An autonomously controllable heterogeneous data-driven agent coordination optimization system and method

Abstract

The application discloses a kind of self-controllable heterogeneous data-driven agent collaborative optimization system and method, by deploying reasoning engine and functional component, construct the private domain knowledge database containing the knowledge information of field of interest, utilize agent development and management collaborative platform to develop multi-task agent based on functional component and reasoning engine cooperation, based on private domain knowledge database, the iteration question and answer of the secondary fine arrangement of the coarse screening result of the iterative question and answer of the intention recognition, query rewriting, dynamic programming, retrieval recall coarse screening of multi-task agent in turn is carried out, to realize multi-task agent based on self-controllable heterogeneous data-driven collaborative optimization.The application is built by private domain knowledge database Agent development and management collaborative platform, realize the full-link self-controllable of multi-task agent, introduce the problem retrieval of multiple rounds iteration of field knowledge and user feedback, collaborative optimization is carried out to retrieval result, significantly improve the efficiency and retrieval precision of problem retrieval.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a self-controllable heterogeneous data-driven agent collaborative optimization system and method. BACKGROUND

[0002] In recent years, the rise and widespread application of large models (LLMs) mark the paradigm shift of artificial intelligence from special narrow domain models to general intelligent systems. Its technical background is rooted in breakthroughs in multiple fields and is driven by the synergy of three core elements: algorithms, computing power, and data. With the rapid development of artificial intelligence technology, agents (Agents) as software entities with autonomous perception, decision-making, and execution capabilities are evolving from traditional rule-driven to data-driven. However, existing agent technology still faces significant bottlenecks in multi-modal interaction, dynamic environment adaptability, accuracy of professional domain retrieval, and large-scale task generalization ability, and needs to be optimized and innovated through technology integration to achieve breakthroughs.

[0003] For example, the invention patent application with publication number CN119578559A discloses an agent automatic configuration method and system based on large models, knowledge bases, and tools. The method includes obtaining training data, inputting natural language to a reasoning model to generate a query task set, and optimizing and inputting the Embedding model to generate a preliminary retrieval result set; inputting the optimization result into a reordering model to generate the final retrieval result; when the training conditions are met, encapsulating the reasoning model, Embedding model, and reordering model into an agent. This technology uses pre-set training data to manually intervene to correct the query task results, which can meet the high-accuracy interaction of the pre-set scenario, but still has problems such as insufficient multi-modal interaction capability, inability to exhaustively preset user query tasks in complex data scenarios, knowledge base professional problems, and weak large-scale task generalization ability.

[0004] Patent application CN118779469A discloses a method for constructing a multimodal knowledge base based on feature representation of a domain-wide model. The method includes the following steps: Step 1, collecting and processing information for constructing the knowledge base; Step 2, extracting multimodal information features based on the information collected in Step 1; Step 3, constructing a multimodal knowledge base based on the information collected in Step 1 and the multimodal information features extracted in Step 2; Step 4, extracting user-inputted demand information based on the multimodal knowledge base constructed in Step 3, constructing a domain intent recognition model, and determining the user's specific demand intent; Step 5, extracting the user-inputted demand information, querying the knowledge base, inputting the queried knowledge blocks into the domain-wide model, and outputting the query results requested by the user using the generation capabilities of the large model. This technology solves the problems of insufficient multimodal interaction capabilities and poor knowledge base professionalism by constructing a domain database for multimodal information feature extraction. However, in complex data scenarios, user query tasks cannot exhaustively cover all presets, resulting in low retrieval accuracy. Furthermore, the large model is easily limited by the length of the context, leading to weak generalization ability for large-scale tasks.

[0005] Currently, while some AI agents attempt to integrate multimodal information, the degree of modal fusion is limited, and the algorithms and engineering challenges of multimodal large-scale models are greater. How to effectively integrate open-source large-scale models with private domain data agent management platforms, and establish a seamless, fully autonomous, and controllable development path for the entire chain of large-scale model-agent development-agent management-agent deployment, conveniently meeting the needs of government and enterprise units to build private large-scale model agent applications, is a pressing issue that needs to be addressed. Summary of the Invention

[0006] The purpose of this invention is to provide an autonomous and controllable heterogeneous data-driven intelligent agent collaborative optimization system and method to solve the problems of existing intelligent agents, such as the inability to exhaustively search presets for query tasks, low retrieval accuracy, and the fact that large models in intelligent agents are easily limited by context length and have weak generalization ability for large-scale tasks.

[0007] To achieve the above-mentioned objectives, an embodiment provides an autonomous and controllable heterogeneous data-driven intelligent agent cooperative optimization system, comprising:

[0008] The environment deployment module is used to deploy the inference engine for user question-answering reasoning based on large models and the functional components for data processing based on small and medium-sized models.

[0009] A private domain knowledge database, which contains knowledge information about areas of interest built based on functional components;

[0010] The intelligent agent development and management collaborative platform is used for developing a multi-task intelligent agent based on functional components and reasoning engines, performing iterative question and answer of intent recognition, query rewriting, dynamic planning, retrieval and recall coarse screening of the multi-task intelligent agent based on a private domain knowledge database, and secondary fine arrangement of coarse screening results, so as to realize collaborative optimization of the multi-task intelligent agent based on autonomous and controllable heterogeneous data driving.

[0011] In one embodiment, the private domain knowledge database is constructed in the following manner:

[0012] Based on a basic database containing a relational database and a knowledge base document, the functional components are used to process the introduced unstructured business data of the field of interest into structured data, and the structured tags are labeled and supplemented to obtain the private domain knowledge database.

[0013] In one embodiment, the small and medium-sized model for data processing includes at least one of a document parsing model, an OCR recognition model, a vector embedding model, a rearrangement model, a speech-to-text model, and a text-to-speech model.

[0014] In one embodiment, the intent recognition and query rewriting optimization of the multi-task intelligent agent based on the private domain knowledge database includes:

[0015] The multi-task intelligent agent uses an intent recognition model to obtain the user intent based on the query question input by the user, and combines the understanding of the private domain knowledge database and the context semantics of the query question to rewrite the intent and the query question.

[0016] In one embodiment, the dynamic planning optimization of the multi-task intelligent agent based on the private domain knowledge database includes:

[0017] Based on the rewritten intent and query question, the multi-task is decomposed into sub-tasks, the execution order of each sub-task is dynamically adjusted, and the execution path is determined based on the real-time feedback of the user to complete the dynamic planning, the execution path including identifying the rewritten intent and query question, and determining to call the private domain knowledge database for retrieval or directly trigger the task according to the identification result.

[0018] In one embodiment, the retrieval and recall coarse screening optimization of the multi-task intelligent agent based on the private domain knowledge database includes:

[0019] The multi-task intelligent agent uses a screening model to perform answer retrieval and recall of the user intent and query question based on the execution path determined by the dynamic planning, and performs statistical coarse screening of the query question answer based on the candidate results of the answer retrieval and recall.

[0020] In one embodiment, the iterative question and answer optimization of the multi-task intelligent agent based on the private domain knowledge database for secondary fine arrangement of coarse screening results includes:

[0021] The multi-task intelligent agent adopts a knowledge question and answer model, and based on the coarse screening result, iteratively asks and answers based on the private domain knowledge database and user feedback, to realize secondary precision screening of the coarse screening result to obtain the final answer document.

[0022] In one embodiment, based on the coarse screening result, iteratively asking and answering based on the private domain knowledge database and user feedback, to realize secondary precision screening of the coarse screening result to obtain the final answer document, comprising:

[0023] In one iteration of the question and answer process, the metadata score is calculated, and the metadata in the question and answer corresponding to the coarse screening result is matched with the data in the private domain knowledge database, and based on the entity matching score and considering the time decay condition and the original score of the question and answer in the coarse screening result based on the small and medium-sized model, the metadata score of the question and answer is obtained;

[0024] At the same time, the content score is calculated, and specifically, the question and answer corresponding to the coarse screening result is divided by sentence, and the sentence score is calculated after supplementing the sentence to improve the coherence of the sentence through a sliding window, and the attention weight of the sentence and the closed-loop feedback weight of the user to the sentence are considered to weight all sentence scores to obtain the content score of the question and answer;

[0025] Based on the original score, metadata score, and content score of the question and answer, the final score of the question and answer is obtained by weighted summation through entropy weight method;

[0026] Based on the final score of the question and answer, after context compression of all sentences in the question and answer, the sentences are sorted again based on the final score and the top K sentences are selected to form the final answer as the output of the current question and answer;

[0027] When the score of the answer of the current question and answer does not meet the preset threshold, the next iteration of the question and answer is performed.

[0028] In one embodiment, the metadata score of the question and answer is obtained based on the entity matching score, considering the time decay condition, and the original score of the question and answer in the coarse screening result, comprising:

[0029] ,

[0030] ,

[0031] wherein, represents the metadata score of the question and answer; represents the data score calculated based on the small and medium-sized model, represents the content score calculated based on the small and medium-sized model, and and are normalized; Time score considering time decay, where day difference represents the difference between the current date and the date of the answer to the question; Day difference threshold, adjusted according to actual conditions; max represents that the day difference is not more than Day difference is calculated according to the day difference, and the day difference is more than Days are calculated as 0; Time decay factor weight, used to ensure the auxiliary nature of the time decay factor and avoid excessive rewards for fresh content, The value range is 0.2-0.5, adjusted according to actual conditions; Entity matching score.

[0032] In one embodiment, the sentence score is calculated after improving the coherence of the sentence by supplementing the sentence with a sliding window, including:

[0033] The sentence supplemented by the sliding window is input into the Transformer, and the sentence score and the attention weight of each sentence are calculated based on the Transformer.

[0034] In one embodiment, the content score of the question answer is obtained by weighting all sentence scores considering the attention weight of the sentence and the closed-loop feedback weight of the user to the sentence, including:

[0035] ,

[0036] ,

[0037] Wherein, The content score of the question answer obtained by weighting all sentence scores; The comprehensive weight of the th sentence is determined by the attention weight of the th sentence and the closed-loop feedback weight of the user to the sentence; The sentence score calculated based on the Transformer after inputting the sentence supplemented by the sliding window into the Transformer.

[0038] In one embodiment, the closed-loop feedback weight is obtained by the following method:

[0039] Input the feedback data of the user to the sentence F={likes, dislikes, dwellTime, timestamp}; where likes represents the number of likes of the user to the sentence, reflecting the positive feedback of the user; dislikes represents the number of dislikes of the user to the sentence, reflecting the negative feedback of the user; dwellTime represents the dwell time of the user on the sentence, reflecting the degree of interest of the user; timestamp represents the timestamp of the user feedback behavior, used to consider the time decay case;

[0040] Based on the number of likes of the user to the sentence likes, the number of dislikes of the user to the sentence dislikes and the dwell time of the user on the sentence dwellTime, the comprehensive reward for reflecting the overall satisfaction of the user to the sentence is calculated

[0041]

[0042] where dwellTime represents the dwell time, is a reasonable threshold value formed based on statistical analysis of user behavior data, reflecting the depth of user participation in content, is 60 seconds or 120 seconds, which can be adjusted appropriately according to the business scenario; min represents that the dwell time does not exceed the set threshold , the dwell time is calculated according to the dwell time, and the dwell time exceeds is calculated as 1; is the weight of the number of dislikes dislikes, ≥ 0.5, which is adjusted according to the actual situation, used to balance the positive and negative feedback of the closed-loop feedback weight;

[0043] At the same time, the time decay coefficient is calculated based on the timestamp of the user feedback behavior

[0044]

[0045] where the time is calculated in days, the day difference represents the interval in days between the current time and the timestamp of the user feedback behavior, represents the knowledge freshness threshold in the selected time, is 30 days, 365 days or 1825 days, used to represent long-term effective knowledge; γ represents the weight reserved for feedback exceeding the threshold , used to balance the long-term stable content; represents that when the day difference does not exceed the threshold , the day difference is calculated, and the day difference exceeds 30 days is calculated, the set threshold and the weight are adjusted according to the actual situation, used to balance the positive feedback of the closed-loop feedback weight;​​​​

[0046] Based on comprehensive rewards and time decay coefficient Calculate the closed-loop feedback weights :

[0047] ,

[0048] Where 1.0 is the base weight, representing no feedback or a feedback contribution of 0; It is used to balance the weight of user likes and dislikes. ≥0.1, which varies depending on whether the user explicitly likes or dislikes the post.

[0049] In one embodiment, the final score of the question answer is obtained by weighted summation using the entropy weighting method based on the original score, metadata score, and content score of the question answer, including:

[0050] Based on the original score of the question answer, the metadata score, and the content score, normalized scores for these three dimensions are calculated. ,in, The dimension index represents the score type. The metrics for calculating scores for each dimension. Indicates the first The raw score of the answer to the question under each calculation indicator. Indicates the first Metadata scores for question answers under each computational metric Indicates the first Content scores for the answers to questions under each calculation indicator; Indicates the first The calculation index is in the first The normalized scores for each dimension are represented in the form of a probability distribution.

[0051] Based on normalized score Calculate the entropy value ,in, These are the three dimensions used when calculating entropy;

[0052] Based on entropy Assign weights to the scores of the three dimensions. Weights of scores across three dimensions The final score for the answer to the question is obtained by weighted summation, where It is a dimensional index. It represents the sum of entropy values ​​across the three dimensions: raw score, metadata score, and content score.

[0053] In one embodiment, the final score of the question answer is used to perform context compression on all sentences in the question answer, including:

[0054] The length of each sentence in the question answer is calculated, the total number of sentences in the question answer and the maximum length limit of the sentence are determined, and the value score of each sentence is calculated considering the attention weight of the sentence;

[0055] The highest score of the sentence within the maximum length limit is calculated by a dynamic programming formula:

[0056] ,

[0057] Wherein, represents the highest score of the sentence within the maximum length limit, represents the number of sentences considered, , represents the total number of sentences; represents the current sentence length limit, , represents the maximum length limit of the sentence; represents the highest score of the first sentences under the length limit , represents the highest score of the first sentences under the length limit and accumulates the value score of the sentence , represents taking the maximum value of and as the highest score of the sentence within the maximum length limit;

[0058] The highest score corresponding to the combination of sentences is screened out to form the final compressed text.

[0059] In one embodiment, when the next round of iterative question answering is performed, the metadata information in the private domain knowledge database is also optimized, specifically by using a knowledge question answering model to optimize the metadata information in the private domain knowledge database through multi-level data tracing, wherein the metadata information includes label information, knowledge graph entity or document name, and the optimized metadata information is used to guide the retrieval of context to generate high-quality question answering answers.

[0060] In one embodiment, the intelligent agent development and management collaborative platform also implements management of multi-task intelligent agents, including:

[0061] The application lifecycle management is applied to determine the application target and architecture of the multi-task agent, and then the multi-task agent is deployed to a production environment, the application process of the multi-task agent is monitored in real time, and abnormal conditions are handled, the multi-task agent is updated in function and optimized in performance according to user feedback and business requirements, and when the multi-task agent cannot meet the business requirements or there is a better alternative, the multi-task agent is retired and replaced;

[0062] The personalized recommendation problem setting is used to collect basic information, historical behavior data and preference data of the user, predict problems interested by the user through the multi-task agent, and generate a problem recommendation list, dynamically adjust the private domain knowledge database and set personalized recommendation problems according to real-time feedback and behavior changes of the user;

[0063] The session record management is used to store the session content when the user uses the multi-task agent to the private domain knowledge database, and classify and index the session content each time, the management personnel of the multi-task agent analyzes the user demand and behavior mode based on the recorded session content, updates the private domain knowledge database, and realizes the optimization of the multi-task agent.

[0064] The application further provides an autonomous controllable heterogeneous data driven agent collaborative optimization method, the agent collaborative optimization method adopts the autonomous controllable heterogeneous data driven agent collaborative optimization system, and comprises the following steps:

[0065] Deploy a reasoning engine based on a large model for knowledge reasoning, and a functional component based on a small and medium-sized model for data processing;

[0066] A private domain knowledge database containing knowledge information of the field of interest constructed based on the functional component is constructed;

[0067] A multi-task agent is developed based on the functional component and the reasoning engine by using the agent development and management collaborative platform, and the multi-task agent is sequentially subjected to intention recognition, query rewriting, dynamic planning, retrieval recall coarse screening, and iteration of secondary fine screening of coarse screening results based on the private domain knowledge database, so that the multi-task agent is subjected to collaborative optimization based on autonomous controllable heterogeneous data driving.

[0068] Compared with the prior art, the application has at least the following beneficial effects:

[0069] (1) The private domain knowledge database provided collects and arranges industry knowledge to form a basic, systematic, correlated and easy-to-use knowledge set, can realize multi-modal data interaction, provides high-quality data for multi-task agent development and optimization, enables the multi-task agent to execute within a limited target data range, and improves retrieval quality.

[0070] (2) The built intelligent agent development and management collaborative platform deploys a reasoning engine based on a large model for knowledge reasoning and functional components based on small and medium-sized models for data processing, ensuring full-link autonomous controllability and enabling full-process management of intelligent agent research and development.

[0071] (3) The multi-task intelligent agent integrated in the intelligent agent development and management collaborative platform can accurately identify the query intention of the user, introduce domain knowledge for fusion query, and perform multi-round iterative retrieval on the retrieval results, effectively reducing document interference and significantly improving the efficiency of problem retrieval, solving the problem of inaccurate retrieval, context length limitation, and reducing the occurrence of incorrect answers to complex problems.

[0072] (4) During the retrieval process, feedback from the user and domain knowledge are used to achieve closed-loop optimization, significantly improving the accuracy of problem retrieval.

[0073] (5) The intelligent agent development and management collaborative platform supports flexible conversation window management, smooth streaming conversation experience, and deep thinking mode, improving the user's experience.

[0074] (6) The intelligent agent development and management collaborative platform has built-in interfaces for interfacing with multiple third-party platforms, allowing for convenient multi-channel publishing and meeting the needs of intelligent agent application services in different business environments. BRIEF DESCRIPTION OF DRAWINGS

[0075] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description.

[0076] Figure 1 The structure diagram of the autonomous controllable heterogeneous data driven intelligent agent collaborative optimization system provided for the embodiment;

[0077] Figure 2 The process diagram of the multi-task intelligent agent using a knowledge question and answer model for iterative question and answer provided for the embodiment;

[0078] Figure 3 The process diagram of the autonomous controllable heterogeneous data driven intelligent agent collaborative optimization method provided for the embodiment. DETAILED DESCRIPTION

[0079] In order to make the purpose, technical solutions and advantages of the present application more clear and explicit, the following will combine the drawings and specific embodiments to further describe the present application.

[0080] The embodiment further describes the present application. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the protection scope of the present application.

[0081] The embodiment provides an autonomous controllable heterogeneous data-driven intelligent agent collaborative optimization system, which comprises an environment deployment module, a private domain knowledge database and an intelligent agent development and management collaborative platform, as shown in Figure 1

[0082] In the embodiment, the environment deployment module is used for deploying corresponding resources in a China software and information service (CSIS) hardware environment. A computing power GPU adopts Huawei Ascend 910B, a database adopts a domestic database such as Renmin University of China Golden Storehouse, Haige and Dameng, an operating system adopts Galaxy Kirin V10, a middleware adopts a domestic middleware such as Dongfang Tong and Kingsoft, a private deployment open source large model such as DeepSeek R1 or Qwen2.5 32B is used as a reasoning engine for knowledge reasoning, and a series of small and medium-sized models required by a business, such as a document parsing model, an OCR identification model, a vector embedding model (Emebedding model), a rearrangement model (Rerank model), speech-to-text and text-to-speech, are used as functional components for data processing, so as to ensure that the reasoning ability of a multi-task intelligent agent and the running environment are autonomous and controllable.

[0083] Then, the private domain knowledge database is constructed. A basic database including a relational database and a knowledge base document is used as a basis. The knowledge base document supports storage according to types, so as to ensure that the business data range searched by a multi-task intelligent agent is autonomous and controllable. Then, business data is imported. The business data includes knowledge information in a field of interest. The field of interest can be a government website, which provides government information query and policy interpretation, such as social security policy, tax policy and administrative examination and approval process, to facilitate citizens to do things and improve the transparency and convenience of government services; a government hotline, which assists government hotline staff to quickly respond to inquiries and complaints of citizens and provides accurate information and solutions; intelligent education, which provides students with course-related question answers such as knowledge point explanation, homework problems and examination preparation, searches appropriate learning resources and tutoring strategies according to the learning situation and problems of students, and provides personalized learning support for students; medical health, which answers common questions of patients about disease symptoms, prevention measures, treatment methods and drug use, and provides preliminary medical advice and health guidance; helps medical researchers to quickly search relevant medical literature, research results and clinical cases; and intelligent office, which helps employees to quickly search information such as company regulations, business processes and project documents, but is not limited to this.

[0084] ETL processing is performed on the imported business data. Functional components are used to process the imported unstructured business data in the field of interest into structured data. For example, OCR identification, speech-to-text and Markdown format conversion are used to retain layout information. Then, structured labels are generated through data labeling. The labeling results are stored in a vector in a database, and the private domain knowledge database is supplemented.

[0085] ​The structured data is analyzed by wide table fusion of a topic library and a special topic library, and documents in a private domain knowledge database are automatically sliced by using multiple document slicing strategies. Different slicing strategies are set, including general segmentation, question and answer pair segmentation, parent-child structure segmentation, and semantic segmentation based on a large model. The segmented result is vectorized and stored in a vector database to support efficient document retrieval and recall, and to optimize the processing efficiency of the multi-task agent.

[0086] In the intelligent agent development and management collaborative platform, the deployed functional components and inference engines can cooperate to develop multi-task agents. The private domain knowledge database is used to iteratively ask and answer the multi-task agent for intent recognition, query rewriting, dynamic planning, coarse screening of retrieval and recall, and secondary fine screening of coarse screening results, to realize collaborative optimization of the multi-task agent based on autonomous and controllable heterogeneous data driving.

[0087] In the embodiment, the multi-task agent uses an intent recognition model to obtain the user's intent based on the user's input query question, and combines the private domain knowledge database and the understanding of the context semantics of the query question to rewrite the intent and the query question.

[0088] Based on the private domain knowledge database, the multi-task agent is dynamically planned and optimized, specifically including: based on the rewritten intent and query question, the multi-task is decomposed into sub-tasks, the execution order of each sub-task is dynamically adjusted, and the execution path is determined by combining the real-time feedback of the user to complete dynamic planning. The execution path includes identifying the rewritten intent and query question, and determining to call the private domain knowledge database for retrieval or directly trigger the task according to the identification result.

[0089] In the embodiment, the multi-task agent uses a screening model, based on the execution path determined by dynamic planning, through natural language to SQL, combines the private domain knowledge database to retrieve and recall the answer of the user's intent and query question, and based on the candidate results of the answer retrieval and recall, the statistical coarse screening of the query question answer is performed.

[0090] In the embodiment, the private domain knowledge database is used to iteratively ask and answer the multi-task agent based on the coarse screening result, including: the multi-task agent uses a knowledge question and answer model, based on the private domain knowledge database and user feedback, iteratively asks and answers based on the coarse screening result, to achieve secondary fine screening of the coarse screening result to obtain the final answer document.

[0091] Specifically, as shown in Figure 2 In the process of one iteration of asking and answering, the metadata score is calculated. Specifically, the metadata in the question answer corresponding to the coarse screening result, such as the time decay factor, the domain classification, the document title, the knowledge base category in the private domain database, the custom label or the knowledge graph entity, is matched with the data in the private domain knowledge database based on the entity list ​query key-value Compute entity matching score , is a scaling coefficient, determined by sample set tuning, based on entity matching score and considering time decay and original score of the primary sorting of the question answer in the coarse screening result, to obtain the metadata score of the question answer, including:

[0092] ,

[0093] ,

[0094] wherein, represents the metadata score of the question answer; represents the data score calculated based on the small and medium models, represents the content score calculated based on the small and medium models, and and are normalized; is a time score considering time decay, wherein the day difference represents the difference between the current date and the publication date of the question answer; is a day difference threshold, which is adjusted according to actual conditions; max represents that the day difference is calculated according to the day difference when the day difference is not more than days, and the day difference is calculated as 0 when the day difference is more than days; is a time decay factor weight, which is used to ensure the auxiliary nature of the time decay factor and avoid excessive reward of fresh content, the value range of which is 0.2-0.5, which is adjusted according to actual conditions; is the entity matching score. In the embodiment, is 30 days, is 0.2. Through entity linking and relationship reasoning, the semantic understanding ability is improved, and the semantic correlation of the multi-hop problem is solved.

[0095] Meanwhile, the content score is calculated, specifically, the question answer corresponding to the coarse screening result is divided into a sentence list ( is a single sentence score), and the sentence is supplemented through a sliding window W to improve the coherence of the sentence, wherein the sliding window W=2, the supplemented sentence list is output to the Transformer, the supplemented sentence score and the attention weight of each sentence are calculated based on the Transformer, the sentence correlation evaluation is enhanced, and the attention weight of the sentence and the closed-loop feedback weight of the user to the sentence are considered to weight all the sentence scores to obtain the content score of the question answer, and the sentence selection is dynamically optimized:

[0096] ,

[0097] ,

[0098] wherein, represents the content score of the question answer obtained by weighting all sentence scores; represents the comprehensive weight of the th sentence, which is determined by the attention weight of the th sentence and the closed-loop feedback weight of the sentence by the user; represents the sentence score calculated based on the input of the sliding window supplemented sentence into the Transformer, , is the single sentence score.

[0099] The closed-loop feedback weight is obtained by inputting the feedback data of the user to the sentence F ={likes, dislikes, dwellTime, timestamp}, wherein likes represents the number of likes of the user to the sentence, reflecting the positive feedback of the user; dislikes represents the number of dislikes of the user to the sentence, reflecting the negative feedback of the user; dwellTime represents the dwell time of the user on the sentence, reflecting the interest degree of the user; and timestamp represents the timestamp of the user feedback behavior, used to consider the time decay condition.

[0100] Based on the number of likes of the user to the sentence likes, the number of dislikes of the user to the sentence dislikes and the dwell time of the user on the sentence dwellTime, the comprehensive reward for reflecting the overall satisfaction of the user to the sentence is calculated:

[0101] ,

[0102] wherein dwellTime represents the dwell time, is a reasonable threshold value formed based on statistical analysis of user behavior data, reflecting the depth participation of the user to the content, is 60 seconds or 120 seconds, which can be appropriately adjusted according to the business scenario; min represents that when the dwell time does not exceed the set threshold value , the dwell time is calculated, and when the dwell time exceeds , 1 is calculated; is the weight of the number of dislikes dislikes, ≥ 0.5, which is adjusted according to the actual situation, used to balance the positive and negative feedback of the closed-loop feedback weight; in the embodiment, is 60 seconds, is 0.5. ​

[0103] Meanwhile, the time decay coefficient is calculated based on the timestamp of the user feedback behavior :

[0104] ,

[0105] wherein the time is calculated in days, the day difference represents the interval days between the current time and the timestamp of the user feedback behavior, represents the knowledge freshness threshold in the selected time, is 30 days, 365 days or 1825 days, used to represent the long-term effective knowledge; represents the weight reserved for the feedback exceeding the threshold , used to balance the long-term stable content; represents the day difference, calculated by the day difference when the day difference does not exceed the threshold , calculated by when the day difference exceeds 30 days, the set threshold and the weight are adjusted according to the actual situation, used for positive feedback of the closed-loop feedback weight; in the embodiment, is 30 days, is 0.5.

[0106] Based on the comprehensive reward and the time decay coefficient , the closed-loop feedback weight is calculated:

[0107] ,

[0108] wherein 1.0 is the basic weight, representing no feedback or feedback contribution of 0; is the weight for balancing the number of likes and dislikes, ≥ 0.1, which changes according to the user's explicit choice of likes or dislikes; in the embodiment, is 0.1.

[0109] Through the above multi-dimensional user feedback, the weight is dynamically adjusted to realize the closed loop, so that the intelligent agent can be cooperatively optimized and autonomously controlled, and can better adapt to the change of user preferences.

[0110] The normalized score of the three dimensions is calculated based on the original score, the metadata score and the content score of the question and answer , wherein is the dimension index representing the score type, is the calculation index of each dimension score, represents the original score of the question and answer under the th calculation index, a metadata score of the answer to the question under a first a content score of the answer to the question under a second a normalized score of the answer to the question under a third dimension, in the form of a probability distribution

[0111] based on the normalized score , an entropy value is calculated , wherein are three dimensions in calculating the entropy value, in the embodiment, d1, d2 and d3 are respectively the original score, the metadata score and the content score is 3.

[0112] based on the information entropy value of the score distribution weights of the three dimensions of scores a final score of the answer to the question is obtained by weighted summation of the weights of the three dimensions of scores, wherein is a dimension index k represents a sum of entropy values under three dimensions of original score, metadata score and content score.

[0113] based on the final score of the answer to the question, context compression is performed on all sentences in the answer to the question, specifically, an occupied length of each sentence in the answer to the question is calculated, a total number of sentences in the answer to the question and a maximum length limit of the sentences are determined, and a value score of each sentence is calculated considering an attention weight of the sentence

[0114] by a dynamic programming formula, a highest score of the sentence within the maximum length limit is calculated

[0115] ,

[0116] wherein represents the highest score of the sentence within the maximum length limit represents a number of considered sentences , represents a total number of sentences represents a current length limit of the sentence , represents a maximum length limit of the sentence represents a highest score of the first k sentences under a length limit of k represents a highest score of the first k sentences under a length limit of k represents a value score of the sentence accumulated , represents a highest score of the first k sentences under a length limit of k represents a highest score of the first k sentences under a length limit of k , represents a maximum length limit of the sentence​​​​ and The maximum value of the two as the highest score of the sentence within the maximum length limit;

[0117] And filter out the highest score corresponding to the sentence combination to form the final compressed text. By selecting the sentence combination under the length constraint, the total score is maximized and the content compression retains the key information, achieving efficient compression of content while retaining key information and getting rid of the context length limit of large models.

[0118] Then the sentences are sorted again based on the final score and the top K sentences are selected to form the final answer as the output of the current question and answer, completing a question and answer.

[0119] The matching degree of the output of the current question and answer and the user question is evaluated by the reasoning engine deployed in the intelligent agent development and management collaborative platform. If the relevance is lower than the preset threshold, a multi-round iterative question and answer is triggered to gradually optimize the document relevance. In the next round of iterative question and answer, the metadata information in the private domain knowledge database is also optimized. Specifically, the knowledge question and answer model is used to optimize the label information, knowledge graph entity or document name through multi-level data tracing in the private domain knowledge database. The optimized metadata information is used to guide the retrieval of context to generate high-quality question and answer answers. And if there is a lot of interference information in the iterative retrieval process, the interference files can be continuously eliminated to further improve the accuracy of document hits. For example, by self-evaluation, the documents related to dog breeding in the opinion and suggestion database and the documents related to dog breeding in the conference file database are eliminated, so that the query task directly hits the documents in the legal regulations database, ensuring the accuracy of the content and reducing the generation of hallucinations. Multi-level data tracing includes document tracing, structured data tracing, and syntax-level tracing of generated text. Through visual methods such as badges and pop-ups in the intelligent agent development and management collaborative platform, users can accurately trace the source of information, and finally accurately generate statistical results, text information and analysis results. The output results are automatically connected to the dialog window and integrated into the interface of the third-party platform. The multi-task intelligent agent displays the reasoning process of the large model through the dialog window and accurately finds the knowledge information of interest and lists it. Finally, the listed knowledge information is provided to the user for review in the form of a file link, making it convenient to realize multi-channel publishing to meet the needs of intelligent agent application services in different business environments.

[0120] The intelligent agent development and management collaborative platform also realizes the management of multi-task intelligent agents to ensure the autonomous controllability of the development, management and release of intelligent agents, including:

[0121] The application lifecycle management is applied to determine the application target and architecture of the multi-task agent, and then the multi-task agent is deployed to a production environment; the application process of the multi-task agent is monitored in real time after deployment, and abnormal conditions are handled; the multi-task agent is updated in function and optimized in performance according to user feedback and business requirements; when the multi-task agent cannot meet the business requirements or there is a better alternative, the multi-task agent is retired and replaced;

[0122] The personalized recommendation problem setting is used to collect the basic information, historical behavior data and preference data of the user, predict the problems interested by the user through the multi-task agent, and generate a problem recommendation list; the private domain knowledge database is dynamically adjusted and the personalized recommendation problem is set according to the real-time feedback and behavior change of the user;

[0123] The session record management is used to store the session content when the user uses the multi-task agent to the private domain knowledge database, and classify and index the session content each time; the management personnel of the multi-task agent analyzes the user demand and behavior mode based on the recorded session content, updates the private domain knowledge database, and realizes the optimization of the multi-task agent.

[0124] The application further provides an autonomous controllable heterogeneous data driven agent collaborative optimization method, as shown in Figure 3 The autonomous controllable heterogeneous data driven agent collaborative optimization system comprises the following steps:

[0125] Deploy a reasoning engine for knowledge reasoning based on a large model, and a functional component for data processing based on a small or medium-sized model;

[0126] A private domain knowledge database containing knowledge information of the field of interest constructed based on the functional component is constructed;

[0127] The multi-task agent is developed based on the cooperation of the functional component and the reasoning engine by using the agent development and management collaborative platform, and the multi-task agent is sequentially subjected to intent recognition, query rewriting, dynamic planning, and candidate results of recall retrieval and coarse screening, and the results after coarse screening are subjected to secondary sorting and iteration of question and answer, so that the multi-task agent is collaboratively optimized based on autonomous controllable heterogeneous data driving.

[0128] The agent development and management collaborative platform deployed by the reasoning engine for knowledge reasoning based on a large model and the functional component for data processing based on a small or medium-sized model breaks through the whole-link private deployment and completely autonomous and controllable development path of a large model-agent development-agent management-agent release, supports efficient integration of private domain data, can conveniently construct a private large model agent, greatly reduces the development threshold of the agent application, and accelerates the research and development and landing process of intelligent application.

[0129] The following is an example of how the algorithm solves the problem with the user question "Please analyze and interpret the information company's work in informationization, digitization, and intelligentization."

[0130] 1. Question

[0131] (1) Query: query="Please analyze and interpret the information company's work in informationization, digitization, and intelligentization

[0132] (2) Input data: callback_results = [{"metadata": {"score": 0.63,"dataset_name": "meeting file library","document_name": "information company annual work report.PDF","scenario": "informationization","labels": "fourteen meetings, information company, annual work report","biz_time": "2024-03-01","entities": ["information company", "work report", "meeting file"]},"content": "Information company is responsible for group informationization and digitization work."},

[0133] {"metadata": {"score":0.82,"dataset_name": "group file library","scenario": "digital reform","document_name": "group 2024 work points.pdf","labels": "plan file, group, work points","biz_time": "2024-03-01","entities": ["group", "plan file"]},"content": "Plan to promote data center construction in 2024, break down data barriers between business systems, and achieve data aggregation. …"},

[0134] {"metadata": {"score": 0.78,"dataset_name": "Institutional Document Repository","scenario": "Government Work","document_name": "(2023) Implementation Measures for Further Improving Algorithm-Driven Intelligent Customer Service.pdf","labels": "Intelligentization, Information Company, Implementation Measures","biz_time": "2024-03-01","entities": ["Information Company", "Implementation Measures"]},"content": "Applying natural language processing, image recognition, knowledge graph and other technologies to achieve intelligent customer service, intelligent quality inspection, public opinion analysis, etc. ……"},

[0135] {"metadata": {"score": 0.45,"dataset_name": "Group Archived Document Library","scenario": "Digital Reform","document_name": "Group General Document No. 41.pdf","biz_time": "2023-12-01","labels": "Group Document, No. 41","entities": ["Group", "General Document"]},"content": "Promote the application of BI tools to achieve various reports, dashboards, and business visualization monitoring. ……"}]

[0136] feedback_data (user feedback data) = {"likes": 5, # 5 likes"dislikes": 1, # 1 dislike"dwell_time": 120, # User dwell time on the sentence 120 seconds"timestamp": "2024-03-20T10:00:00" # Timestamp of the user feedback behavior}.

[0137] 2. Solution Steps

[0138] Step 1: Calculate the metadata score (Rerank_metadata)

[0139] (1) Input: Metadata in the query q (query) and the question answers corresponding to the coarse screening results.

[0140] (2) Process:

[0141] Time decay: ;

[0142] First three (2024-03-01): about 26 days ago, ;

[0143] Fourth (2023-12-01): about 117 days ago, ;

[0144] Knowledge graph enhanced computing entity matching score:

[0145] First: match "Information Company" "Work Report" "Meeting Document", ;

[0146] Second: match "Group" "Plan Document", ;

[0147] Third: match "Information Company" "Implementation Measures", ;

[0148] Fourth: match "Group" "Pu Fa Document", ;

[0149] Result (assuming the Rerank service returns the base score consistent with score):

[0150] First: ;

[0151] Second: ;

[0152] Third: ;

[0153] Fourth: ;

[0154] Step 2: Compute feedback weight (compute_feedback_weight)

[0155] (1) Input: feedback_data (user feedback data)

[0156] (2) Process:

[0157] Comprehensive reward:

[0158] Time decay coefficient: 7 days ago,

[0159] Closed-loop feedback weight:

[0160] Limit range:

[0161] Step 3: Calculate content score and compression (Rerank_content)

[0162] (1) Input: query q and question answers corresponding to the coarse screening results.

[0163] (2) Process:

[0164] Sliding window: divide the question answers corresponding to the coarse screening results into a sentence list by sentence, and supplement each sentence in the sentence list with a window.

[0165] First: "Information Company is responsible for the group's informationization and digitization." → "Information Company is responsible for the group's informationization and digitization. Currently, the network, server, database, middleware, and other infrastructure have been built, realizing office automation and electronic business systems."

[0166] Self-attention: calculate the attention weight of each sentence based on Transformer .

[0167] Reinforcement learning: apply closed-loop feedback weight , adjust the weight .

[0168] Dynamic programming: select the highest scoring sentence combination.

[0169] Result:

[0170] First: , compressed to "Information Company is responsible for the group's informationization and digitization. Currently, office automation and electronic business systems have been realized. …"

[0171] Second: , compressed to "Plan to promote data platform construction in 2024, break down data barriers between business systems, and realize data aggregation. …"

[0172] Third: , compressed to "Apply natural language processing, image recognition, knowledge graph, and other technologies to realize intelligent customer service, intelligent quality inspection, and public opinion analysis. …"

[0173] Fourth: , compressed to "Promote the application of BI tools to realize various reports, dashboards, and business visualized monitoring. …"

[0174] Step 4: Fusion score (compress_content)

[0175] (1) Input: original score, metadata score, content score.

[0176] (1) Process:

[0177] Score list:

[0178]

[0179]

[0180]

[0181] Entropy weight: , ,

[0182] Synthetic score:

[0183] First:

[0184] Second:

[0185] Third:

[0186] Fourth:

[0187] Step 5: Normalize and sort (normalize_scores)

[0188] (1) Process:

[0189] Normalization: , , ;

[0190] First: ;

[0191] Second: ;

[0192] Third: ;

[0193] Fourth: ;

[0194] (2) Result: The second article has the highest score and ranks at the front.

[0195] Step 6: Output Top K results

[0196] (1) Input: ;

[0197] (2) Output: {"result": [{"metadata": {"score": 0.82, "dataset_name": "Group File Library", "scenario": "Digital Reform", "document_name": "Group General Document No. 41.pdf", "labels": "Plan Document, Group, Key Points", "biz_time": "2024-03-01", "entities": ["Group", "Plan Document"]}, "content": "xxxxxxxx", "final_score": 1.0}]}.

[0198] In summary, the autonomous controllable heterogeneous data driven agent collaborative optimization method and device designed by the present application has the core optimization functions of integrated model unified management, visual workflow engine, flexible tool calling mechanism, powerful knowledge base engine, convenient question and answer quality feedback mechanism, intelligent intent recognition management, fine knowledge base management, personalized document slicing strategy, multiple domain knowledge recall algorithm, natural language to SQL, multi-path recall, recall result self-supervision evaluation, user query rewriting, named entity recognition, etc. It solves the problem of inaccurate retrieval and context length limitation in current knowledge base retrieval, which leads to many complex problem answers being incorrect. At the same time, it ensures the autonomy and controllability of the development, management and release of the intelligent agent, greatly reduces the development threshold of intelligent agent application, and accelerates the research and development and landing process of intelligent application.

[0199] The above specific embodiments have described the technical solutions and advantages of the present application in detail. It should be understood that the above description is only the most preferred embodiment of the present application and is not intended to limit the present application. Any modifications, supplements and equivalent replacements made within the principle range of the present application should be included in the protection scope of the present application.

Claims

1. An autonomously controllable heterogeneous data-driven agent coordination optimization system, characterized in that, Comprise: An environmental deployment module for deploying an inference engine based on large models for user question and answer reasoning, and functional components based on small and medium-sized models for data processing; A private domain knowledge database containing knowledge information of the field of interest based on the functional components, including: based on the basic database containing relational databases and knowledge base documents, using the functional components to process the introduced unstructured business data of the field of interest into structured data, and supplementing after labeling structured tags to obtain a private domain knowledge database; An intelligent agent development and management collaborative platform for developing multi-task intelligent agents based on the functional components and the inference engine, and iteratively questioning and answering the multi-task intelligent agents based on the private domain knowledge database to realize the collaborative optimization of the multi-task intelligent agents driven by autonomous and controllable heterogeneous data. Wherein, the iterative questioning and answering comprises: in one iteration of the questioning and answering process, calculating the metadata score, specifically matching the metadata in the question answer corresponding to the rough screening result with the data in the private domain knowledge database, and based on the entity matching score and considering the time decay condition and the original score of the question answer in the rough screening result based on the small and medium-sized models, obtaining the metadata score of the question answer; Meanwhile, the content score is calculated, specifically the question answer corresponding to the rough screening result is divided by sentence, and the sentence score is calculated after supplementing the sentence to improve the coherence of the sentence through the sliding window, and the content score of the question answer is obtained by weighting all sentence scores considering the attention weight of the sentence and the closed-loop feedback weight of the user to the sentence; The final score of the question answer is obtained by weighting and summing the original score, metadata score and content score of the question answer based on the entropy weight method; After the context compression of the question answer based on the final score, the final answer is obtained by secondary sorting and screening the top K sentences as the output of the current iteration of the questioning and answering process; When the score of the answer of the current iteration of the questioning and answering process does not meet the preset threshold, the next iteration of the questioning and answering process is performed; In the next iteration of the questioning and answering process, the metadata information in the private domain knowledge database is also optimized, specifically the knowledge questioning and answering model is used to optimize the metadata information by performing multi-level data tracing in the private domain knowledge database, wherein the metadata information includes label information, knowledge graph entity or document name, and the optimized metadata information is used to guide the retrieval of context to generate high-quality question and answer answers.

2. The autonomously controllable heterogeneous data-driven agent collaborative optimization system in accordance with claim 1, wherein, The small and medium-sized models for data processing include at least one of the document parsing model, the OCR recognition model, the vector embedding model, the rearrangement model, the speech to text model, and the text to speech model.

3. The autonomously controllable heterogeneous data-driven intelligent agent collaborative optimization system in accordance with claim 1, wherein, The intent recognition and query rewriting optimization of the multi-task intelligent agent based on the private domain knowledge database comprises: The multi-task intelligent agent uses an intent recognition model to obtain the user's intent based on the user's input query question, and combines the private domain knowledge database and the understanding of the context semantics of the query question to rewrite the intent and the query question.

4. The autonomously controllable heterogeneous data-driven intelligent agent collaborative optimization system in accordance with claim 1, wherein, The dynamic planning optimization of the multi-task intelligent agent based on the private domain knowledge database comprises: Based on the rewritten intent and query question, the multi-task is decomposed into sub-tasks, the execution order of each sub-task is dynamically adjusted, and the execution path is determined to complete dynamic planning in combination with real-time feedback of the user, the execution path including identifying the rewritten intent and query question, and determining to call a private domain knowledge database for retrieval or directly triggering a task according to the identification result.

5. The autonomously controllable heterogeneous data-driven intelligent agent collaborative optimization system in accordance with claim 1, wherein, The multi-task agent is retrieved and recalled based on the private domain knowledge database for optimization, including: The multi-task agent uses a screening model to retrieve and recall answers to the user's intent and query question based on the execution path determined by dynamic planning and the private domain knowledge database, and then performs statistical screening of the query question answers based on the candidate results of the answer retrieval and recall.

6. The autonomously controllable heterogeneous data-driven intelligent agent collaborative optimization system in accordance with claim 1, wherein, The multi-task agent is iteratively questioned and answered based on the coarse screening result for secondary fine sorting, including: The multi-task agent uses a knowledge question and answer model to iteratively question and answer based on the private domain knowledge database and user feedback based on the coarse screening result, to achieve secondary fine sorting of the coarse screening result to obtain the final answer document.

7. The autonomously controllable heterogeneous data-driven intelligent agent collaborative optimization system in accordance with claim 1, wherein, The metadata score of the question answer is obtained based on the entity matching score and considering the time decay and the initial score of the question answer in the coarse screening result, including: ， ， wherein, a metadata score representing the question answer; a data score calculated based on the small and medium models, a content score calculated based on the small and medium models, and and normalized; a time score considering time decay, wherein the day difference represents the difference between the current date and the publication date of the question answer, and max represents that the day difference does not exceed days, the day difference is calculated according to the day difference, and the day difference exceeding days is calculated as 0, the day difference threshold value; the time decay factor weight, the value range of is 0.2~0.5; an entity matching score.

8. The autonomously controllable heterogeneous data-driven intelligent agent collaborative optimization system in accordance with claim 1, wherein, The sentence score is calculated after supplementing the sentence in the sliding window to improve the coherence of the sentence, including: The sentence score and the attention weight of each sentence are calculated based on the input of the supplemented sentence in the sliding window into the Transformer.

9. The autonomously controllable heterogeneous data-driven intelligent agent collaborative optimization system in accordance with claim 1, wherein, The content score of the question answer is obtained by weighting all sentence scores considering the attention weight of the sentence and the closed-loop feedback weight of the user, including: ， ， wherein, represents the content score of the question answer weighted by all sentence scores; represents the comprehensive weight of the th sentence, which is jointly determined by the attention weight of the th sentence and the user's closed-loop feedback weight to the sentence; represents the sentence score calculated based on the Transformer after the sliding window supplement sentence is input into the Transformer.​ 10. The autonomously controllable heterogeneous data-driven agent-cooperative optimization system according to claim 1 or 9, characterized in that, The closed-loop feedback weight is obtained by the following method: Input user feedback data for a sentence F = {likes, dislikes, dwellTime, timestamp}, where likes represents the number of likes of the user for the sentence, reflecting the positive feedback of the user; dislikes represents the number of dislikes of the user for the sentence, reflecting the negative feedback of the user; dwellTime represents the dwell time of the user on the sentence, reflecting the degree of interest of the user; timestamp represents the timestamp of the user feedback behavior, used to consider the time decay case; Based on the number of likes of the sentence by the user, the number of dislikes of the sentence by the user, and the dwell time of the user on the sentence, a comprehensive reward is calculated to reflect the overall satisfaction of the user with the sentence : ， Wherein, dwellTime represents the dwell time, min represents that the dwell time is not more than the set threshold , and when the dwell time is more than , 1 is calculated according to the dwell time; The weight of the point click number dislikes, , for the weight of the closed-loop feedback negative feedback; Simultaneously computing a time decay coefficient based on a timestamp of when user feedback behavior occurred : ， wherein time is calculated in days, and the day difference represents the interval in days between the current time and the timestamp of the user feedback behavior, represents the knowledge freshness threshold within the selected time, represents the weight reserved for feedback exceeding the threshold ; max represents that when the day difference does not exceed the threshold , the day difference is calculated by the day difference, and when the day difference exceeds 30 days, it is calculated by , which is used for the positive feedback of the closed-loop feedback weight; Based on comprehensive rewards and a time decay factor , calculate closed loop reward weight : ， where 1.0 is the base weight, representing no feedback or a feedback contribution of 0; is a weight for balancing the user likes and dislikes of the book, .

11. The autonomously controllable heterogeneous data-driven intelligent agent collaborative optimization system in accordance with claim 1, wherein, The final score of the question answer is obtained by weighting and summing the initial score, metadata score, and content score of the question answer by the entropy weight method, including: Based on the raw scores, metadata scores, and content scores of the question answers, the normalized scores of the three dimensions are calculated wherein, is the dimension index representing the score type, is the calculation indicator of each dimension score, is the raw score of the question answer under the th calculation indicator, is the metadata score of the question answer under the th calculation indicator, is the content score of the question answer under the th calculation indicator; is the normalized score of the th calculation indicator under the th dimension, represented in the form of a probability distribution; based on the normalized scores , calculating the entropy value wherein, three dimensions for calculating the entropy value; based on the entropy values assigning weights to the three dimension scores by the weights of the three dimension scores weighting the sum to obtain a final score for the answer to the question, wherein is a dimension index, denotes the sum of the entropy values in the original score, the metadata score, and the content score.

12. The autonomously controllable heterogeneous data-driven intelligent agent collaborative optimization system in accordance with claim 1, wherein, The context of all sentences in the question answer is compressed based on the final score of the question answer, including: The length occupied by each sentence in the question answer is calculated to determine the total number of sentences and the maximum length limit of the sentence in the question answer, and the value score of each sentence is calculated considering the attention weight of the sentence; The highest score of the sentence within the maximum length limit is calculated by the dynamic programming formula: ， wherein, represents the highest score of the sentence within the maximum length limit, represents the number of sentences considered, , represents the total number of sentences; represents the current sentence length limit, , represents the maximum length limit of the sentence; represents the highest score of the first sentences under the length limit of , represents the highest score of the first sentences under the length limit of and accumulates the value score of the sentence , max represents taking the maximum value of and as the highest score of the sentence within the maximum length limit; And the sentence combination corresponding to the highest score is selected to form the final compressed text.

13. The autonomously controllable heterogeneous data-driven intelligent agent collaborative optimization system in accordance with claim 1, wherein, The intelligent agent development and management collaborative platform also realizes the management of the multi-task intelligent agent, including: The application lifecycle management is used to clarify the application target and architecture of the multi-task intelligent agent, which is then deployed to the production environment. After deployment, the application process of the multi-task intelligent agent is monitored in real time and abnormal situations are handled. At the same time, the multi-task intelligent agent is functionally updated and performance optimized according to user feedback and business requirements. When the intelligent agent cannot meet the business requirements or there is a better alternative, it will be retired and replaced. Personalized recommendation question setting is used to collect basic information, historical behavior data and preference data of users, predict questions of interest of users by multi-task agent and generate question recommendation list, dynamically adjust private domain knowledge database and set personalized recommendation question according to real-time feedback and behavior change of users; Session record management is used to store session content of users when using multi-task agent to private domain knowledge database, classify and index each session content, and analyze user demand and behavior mode based on recorded session content to update private domain knowledge database and realize optimization of multi-task agent.

14. An autonomously controllable heterogeneous data-driven agent coordination optimization method, characterized in that, The method adopts the autonomous controllable heterogeneous data driven agent collaborative optimization system of any one of claims 1-13, comprising the following steps: Deploying a reasoning engine for knowledge reasoning based on a large model and a functional component for data processing based on a small or medium-sized model; Building a private domain knowledge database containing knowledge information of fields of interest based on the functional component; Using the agent development and management collaborative platform to deploy a multi-task agent developed by the functional component and the reasoning engine, and iteratively performing intent recognition, query rewriting, dynamic planning, rough screening of retrieval and recall, and secondary precision sorting of rough screening results on the multi-task agent based on the private domain knowledge database to realize the autonomous controllable heterogeneous data driven collaborative optimization of the multi-task agent.

Images