update based on the M first iterative memory units. Keyword extraction is performed on M second initial reasoning paradigms randomly extracted without replacement from the set of M initial reasoning paradigms, and the M first iterative memory units are updated based on the extraction results to obtain M second iterative memory units; an iterative subunit is used to sequentially extract keywords and update iterative memory units on the remaining initial reasoning paradigms in the set of M initial reasoning paradigms based on the M second iterative memory units to obtain M target iterative memory units; a filtering subunit is used to filter the M target iterative memory units to determine the M reasoning paradigms.

[0052] In this embodiment of the application, firstly, a random sub-unit is randomly extracted to perform random abstraction without replacement on the M initial inference paradigm sets, and M first initial inference paradigms are extracted respectively.

[0053] Subsequently, the keyword extraction subunit loads a preset set of inference node keywords, which includes keywords representing causal logical relationships such as "because," "leading to," "causing," and "therefore." The keyword extraction subunit sequentially performs keyword extraction operations on the aforementioned M initial inference paradigms, using pattern matching and dependency parsing methods to identify linguistic expressions with causal attributes in each inference paradigm, and constructs M sets of first inference node keywords accordingly. During this process, to ensure the traceability of inference nodes in structured processing, each extracted inference node keyword is assigned a unique node identifier, including its position in the original inference text, its task category, and its node hierarchy. After completing keyword extraction, the keyword extraction subunit writes the preset set of inference node keywords and the corresponding M sets of first inference node keywords into M first iterative memory units, thereby establishing the first round of iterative memory structure.

[0054] Next, the updating sub-unit continues to perform a random extraction operation without replacement in each initial inference paradigm set, obtaining M second initial inference paradigms. Based on this, the updating sub-unit performs keyword extraction operations on the extracted M second initial inference paradigms, using the node identifiers, keyword distributions, and structural patterns from the aforementioned M first iterative memory units. This ensures that the extraction process completes, aligns, and enhances the existing semantic framework. The extracted keyword information is used to update the corresponding first iterative memory units, thus forming M second iterative memory units. Each unit integrates keyword node information from two different inference paradigm samples, resulting in a more complete structure.

[0055] The iterative subunit then continues to extract the remaining unprocessed initial reasoning paradigms from the M initial reasoning paradigm sets sequentially. Based on the second iterative memory unit obtained in the current iteration round, it performs keyword extraction and memory update operations. After each initial reasoning paradigm is processed, the corresponding iterative memory unit is updated in real time until all reasoning paradigms in each set have been traversed, ultimately outputting M target iterative memory units.

[0056] Finally, the selection sub-unit filters the M target iterative memory units. Specifically, it clusters the inference node keywords based on node identifiers to obtain M inference node keyword clusters. Then, it extracts the most frequent inference node keywords from each keyword cluster and concatenates them according to the order of their corresponding node identifiers to finally output M inference paradigms.

[0057] Furthermore, in the intelligent agent provided in the application embodiments, the filtering subunit further includes:

[0058] Clustering micro-units are used to cluster the inference node keywords in the M target iterative memory units based on node identifiers to obtain M inference node keyword clusters; concatenating micro-units are used to extract the inference node keyword with the highest frequency in the inference node keyword set corresponding to each node identifier in the M inference node keyword clusters, and concatenate them according to the order of the node identifiers to obtain the M inference paradigms.

[0059] In this embodiment, the clustering micro-unit first performs clustering processing on the inference node keywords contained in the M target iterative memory units based on node identifiers. Specifically, each target iterative memory unit contains multiple inference node keywords, and each inference node keyword is bound to a unique node identifier, such as node 1, node 2, etc. The clustering micro-unit adopts a clustering method based on identifier hash mapping to construct a clustering dictionary with node identifiers as keys and the set of inference node keywords as values, and divides all inference node keywords into corresponding inference node keyword clusters according to their node identifiers, ultimately obtaining M inference node keyword clusters.

[0060] After clustering, the concatenated micro-units perform keyword filtering and sequence construction operations on the M inference node keyword clusters. During this process, the concatenated micro-units use word frequency statistics to calculate the frequency of each keyword under its corresponding node identifier within each keyword cluster. Then, the keyword with the highest frequency in each cluster is selected as the representative keyword corresponding to that node identifier. To ensure the sequential nature of logical reasoning, the concatenated micro-units sequentially concatenate these representative keywords according to the node identifiers (e.g., node 1 to node n) to form a complete reasoning path structure, ultimately resulting in M ​​inference paradigms.

[0061] Furthermore, in the intelligent agent provided in the application embodiment, the inference module 12 further includes:

[0062] The identifier extraction unit is used to extract the cue structure identifier in the chain-like inference scheduling mechanism; the cue structure identification unit is used to use the cue structure identifier to identify the cue structure of each inference node of each inference paradigm in the inference paradigm set, and obtain the multi-stage cue structure set, wherein each stage cue structure in the multi-stage cue structure set is a sub-target description-analysis element-expected output format.

[0063] In this embodiment, the recognizer extraction unit first extracts the cue structure recognizer from the chain-based inference scheduling mechanism. The cue structure recognizer is a pre-built neural network model with a multi-layer Transformer architecture. During the training phase, the cue structure recognizer is trained on data including intelligence analysis documents, scientific paper paragraphs, and inference sample sets with manual annotations. Each training sample contains corresponding cue structure annotation information for supervised learning. The cue structure annotation uses a three-segment structure label, including sub-target description, analysis elements, and expected output format. Parameter convergence is achieved by minimizing the structure analysis loss function, ultimately obtaining a cue structure recognizer model with structure recognition capabilities.

[0064] Subsequently, the prompt structure recognition unit utilizes the prompt structure recognizer to perform structural parsing on each inference node of each inference paradigm in the inference paradigm set. During this process, the prompt structure recognition unit inputs the inference node text into the prompt structure recognizer one by one, and the recognizer outputs structured prompts. This prompt structure follows a standard triplet format, meaning that each stage's prompt structure consists of three parts: a sub-goal description, analytical elements, and an expected output format. The sub-goal description describes the intent and target of the inference node's current task; the analytical elements clarify the semantic dimensions that the task needs to focus on, such as event elements, evidence types, or temporal / spatial clues; and the expected output format specifies the content format that the task is expected to return, such as a natural language answer, label identification, or nested structural templates.

[0065] Finally, the structure recognition unit performs the above structure extraction process on all inference nodes and outputs a multi-stage prompt structure set, where each stage prompt structure is a combination structure of sub-target description, analysis elements, and expected output format.

[0066] The credibility supervision module 13 is used to perform consistency verification between reasoning conclusions and external knowledge sources using the thinking chain prompting engineering technology. When the consistency verification fails, a dynamic error correction mechanism is triggered to form a feedback loop.

[0067] In this embodiment, when the credibility supervision module 13 uses the mind chain prompting engineering technology to verify the consistency between the reasoning conclusion and the external knowledge source, it inputs the reasoning conclusion, supporting evidence from the knowledge source, and judgment instructions into the consistency verification model to generate a consistency verification result. When the judgment result is unsuccessful, a dynamic error correction mechanism is activated, a two-stage optimization process is executed, and the reasoning conclusion is regenerated and updated, thereby realizing a feedback loop for the reasoning output.

[0068] Furthermore, in the intelligent agent provided in the application embodiment, the trustworthiness supervision module 13 further includes:

[0069] The consistency verification unit uses the thought chain prompting engineering technique to input the reasoning conclusion, supporting evidence from external knowledge sources, and judgment instructions into the consistency verification model and output the consistency verification result. The optimization unit is used to trigger the two-stage optimization process in the dynamic error correction mechanism when the consistency verification result is that the consistency verification fails, and obtain an updated reasoning conclusion.

[0070] Furthermore, the intelligent agent provided in the application embodiments also includes:

[0071] The two-stage optimization process includes expanding supporting evidence from external knowledge sources using a weighted retrieval mechanism based on multidimensional quality evaluation indicators to obtain new evidence, and re-reasoning based on the new evidence and external knowledge sources to update the reasoning conclusion and obtain an updated reasoning conclusion.

[0072] In this embodiment, the consistency verification unit employs a thought chain prompting engineering technique, inputting the reasoning conclusion, supporting evidence from external knowledge sources, and judgment instructions into the consistency verification model. Supporting evidence refers to cited external knowledge paragraphs, such as sentences from papers, reports, or databases, whose content must be factually relevant to the reasoning conclusion. Judgment instructions guide the model to focus on the supporting relationships among the three, such as "Judging whether the above evidence supports the reasoning conclusion." Before input, the three pieces of information are organized into structured prompts. These structured prompts are then input into the consistency verification model, a trained scoring model that outputs a consistency score based on the semantic relationships of the input triples. This score is a continuous value between 0 and 1, used to quantify whether the current reasoning conclusion is consistent with the supporting evidence. This consistency score is compared with a preset consistency threshold; if the score is less than the preset threshold, the consistency verification fails. This consistency verification model is constructed using a BERT-based three-stage attention mechanism, and during training, a large number of labeled samples with consistency score labels are used for supervised learning of the reasoning-evidence-instruction triples. The consistency score label is a real value between 0 and 1, representing the strength of support for the inference conclusion, where 1 represents complete support and 0 represents complete lack of support. The labels are obtained through expert annotation. The model training objective is to minimize the mean squared error (MSE) loss between the predicted score and the true label score, in order to optimize the model parameters and improve the accuracy of the consistency assessment.

[0073] When the consistency score output by the consistency verification model is lower than a preset threshold (e.g., 0.8), the consistency verification is deemed to have failed. This triggers a two-stage optimization process in the dynamic error correction mechanism to revise and update the current inference conclusion. In the first stage of the revision and update, a weighted retrieval mechanism based on multi-dimensional quality evaluation indicators is used to select text paragraphs relevant to the current problem from the candidate literature set (excluding original supporting evidence) and calculate scores based on the constructed literature quality scoring function. Paragraphs with scores higher than the pre-set scoring threshold by technical experts are selected as new evidence. Subsequently, a new knowledge context is constructed based on the new evidence and the original supporting evidence, and the thought chain prompt structure is reconstructed and input into the inference module 12 to generate an updated inference conclusion. For example, if the original conclusion "This system improves the credibility of inference" is deemed inconsistent due to insufficient evidence, a description from a peer-reviewed paper is added as new evidence. This new evidence is then recombined and input, and the large language model is re-called using a unified prompt template for chain-like inference to generate a new inference conclusion with stronger logical support.

[0074] In summary, the embodiments of this application have at least the following technical effects:

[0075] This application employs a weighted retrieval mechanism based on multidimensional quality evaluation indicators to score and rank a candidate set of documents in response to target task instructions, thereby obtaining external knowledge sources. It then matches the target task instructions with a reasoning database to obtain target reasoning paradigms and target multi-stage prompting structures. This drives a large language model to perform structured reasoning on the external knowledge sources based on the target reasoning paradigms and target multi-stage prompting structures, resulting in a reasoning conclusion. The reasoning database stores a set of task types and a set of reasoning paradigms obtained after extracting reasoning paradigms using a scientific thinking chain construction mechanism, as well as a set of multi-stage prompting structures obtained after mapping and transforming the reasoning paradigm set using a chain-based reasoning scheduling mechanism. Finally, it uses thinking chain prompting engineering technology to verify the consistency between the reasoning conclusions and external knowledge sources. If the consistency verification fails, a dynamic error correction mechanism is triggered, forming a feedback loop. This invention addresses the technical problems of uncontrollable reasoning processes and high noise levels in knowledge sources leading to low reliability of conclusions in existing technologies. By constructing a multi-module collaborative intelligent agent architecture that includes an external tool support module, a reasoning module, and a reliability supervision module, a closed-loop reasoning control process of retrieval-reasoning-supervision is formed. This achieves the technical effect of improving the structuring of the reasoning process, enhancing the authority of knowledge sources and the verification capability of conclusions, thereby improving the reliability of intelligence analysis.

[0076] Example 2: Based on the inventive concept of a multi-module collaborative intelligent agent that enhances the credibility of intelligence analysis in the foregoing embodiments, this application also provides an electronic device, including: at least one processor; a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the intelligent agent described in any of the above Examples 1.

[0077] Figure 2 This is a schematic diagram of the structure of an exemplary electronic device of this application. Figure 2 In this document, the bus architecture is represented by bus 300. Bus 300 may include any number of interconnected buses and bridges, and bus 300 connects various circuits including one or more processors represented by processor 302 and memory represented by memory 304. Bus 300 may also connect various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. Bus interface 305 provides an interface between bus 300 and receiver 301 and transmitter 303. Receiver 301 and transmitter 303 may be the same element, i.e., a transceiver, providing a unit for communicating with various other devices over a transmission medium. Processor 302 is responsible for managing bus 300 and general processing, while memory 304 can be used to store data used by processor 302 during operation.

[0078] Example 3: Based on the multi-module collaborative intelligent agent that enhances the credibility of intelligence analysis in the foregoing embodiments, and with the same inventive concept, this application also provides a computer-readable storage medium storing a computer program, which, when executed, implements the steps of the intelligent agent described in any one of the above-described Examples 1.

[0079] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0080] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0081] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A multi-module collaborative intelligent agent for enhancing the credibility of intelligence analysis, characterized in that, include: The external tool support module is used to score and rank the candidate literature set in response to the target task instructions by a weighted retrieval mechanism based on multidimensional quality evaluation indicators, thereby obtaining external knowledge sources. The reasoning module is used to match the target task instructions with the reasoning database to obtain the target reasoning paradigm and the target multi-stage prompting structure, and drive the large language model to perform structured reasoning on the external knowledge source according to the target reasoning paradigm and the target multi-stage prompting structure to obtain a reasoning conclusion. The reasoning database is used to store the task type set and reasoning paradigm set obtained after the reasoning paradigm is extracted according to the scientific thinking chain construction mechanism, and the multi-stage prompting structure set obtained after the prompting structure is mapped and transformed by the chain reasoning scheduling mechanism. The credibility supervision module is used to verify the consistency between reasoning conclusions and external knowledge sources using the thinking chain prompting engineering technology. When the consistency verification fails, a dynamic error correction mechanism is triggered to form a feedback loop. The reasoning module includes: The prior classification unit is used to perform prior classification on a typical intelligence task set to obtain M task types and M task material sets, where M is a positive integer; The reasoning paradigm extraction unit is used to extract reasoning paradigms from the M task material sets based on the scientific thinking chain construction mechanism to obtain M initial reasoning paradigm sets. An iterative memory extraction unit is used to perform iterative memory extraction on the M initial inference paradigm sets to determine the M inference paradigms; The iterative memory extraction unit includes: Randomly extract sub-units to randomly and without replacement draw an initial inference paradigm from each initial inference paradigm set, resulting in M ​​first initial inference paradigms; The keyword extraction subunit is used to extract keywords from the M first initial reasoning paradigms according to the preset reasoning node keyword set, determine the M first reasoning node keyword sets, and write the preset reasoning node keyword set and the M first reasoning node keyword sets into the M first iterative memory units, wherein each first reasoning node keyword in the first reasoning node keyword set has a node identifier; The update subunit is used to extract keywords from M second initial inference paradigms randomly extracted from the M initial inference paradigm sets without replacement, based on the M first iterative memory units, and update the M first iterative memory units according to the extraction results to obtain M second iterative memory units. An iterative subunit is used to extract keywords and update iterative memory units sequentially for the remaining initial reasoning paradigms in the set of M initial reasoning paradigms according to the M second iterative memory units, thereby obtaining M target iterative memory units. A filtering subunit is used to filter the M target iterative memory units and determine the M inference paradigms; The summarization unit is used to summarize the M task types and M reasoning paradigms respectively to obtain the task type set and reasoning paradigm set, wherein each task type corresponds to one reasoning paradigm.

2. The multi-module collaborative intelligent agent for enhancing intelligence analysis credibility as claimed in claim 1 wherein, The external tool support module includes: The function extraction unit is used to extract the multidimensional document quality scoring function from the weighted retrieval mechanism constructed based on multidimensional quality evaluation indicators, wherein the multidimensional document quality scoring function is: ； in, This represents the overall score of candidate document d. The scores are the normalized scores of candidate document d on four dimensions: topic relevance, citation count, source authority, and time novelty. , , The weights of topic relevance, citation count, source authority, and time novelty in the quality scoring are respectively: The scoring unit is used to score the candidate documents by traversing the document candidate set based on the multidimensional document quality scoring function, and sort them in descending order of score, adding the top K documents to the external knowledge source, where K is a positive integer.

3. The multi-module collaborative intelligent agent for enhancing intelligence analysis trustworthiness of claim 1, wherein, The filtering subunit includes: Clustering micro-units are used to cluster inference node keywords within the M target iterative memory units based on node identifiers to obtain M inference node keyword clusters; The serial micro-units are used to extract the most frequently occurring inference node keyword from the inference node keyword set corresponding to each node identifier in the M inference node keyword clusters, and to serialize them according to the order of the node identifiers to obtain the M inference paradigms.

4. The multi-module collaborative intelligent agent for enhancing intelligence analysis trustworthiness of claim 1, wherein, The reasoning module also includes: The recognizer extraction unit is used to extract the prompt structure recognizer in the chain reasoning scheduling mechanism; The prompt structure recognition unit is used to perform prompt structure recognition on each inference node of each inference paradigm in the inference paradigm set using the prompt structure recognizer, to obtain the multi-stage prompt structure set, wherein each stage prompt structure in the multi-stage prompt structure set is a sub-target description-analysis element-expected output format.

5. The multi-module collaborative intelligent agent for enhancing intelligence analysis trustworthiness of claim 1, wherein, The credibility supervision module includes: The consistency verification unit is used to input reasoning conclusions, supporting evidence from external knowledge sources, and judgment instructions into the consistency verification model using the mind chain prompting engineering technology, and output the consistency verification result. The optimization unit is used to trigger the two-stage optimization process in the dynamic error correction mechanism when the consistency check result is that the consistency check fails, so as to obtain the updated inference conclusion.

6. The multi-module collaborative intelligent agent for enhancing intelligence analysis trustworthiness of claim 5, wherein, The two-stage optimization process includes expanding supporting evidence from external knowledge sources using a weighted retrieval mechanism based on multidimensional quality evaluation indicators to obtain new evidence, and re-reasoning based on the new evidence and external knowledge sources to update the reasoning conclusion and obtain an updated reasoning conclusion.

7. An electronic device, comprising: The electronic device includes: Memory, used to store executable instructions; A processor, when executing executable instructions stored in the memory, implements a multi-module collaborative intelligent agent that enhances the credibility of intelligence analysis as described in any one of claims 1-6.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that, When executed by the processor, the program implements a multi-module collaborative agent that enhances the credibility of intelligence analysis as described in any one of claims 1-6.

Images