Data analysis method and device based on multi-agent cooperation and storage medium

Abstract

The invention discloses a data analysis method and device based on multi-agent cooperation and a storage medium. The method comprises the following steps: receiving a natural language request by a workflow arrangement agent, and generating an execution plan comprising a plurality of subtasks and corresponding task processing units by consulting a data exploration agent and a data engine understanding agent; in the execution stage, the professional agent matched with the subtask converts the subtask into an executable instruction, and the data engine dispatch agent calls and coordinates the corresponding task processing unit to execute the instruction. According to the method, through a layered collaborative architecture, the separation of planning, translation and execution is realized, and the automation level, modularization and expandability of data analysis are remarkably improved.

Description

Technical Field

[0001] This application relates to the field of information retrieval technology, and in particular to a data analysis method, device and storage medium based on multi-agent collaboration. Background Technology

[0002] In recent years, significant progress has been made in the intersection of artificial intelligence (AI) technology and data management (Data+AI), with development mainly in two directions. On the one hand, AI is being used to empower traditional data management. For example, machine learning models are used to handle NP-hard problems such as index recommendation and automatic parameter tuning, as well as regression problems like query cost estimation. However, these methods heavily rely on expert tuning for specific scenarios and lack adaptability to dynamic changes in data, queries, and the environment, still facing considerable challenges. On the other hand, data management technologies are being used to support the deployment of AI applications. For example, database optimization techniques are used to support machine learning training, data cleaning, and model management. The core challenge of this approach lies in how to automate system workflows to reduce significant manual intervention.

[0003] The core issue hindering the adaptation of existing Data+AI technologies to diverse scenarios lies in their inherent limitations in semantic understanding, reasoning, and planning capabilities. Fortunately, the emergence of Large Language Models (LLMs) has brought new opportunities, with their powerful language capabilities promising to drive the transformation of Data+AI systems. However, directly utilizing LLMs to coordinate the entire Data+AI ecosystem to handle complex data analysis tasks still faces a series of significant technical challenges. These include how to enable models to effectively understand users' fuzzy queries, heterogeneous data, and the capabilities of various agents and tools within the system; how to automatically orchestrate efficient and reasonable workflows to bridge the gap between user needs and underlying data; and how to schedule and coordinate agents and tools with different functionalities to improve task performance.

[0004] Currently, there is a lack of system frameworks in this field capable of comprehensively addressing the aforementioned challenges and autonomously completing end-to-end data analysis tasks. Therefore, proposing a new technological solution that can overcome the shortcomings of existing technologies and effectively fill this gap has become an urgent problem to be solved in this field. Summary of the Invention

[0005] The purpose of this application is to provide a systematic solution that integrates semantic understanding, automatic planning, and multi-agent collaboration when using large language models for data analysis.

[0006] According to one embodiment of this application, a data analysis method based on multi-agent cooperation is proposed, comprising:

[0007] The workflow orchestration agent receives users' natural language data analysis requests;

[0008] The workflow orchestration agent calls the data exploration agent to obtain metadata information from the data source related to the request;

[0009] The workflow orchestration agent invokes the integrated large language model, and based on the predefined atomic data operation subtask types and the acquired metadata information, decomposes and orchestrates the request into a workflow framework containing subtasks with logical dependencies.

[0010] The workflow orchestration agent calls the data engine to understand the agent, determines the task processing unit corresponding to each subtask based on the capability profile of the task processing unit, and generates a data processing execution plan in conjunction with the workflow framework.

[0011] The workflow orchestration agent drives the execution of subtasks according to the execution plan. For each subtask to be executed, the data exploration agent or data engine understanding agent matched with each subtask converts the subtask into executable instructions supported by the corresponding task processing unit. The data engine scheduling agent calls and coordinates each task processing unit to execute the converted executable instructions and reports the execution status to the workflow orchestration agent.

[0012] The workflow orchestration agent receives the execution status of subtasks and drives the completion of all subtasks in the execution plan to obtain the final data analysis results.

[0013] In some implementations, when the workflow orchestration agent orchestrates the workflow framework, it provides the acquired metadata information as context to the large language model to assist in generating a workflow framework that matches the data source.

[0014] In some implementations, the data exploration agent obtains structured metadata information by accessing a unified data semantic catalog and uses data weaving techniques to provide a unified view of heterogeneous data.

[0015] In some implementations, the workflow orchestration agent invokes a data engine to understand the agent and determines the task processing unit corresponding to each subtask based on the capability profile of the task processing unit, including:

[0016] The workflow orchestration agent provides the characteristics of subtasks to the data engine understanding agent. The data engine understanding agent matches the characteristics of the subtasks with the capability profiles of various task processing units to determine the task processing unit corresponding to the subtask. The subtask characteristics include some or all of the subtask type, data size, data distribution, and data type.

[0017] In some implementations, the data engine understands the agent by accessing a unified data engine catalog to obtain capability profiles of various task processing units. These capability profiles are used to record the capability characteristics, applicable scenarios, and connection information of various task processing units.

[0018] In some embodiments, the method further includes:

[0019] The workflow orchestration agent dynamically optimizes the unexecuted portions of the execution plan based on actual runtime information in the execution state.

[0020] And / or, the workflow orchestration agent, based on indications of unsatisfactory execution status or user feedback, invokes a large language model to re-decompose and orchestrate the request to generate a new execution plan.

[0021] In some implementations, the workflow orchestration agent, data exploration agent, data engine understanding agent, and / or data engine scheduling agent achieve self-improvement through continuous learning, which includes at least one of self-reflection, reinforcement learning, or reward models.

[0022] In some implementations, for each subtask to be executed, a data exploration agent or data engine understanding agent matched with each subtask converts the subtask into executable instructions supported by the corresponding task processing unit, including:

[0023] If the subtask is an exploratory subtask, the data exploration agent matching the subtask will convert the subtask into executable instructions supported by the corresponding task processing unit;

[0024] If the subtask is a computation / verification subtask, the data engine understanding agent matching the subtask converts the subtask into executable instructions supported by the corresponding task processing unit.

[0025] In some embodiments, the method further includes:

[0026] A low-level perception component is provided for multiple agents to understand the environment, data, agents and tools. The perception component aligns the agents by offline fine-tuning or by preparing offline prompt templates.

[0027] In some embodiments, the method further includes:

[0028] Provides a low-level memory component for multi-agent systems to manage long-term and short-term memory, wherein the long-term memory includes domain knowledge and environmental knowledge, and the short-term memory includes user context and reflection context.

[0029] In some embodiments, the method further includes:

[0030] The system provides a low-level agent-to-agent component for multi-agent interaction, used to coordinate and collaborate multiple agents to complete decomposed sub-tasks. The agent-to-agent interaction component further includes:

[0031] The agent profiling and selection sub-component is used to build capability profiles for each agent and determine the agent that matches the sub-task.

[0032] The Agent Interaction and Collaboration Sub-component is used to communicate and synchronize states between agents through the Agent-to-Agent (A2A) protocol. Subtasks with direct dependencies pass information through the Agent Interaction and Collaboration Sub-component.

[0033] The agent execution sub-component is used to execute multiple agents in a serial, parallel, or pipelined manner.

[0034] In some embodiments, the method further includes:

[0035] A low-level agent-tool component is provided for multi-agent systems to select and invoke task processing units. The agent-tool component constructs a capability profile for each task processing unit and supports the invocation and interaction between the data engine scheduling agent and the task processing unit, as well as the integration of new task processing units, through the Model Context Protocol (MCP).

[0036] According to one embodiment of this application, an electronic device is provided, the device including a memory and a processor, the memory being used to store computer instructions executable on the processor, and the processor being used to implement the method as described in any of the preceding claims when executing the computer instructions.

[0037] According to one embodiment of this application, a computer-readable storage medium is provided that stores a computer program thereon, which, when executed by a processor, implements the method as described in any of the preceding claims.

[0038] This application proposes a multi-agent collaborative data analysis scheme. By constructing a collaborative framework consisting of a workflow orchestration agent, specialized agents (including a data exploration agent and a data engine understanding agent), and a data engine scheduling agent, it achieves a high degree of automation and intelligence in the data analysis process. Through a hierarchical collaborative planning process, the workflow orchestration agent consults with specialized agents from different domains to decompose and orchestrate requests and determine specific task processing units, forming a complete execution plan. This makes end-to-end automation of complex data analysis tasks possible, significantly reducing the need for and complexity of manual intervention. During the execution phase of this plan, each specialized agent is responsible for converting abstract subtasks into concrete executable instructions supported by backend task processing units. The dedicated data engine scheduling agent is responsible for coordinating the final invocation and execution of these instructions, achieving separation of task translation and execution, and significantly improving the modularity, scalability, and maintainability of the entire system.

[0039] The various technologies in this application further enhance the system's capabilities and effectiveness. By providing the acquired metadata as context to the large language model, the quality of the workflow framework is improved; through a unified data semantic catalog and data weaving technology, the system can effectively understand and integrate heterogeneous data sources, providing a comprehensive and consistent data view for subsequent analysis; by matching the capability profiles of task processing units based on the data scale and type of subtasks to determine specific task processing units, intelligent and refined selection of task processing units is achieved, improving the execution efficiency and resource utilization of the execution plan; by centrally storing the capability profiles of various task processing units in a unified data engine catalog, a centralized tool knowledge base is provided for the system, making the addition or updating of tools more convenient. The system's simplicity greatly enhances its scalability and maintainability; the introduction of failure feedback and replanning mechanisms enables dynamic self-correction, significantly improving its robustness in complex real-world environments; the introduction of continuous learning mechanisms (such as self-reflection and reinforcement learning) allows the system to continuously improve from experience, enhancing its intelligence; the introduction of underlying components such as perception and memory provides the agent with unified environmental understanding and knowledge management capabilities, enhancing decision-making accuracy; and the definition of standardized interaction protocols (such as A2A and MCP) ensures efficient communication and convenient expansion among components, improving the overall maintainability and scalability of the system.

[0040] Other features and advantages of the technical solution proposed in this application are described below. Attached Figure Description

[0041] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this specification and, together with the description, serve to explain the principles of this specification.

[0042] Figure 1 A flowchart of a data analysis method based on multi-agent collaboration according to an embodiment of this application is shown.

[0043] Figure 2 A schematic diagram of the overall architecture of a data analysis method according to an exemplary embodiment of this application is shown.

[0044] Figure 3 The diagram illustrates the workflow and component interactions of the data analysis method according to an exemplary embodiment of this application.

[0045] Figure 4 A schematic diagram of a workflow framework according to an exemplary embodiment of this application is shown.

[0046] Figure 5 This is a schematic diagram of the structure of an electronic device shown in at least one embodiment of this application. Detailed Implementation

[0047] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0048] In the context of this application, the term "agent" refers to a computational entity possessing a certain degree of autonomy. This entity is capable of reasoning and making decisions based on its internal knowledge, rules, or models by perceiving its environment (e.g., receiving user requests, data status, or system feedback), and autonomously executing a series of actions to achieve a preset goal. In multi-agent systems, the agents also possess the ability to communicate and collaborate with each other to jointly accomplish complex tasks that are difficult for a single agent to solve.

[0049] The purpose of this application is to provide an automated data analysis method based on large language models and multi-agent collaboration, to construct a collaborative framework with a workflow orchestration agent integrating a large language model (LLM) at its core for unified planning and scheduling. This method employs a layered collaborative process. First, the workflow orchestration agent invokes a data exploration agent to deeply understand and locate relevant data, decomposing and orchestrating the user's natural language request into a high-level workflow framework containing multiple logical sub-tasks. Subsequently, the workflow orchestration agent consults with the data engine understanding agent to select the optimal task processing unit, forming a complete and executable data processing plan. During the execution phase of this plan, various specialized agents are responsible for converting abstract sub-tasks into concrete executable instructions supported by backend task processing units, while a dedicated data engine scheduling agent coordinates the final invocation and execution of these instructions. Supported by underlying memory, perception, and interaction components, the entire method achieves end-to-end automation from understanding the user's ambiguous intent to obtaining precise analysis results.

[0050] To make the above-mentioned objectives, features, and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. It should be noted that, unless otherwise specified, the embodiments / implementations and features therein in this application can be combined with each other.

[0051] Reference Figure 1 This application provides a data analysis method based on multi-agent collaboration, which includes the following steps 101 to 106.

[0052] Step 101: The workflow orchestration agent receives the user's natural language data analysis request.

[0053] In this embodiment, this step is the starting point of the entire automated data analysis process. The workflow orchestration agent is the core and overall commander of the entire data analysis system. The workflow orchestration agent is an intelligent agent built on a large language model, possessing powerful semantic understanding, reasoning, planning, and self-reflection capabilities. Its core responsibility is to transform the high-level, fuzzy data analysis goals proposed by the user into a structured workflow consisting of a series of specific, machine-executable subtasks.

[0054] Natural language processing (NLP) data analysis requests can be queries entered by users in a conversational manner, potentially containing complex logic and specific business context. For example, a request like, "Among questions with over 500 views, which involves ball sports and has the highest ratio of injury-related questions to training-related questions?" might be problematic. The true intent of an NLP request may be implicit, and its direct mapping to the underlying heterogeneous data is unclear. Upon receiving such a request, the workflow orchestration agent will utilize its built-in large-scale language model capabilities to perform preliminary intent analysis to initiate subsequent planning processes. The aim is to bridge the gap between the user's advanced requirements and the underlying data.

[0055] Step 102: The workflow orchestration agent calls the data exploration agent to obtain metadata information of the data source related to the request.

[0056] According to this embodiment, upon receiving a user request, the workflow orchestration agent first understands what data is available in the current system to answer the question. Therefore, the workflow orchestration agent invokes a dedicated data exploration agent responsible for understanding and discovering data. In this application, the data exploration agent refers to a specialized agent used to organize and understand data so that other agents can better discover and access it. In some implementations, the data exploration agent can be implemented using a traditional rule engine; in other implementations, it can be built based on an LLM (Local Level Management Model). Data exploration agents built on an LLM, especially those fine-tuned with domain knowledge, are better able to understand more ambiguous semantic requests, such as "find the table related to 'customer churn'," thereby significantly improving system capabilities.

[0057] Step 102 is the targeted information acquisition process. Specifically, the workflow orchestration agent will pass the keywords or semantic concepts (such as "problem," "views," "injury," etc.) parsed from the user request to the data exploration agent. The data exploration agent will then use this information to query and retrieve the knowledge base it manages, identify all potentially relevant available data sources, and return their metadata information to the workflow orchestration agent.

[0058] Metadata information refers to structured information used to describe and define data assets; in other words, data about data. Metadata information can include technical metadata describing the physical and technical attributes of the data, such as table names, column names, and data types. It can also include business metadata describing the data from a business perspective, such as the business definition of the data. Furthermore, metadata information can also include semantic metadata that reveals the deeper meaning and relationships of the data, such as business tags attached to the data and the lineage relationships between data points. In multi-agent systems, this metadata information collectively constitutes the basic knowledge base that enables agents to understand, discover, and plan how to manipulate data.

[0059] In a preferred embodiment, the data exploration agent can obtain structured metadata information by accessing a unified data semantic catalog and provide a unified view of heterogeneous data through data weaving technology. The data semantic catalog is a knowledge base that provides a well-structured metadata system. It can contain technical information about the data (such as table names and column names) as well as semantic information such as business definitions and data tags. Data weaving is a technique that provides a unified access view for heterogeneous data from diverse sources and in different formats. Through the combination of these two techniques, the data exploration agent can effectively provide the workflow orchestration agent with a comprehensive, consistent, and highly relevant list of available data and its detailed descriptions from a complex data environment.

[0060] To address the challenges of inconsistent data quality and diverse data sources in the real world, the data understanding and exploration agent in this application integrates key data preprocessing capabilities in addition to data discovery. The components comprising the data understanding and exploration agent may include multiple tools for data preparation, cleaning, and integration. Data preparation tools can perform preparatory operations such as format conversion, sampling, or structural adjustment of data before formal analysis. Data cleaning can be specifically used to identify and correct errors, inconsistencies, or missing values ​​in the data to improve data quality. Data integration tools can merge and unify interrelated data from different data sources to form a more comprehensive dataset for subsequent analysis.

[0061] When developing an execution plan, the workflow orchestration agent can leverage this capability of the data exploration agent. For example, if quality issues are discovered in the required data source during the exploration phase, or if merging with another data source is necessary, the workflow orchestration agent can automatically and proactively plan one or more data cleaning or integration subtasks utilizing these tools within its generated workflow framework. By developing efficient strategies to fully leverage the capabilities of these tools, the system ensures that the data input to subsequent computational / validation subtasks is of high quality and well-organized, thereby significantly improving the accuracy and robustness of the entire data analysis process.

[0062] To make subsequent planning more intelligent and accurate, in some implementations, when the workflow orchestration agent orchestrates the workflow framework, it provides the acquired metadata information as context to the large language model to assist in generating a workflow framework that matches the data source. According to this implementation, by inputting accurate metadata information relevant to currently available data as context into the large language model, it ensures that the generated plan (i.e., the workflow framework) is closely integrated with real-world data, thereby significantly improving the accuracy, feasibility, and final execution success rate of the subsequently generated plan.

[0063] Step 103: The workflow orchestration agent calls the large language model, and based on the predefined atomic data operation subtask types and the acquired metadata information, decomposes and orchestrates the request into a workflow framework containing one or more subtasks with logical dependencies.

[0064] After acquiring metadata information about available data sources related to the user request, the workflow orchestration agent enters the workflow planning phase to generate a structured workflow framework.

[0065] The workflow orchestration agent can invoke its built-in large language model. Based on the metadata information obtained in step 102 and provided as context, as well as the predefined atomic data operation subtask types in the system, the large language model can perform deep reasoning and decomposition of the user's natural language request. Atomic data operation subtask types refer to the smallest logically indivisible operation units predefined in the system, such as Scan, Filter, Extract, GroupBy, Count, and Compute, etc.

[0066] Large language models, through their powerful understanding and reasoning capabilities, transform complex, unstructured user questions into a sequence of subtasks or a graph structure composed of atomic operations with logical dependencies. Logical dependencies refer to the execution order and data flow between subtasks. For example, a Filter subtask can only execute after a Scan subtask is completed because the Filter subtask needs the output of the Scan as its input. These subtasks can be executed sequentially or in parallel. Figure 4 It demonstrates an execution flow that includes sequential, parallel, and converging paths, with a directed acyclic graph (DAG) structure that defines each atomic step required to complete the entire data analysis task and their interrelationships.

[0067] Step 104: The workflow orchestration agent calls the data engine to understand the agent, determines the task processing unit corresponding to each subtask based on the capability profile of the task processing unit, and generates a data processing execution plan in conjunction with the workflow framework.

[0068] After generating the workflow framework, the workflow orchestration agent further determines the execution unit corresponding to each decomposed subtask. This embodiment uses collaborative technology selection to choose the most suitable task processing unit for each subtask from a large number of available task processing units to form a data processing execution plan.

[0069] The task processing unit includes all backend tools in the system. For example, it may include data processing engines for computation (such as Spark, DBMS, Pandas, PyData, etc.) and data discovery engines for data discovery (such as data catalogs, vector databases, etc.).

[0070] To make the optimal choice, according to this embodiment, the workflow orchestration agent can invoke a dedicated data engine understanding agent responsible for understanding the capabilities of the tools. In some implementations, the workflow orchestration agent provides the characteristics of subtasks to the data engine understanding agent. The data engine understanding agent matches the characteristics of the subtasks with the capability profiles of various task processing units to determine the corresponding task processing unit for the subtask. The subtask characteristics may include some or all of the following: subtask type, data size, data distribution, and data type. The subtask type typically determines the operation to be performed, such as filtering, aggregation (GroupBy), or joining. Data size can also be an important consideration in selecting a task processing unit. For example, small-scale data can use a lightweight in-memory computing engine (such as Pandas), while large-scale data requires a distributed computing engine (such as Spark). Data distribution can affect execution efficiency. For example, for an aggregation (GroupBy) subtask, if the data distribution is highly skewed, the data engine understanding agent may choose a distributed computing engine with more advanced skew handling mechanisms (such as skew join optimization in adaptive query execution) instead of a simple hash aggregation engine. Data types can be directly associated with specific optimization engines. For example, if a subtask involves searching for similarity in vectorized text, the data engine understanding agent can prioritize a vector database specifically optimized for this purpose as the task processing unit, rather than a traditional relational database. The workflow orchestration agent can send these detailed subtask features as consultation requests to the data engine understanding agent to obtain optimal technology selection recommendations.

[0071] In some implementations, upon receiving a query, the data engine understanding agent can access a unified data engine catalog to obtain capability profiles of various task processing units. These capability profiles record the capability characteristics, applicable scenarios, and connectivity information of each task processing unit. The data engine understanding agent then matches the input subtask characteristics with the capability profiles of each task processing unit within the data engine catalog, thereby recommending the most suitable task processing unit for that subtask. For example, it might recommend using the Spark engine for a large-scale GroupBy task.

[0072] In some implementations, the data engine understanding agent can be implemented using a traditional rule engine. In other implementations, it can be built based on an LLM (Local Level Management Model), particularly an LLM fine-tuned with domain knowledge. An LLM-based data engine understanding agent is better able to understand analytical intent and match optimal task processing units.

[0073] Then, the workflow orchestration agent can fill the task processing units obtained from the selection suggestions returned by the data engine into the corresponding subtask nodes of the workflow framework, thereby forming a complete data processing execution plan that includes each step and the execution tools required to execute that step.

[0074] Step 105: The workflow orchestration agent drives the execution of subtasks according to the final data processing execution plan. For the subtask to be executed, the execution process includes: the specialized agent (data exploration agent or data engine understanding agent) assigned to the subtask converts the subtask into executable instructions supported by its corresponding task processing unit; the data engine scheduling agent calls and coordinates the task processing units corresponding to each subtask to be executed to execute the converted executable instructions.

[0075] After generating a complete data processing execution plan, the system can enter the actual execution phase. The workflow orchestration agent acts as the driver, sequentially launching each subtask according to the subtask dependencies defined in the plan. Subtasks that have no dependencies on each other can be executed simultaneously.

[0076] For each subtask to be executed, the execution process can be divided into two parts: translation and invocation. Subtasks in the execution plan are high-level, abstract atomic operations. To enable the backend task processing unit to better understand and execute the subtask, it can first be converted into specific executable instructions supported by that task processing unit. This is then completed by a specialized intelligent agent matched with the task processing unit corresponding to the subtask. In some implementations, if the subtask is an exploratory subtask, the data exploration intelligent agent matched with the subtask converts it into executable instructions supported by the corresponding task processing unit; if the subtask is a computation / verification subtask, the data engine understanding intelligent agent matched with the subtask converts it into executable instructions supported by the corresponding task processing unit. Exploratory subtasks typically refer to atomic operations focused on data discovery and understanding, aiming to provide a basis for subsequent planning, such as finding relevant data sources or obtaining their metadata. Computation / verification subtasks typically refer to atomic operations that specifically process, transform, and compute data, aiming to transform data into final results, such as filtering, aggregating, or joining data.

[0077] There is a pre-defined correspondence between specialized intelligent agents (data exploration agents or data engine understanding agents) acting as translators and the task processing units they support. This correspondence can be one-to-one, where a highly specialized agent is responsible for generating instructions only for a specific task processing unit (e.g., only for the Spark engine); it can be one-to-many, where a more generalized agent can generate instructions for a class of task processing units with similar interaction methods (e.g., all relational databases that support standard SQL); or it can be many-to-one, for example, for the same task processing unit (such as a Spark cluster), the most suitable agent is selected to generate the optimal instructions based on the characteristics of the subtasks or the optimization objectives. When a new task processing unit needs to be integrated into the system, specialized agents capable of generating instructions that the task processing unit can support can be provided or configured without modifying the system's upper-level planning and scheduling logic.

[0078] The data engine scheduler receives executable instructions translated by a specialized agent, and then is responsible for communicating and interacting with the corresponding backend task processing unit, calling and coordinating its execution of the instructions.

[0079] The entire execution process is supported by multiple underlying components. In some implementations, coordination and cooperation among multiple agents are supported by an agent-agent interaction component. This component includes an agent profiling and selection subcomponent for building capability profiles for agents and making selections; an agent interaction and coordination subcomponent for facilitating communication and state synchronization through an agent-to-agent (A2A) protocol; and an agent execution subcomponent for efficiently executing multiple agents in a serial, parallel, or pipelined manner. When there are direct dependencies between subtasks (e.g., the output of Scan is the input of Filter), the data results between them can be directly passed through the agent interaction and coordination subcomponent, thereby improving the efficiency of data flow.

[0080] In some implementations, the calls and interactions between the data engine scheduling agent and various task processing units are supported by an agent-tool invocation component. This component enables the convenient integration of new tools by building capability profiles for various task processing units and standardizing interactions through a Model Context Protocol (MCP).

[0081] Step 106: The workflow orchestration agent receives the execution status of the subtasks and drives the completion of all subtasks in the execution plan to obtain the final data analysis results.

[0082] The data engine scheduling agent can report the execution status to the workflow orchestration agent. The execution status can include metadata such as success, failure, execution time, and output data size.

[0083] The workflow orchestration agent can determine and start the next or next batch of subtasks to be executed based on the received execution status and the logical dependencies in the execution plan, until the entire execution plan is completed, so as to obtain the final data analysis results of the request.

[0084] In some implementations, the workflow orchestration agent can dynamically optimize the unexecuted portions of the execution plan based on actual runtime information in the execution status. Specifically, after receiving the subtask execution status, the workflow orchestration agent can analyze the actual runtime information contained therein (e.g., output data scale) and use this information to determine whether the unexecuted portions of the original plan are still optimal. If not, subsequent plans can be dynamically optimized to pursue higher execution efficiency (e.g., improving latency, reducing costs) or accuracy.

[0085] For example, suppose the workflow orchestration agent initially plans a sequence of Scan and Filter subtasks. During planning, based on metadata estimates, the agent anticipates the Scan task will output a medium-sized dataset of approximately 100MB. Therefore, during the technology selection phase, the data engine understanding agent chooses the lightweight Pandas in-memory computing engine as the task processing unit for the subsequent Filter subtask. However, after the Scan task is actually completed, the execution status reported by the data engine scheduling agent shows that the actual output data volume is as high as 10TB. Based on this actual runtime information, the workflow orchestration agent determines that the original plan to use Pandas to process 10TB of data is infeasible or extremely inefficient. Therefore, the workflow orchestration agent can consult the data engine understanding agent again based on the updated subtask characteristics (task type = Filter, data size = 10TB). The data engine understanding agent, based on the new information, can recommend a more suitable task processing unit for handling large-scale data, such as the Spark engine. The workflow orchestration agent can update its execution plan based on this suggestion and distribute the optimized Filter subtasks for execution.

[0086] In some implementations, the workflow orchestration agent invokes a large language model based on an indication of an unsatisfactory execution status or user feedback to re-decompose and orchestrate the request to generate a new execution plan. If the workflow orchestration agent receives an unsatisfactory execution status (e.g., an indication of execution failure) or user feedback, it can invoke a large language model to re-decompose and orchestrate the request to generate a new execution plan, further improving the robustness and intelligence of the system.

[0087] Once the workflow orchestration agent confirms that all subtasks in its execution plan have been successfully completed, it can integrate the outputs of the last subtask to form the final answer to the user's initial request for natural language data analysis and provide it to the user.

[0088] In some implementations, to support the complex operation of the entire system, this embodiment can provide multiple underlying components to support multiple agents. A low-level perception component can be provided for multiple agents to understand the environment, data, agents, and tools, wherein the perception component aligns the agents through offline fine-tuning or preparing offline prompt templates. A low-level memory component can be provided for multiple agents to manage long-term and short-term memory, wherein the long-term memory includes domain knowledge and environmental knowledge, and the short-term memory includes user context and reflective context. Furthermore, as described above, this embodiment can also provide agent-agent interaction components for coordinating and collaborating among multiple agents, and agent-tool invocation for interaction between agents and various task processing units.

[0089] In some embodiments of this application, one or more agents in the system may also achieve self-improvement through continuous learning, which includes at least one of self-reflection, reinforcement learning, or reward models.

[0090] The following will combine Figure 2 This paper introduces a data analysis method proposed according to an exemplary embodiment of the present application from the perspective of overall system architecture. The architecture is centered on a large language model and operates collaboratively through six key functional modules to achieve intelligent and automated processing of data analysis tasks.

[0091] First, the perception component is the input and understanding hub of the system. The perception component has a comprehensive understanding of the external environment and internal state, which can include data / query understanding of complex queries and heterogeneous data input by users, environmental understanding of the system operating environment, available resources and constraints, and cognition of the current task optimization goals (such as low latency and low cost).

[0092] The reasoning and planning component acts as the "brain" of the system, formulating action plans based on information acquired by the perception module. It decomposes natural language data analysis requests into task plans consisting of multiple dependent sub-tasks, optimizes these plans, and drives task execution.

[0093] During planning and execution, the system interacts with external tools through a tool invocation component. This component has a deep understanding of available tools and can select tools based on task requirements. During execution, this component is responsible for handling interactions with tools and ensures the reliability and scalability of these interactions through standardized tool interfaces (such as the MCP protocol).

[0094] The memory component provides the agent with a knowledge base to store fixed domain knowledge, contextual knowledge related to the current user's dialogue, environmental knowledge of the system's operation, and reflective knowledge summarized from past experiences, providing a basis for the agent's decision-making.

[0095] Continuous learning components endow the system with the ability to improve itself, enabling it to continuously evolve. Through reflection mechanisms, continuous learning components analyze the successes and failures of task execution; and by utilizing techniques such as reinforcement learning and reward models, they continuously optimize their strategies and models based on execution feedback.

[0096] Finally, the entire architecture is organized as a multi-agent system to overcome the limitations of a single agent in handling complex real-world tasks. This component is responsible for managing the capabilities of agents (agent profiles), coordinating interactions between agents, agents' invocation of external tools, and communication between agents and the network, enabling multi-agent collaboration to achieve complex analytical goals.

[0097] The following will refer to Figure 3 The specific workflow and component interaction relationships of a data analysis method according to an exemplary embodiment of this application are described.

[0098] The entire system consists of three core intelligent agents (labeled ①, ②, ③) working together, and is supported by four underlying components (labeled (1), (2), (3), (4)).

[0099] The system operates on an underlying data lake, upon which various tools that directly interact with data are deployed, such as data weaving tools, data access tools, data processing tools, and data visualization tools. The system employs two types of specialized intelligent agents to intelligently manage and utilize these resources. The first agent on the left, designated "Data Understanding and Exploration," acts as the system's data expert. It understands the business meaning and technical attributes of data by accessing the data semantic catalog and invokes data access and weaving tools to address the questions of what data is available and how to access it. The second agent on the right, designated "Data Engine Understanding and Scheduling," (including the data engine understanding agent and data engine scheduling agent mentioned above), acts as the system's tool expert. It manages the capability profiles of various task processing units by referring to the data engine catalog and invokes data processing and visualization tools to address the questions of which tools to use and how to process data.

[0100] Above and interacting bidirectionally with these two types of specialized intelligent agents is the third type, the workflow orchestration intelligent agent. The workflow orchestration intelligent agent is the planner in the system, responsible for receiving user requests and integrating all information about data and tools by consulting with data experts (first type of intelligent agent) and tool experts (second type of intelligent agent) respectively, ultimately generating and driving a complete, end-to-end data processing execution plan.

[0101] Four underlying components provide the basic capabilities to support the operation of the entire system. The memory component (1) provides long-term and short-term knowledge storage for all agents. The perception component (2) enables the system to fully understand the environment and tasks. The agent-agent interaction component (3) coordinates the collaboration between the three core agents, and further includes an agent profiling and selection sub-component (3-1) for agent capability assessment and selection, an agent interaction and coordination sub-component (3-2) for ensuring communication between agents, and an agent execution sub-component (3-3) responsible for scheduling agent execution. The agent-tool invocation component (4) provides all agents with a standardized interface for interacting with external tools and task processing units.

[0102] To illustrate the workflow of this invention more specifically, the following will be combined with... Figure 4 This provides an end-to-end example of handling complex data analysis tasks.

[0103] In this example, the workflow orchestration agent receives a user's natural language data analysis request: "Among questions with more than 500 views, which one involves ball sports and has the highest ratio of injury-related questions to training-related questions?"

[0104] In this example, multiple data exploration agents provide different ways to organize and access multimodal data in the system. For instance, for tabular and text data that may be involved in a request, a multimodal data agent can be built that divides the tabular data by row and the text data by paragraph, converts them into vectors using a unified embedding model, and then constructs a unified vector index by clustering and chunking the vectors. Simultaneously, a structured agent can be built that can directly utilize the patterns in the tabular data to select the most appropriate tables and columns for access.

[0105] For insights into intelligent agents, refer to the Data Engine Catalog. Based on the data scale and task type of the current task (e.g., involving semantic similarity calculation or structured aggregation), select the most suitable task processing unit, such as a vector database or a relational database, to provide efficient storage and computational support for subsequent data analysis tasks.

[0106] Workflow orchestration agents can decompose and orchestrate the complex analytical problems described above into a specific execution plan consisting of multiple atomic data operation subtasks. An example of such a plan is as follows: Figure 4 As shown. (Refer to...) Figure 4 The execution plan is a directed acyclic graph (DAG), starting with the Scan operation at the bottom and proceeding sequentially through Filter, Extract, and GroupBy operations. Then, to calculate the number of "injury-related issues" and "training-related issues" separately, the plan branches into two parallel Filter->Count branches. Finally, the results from the two parallel branches are merged into the Compute node to calculate their proportions, and the final answer is found through the Max node.

[0107] During the execution of this plan, the agent profiling and selection subcomponent selects the most suitable specialized agent for each subtask (such as Scan, Filter, etc.). For example, when processing the Scan subtask, the agent profiling and selection subcomponent can choose the most suitable agent from multimodal data agents and structured data agents based on data characteristics to transform the subtask. The selected specialized agent is responsible for translating the Scan subtask into executable instructions supported by its backend task processing unit. These instructions are then handed over to a dedicated data engine scheduling agent for final execution. The agent interaction and collaboration subcomponent is responsible for passing data results between dependent subtasks (e.g., passing Scan output data to Filter as input) until the entire plan is completed and the final result is output.

[0108] During execution, the workflow orchestration agent can also optimize the plan based on the feedback execution status, such as adjusting the conditions and order of filter operations.

[0109] In summary, this application provides an advanced data analysis scheme based on a large-scale language model and multi-agent collaboration. This scheme centers on a workflow orchestration agent as the overall commander, achieving end-to-end automation through a layered collaborative process. The workflow orchestration agent is responsible for understanding the user's natural language requests and, by consulting a dedicated data exploration agent (which utilizes a data semantic catalog), locates and understands relevant data, thereby constructing a high-level workflow framework. The workflow orchestration agent further consults a data engine understanding agent (which references capability profiles in the data engine catalog) to select the optimal task processing unit for each subtask in the workflow, forming the final execution plan.

[0110] During the execution phase, this solution separates task translation from task execution. Selected specialized agents (data exploration or data engine understanding agents) are responsible for translating abstract subtasks into specific instructions that receive optimal support from the corresponding task processing units. The data engine scheduling agent, acting as the unified executor, is responsible for invoking task processing units to execute these instructions. The entire system also possesses advanced capabilities such as dynamic optimization and replanning after failure, supported by underlying perception, memory, and interaction components, forming a professional, highly modular, scalable, and intelligent data analysis solution.

[0111] Figure 5 An electronic device provided in at least one embodiment of this application includes a memory and a processor. The memory is used to store computer instructions that can be executed on the processor. The processor is used to implement the data analysis method based on multi-agent cooperation as described in any embodiment or implementation of this application when executing the computer instructions.

[0112] At least one embodiment of this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the data analysis method based on multi-agent cooperation described in any embodiment or implementation of this application.

[0113] Those skilled in the art will understand that one or more embodiments of this specification can be provided as a method, system, or computer program product. Therefore, one or more embodiments of this specification may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, one or more embodiments of this specification may take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0114] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the data processing device embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0115] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired results. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0116] The embodiments of the subject matter and functional operation described in this specification can be implemented in the following ways: digital electronic circuits, tangibly embodied computer software or firmware, computer hardware including the structures disclosed in this specification and their structural equivalents, or combinations thereof. Embodiments of the subject matter described in this specification can be implemented as one or more computer programs, i.e., one or more modules of computer program instructions encoded on a tangible, non-transitory program carrier for execution by a data processing apparatus or for controlling the operation of a data processing apparatus. Alternatively or additionally, the program instructions may be encoded on artificially generated propagation signals, such as machine-generated electrical, optical, or electromagnetic signals, which are generated to encode information and transmit it to a suitable receiving device for execution by the data processing apparatus. The computer storage medium may be a machine-readable storage device, a machine-readable storage substrate, a random or serial access memory device, or combinations thereof.

[0117] The processing and logic flow described in this specification can be executed by one or more programmable computers that execute one or more computer programs to perform corresponding functions by operating on input data and generating output. The processing and logic flow can also be executed by dedicated logic circuitry—such as FPGAs (Field-Programmable Gate Arrays) or ASICs (Application-Specific Integrated Circuits), and the device can also be implemented as dedicated logic circuitry.

[0118] Suitable computers for executing computer programs include, for example, general-purpose and / or special-purpose microprocessors, or any other type of central processing unit. Typically, the central processing unit receives instructions and data from read-only memory and / or random access memory. The basic components of a computer include a central processing unit for implementing or executing instructions and one or more memory devices for storing instructions and data. Typically, a computer will also include one or more mass storage devices for storing data, such as disks, magneto-optical disks, or optical disks, or the computer will be operatively coupled to such mass storage devices to receive data from or transfer data to them, or both. However, a computer is not required to have such devices. Furthermore, a computer can be embedded in another device, such as a mobile phone, a personal digital assistant (PDA), a mobile audio or video player, a game console, a global positioning system (GPS) receiver, or a portable storage device such as a universal serial bus (USB) flash drive, to name a few.

[0119] Computer-readable media suitable for storing computer program instructions and data include all forms of non-volatile memory, media, and memory devices, such as semiconductor memory devices (e.g., EPROM, EEPROM, and flash memory devices), magnetic disks (e.g., internal hard disks or removable disks), magneto-optical disks, and CD-ROM and DVD-ROM disks. Processors and memory may be supplemented by or incorporated into dedicated logic circuitry.

[0120] While this specification contains numerous specific implementation details, these should not be construed as limiting the scope of any invention or the scope of the claims, but rather are primarily intended to describe features of specific embodiments of a particular invention. Certain features described in the various embodiments herein may also be implemented in combination in a single embodiment. Conversely, various features described in a single embodiment may also be implemented separately in various embodiments or in any suitable sub-combination. Furthermore, while features may function in certain combinations as described above and even initially claimed in this way, one or more features from a claimed combination may be removed from that combination in some cases, and a claimed combination may refer to a sub-combination or a variation thereof.

[0121] Similarly, although the operations are depicted in a specific order in the accompanying drawings, this should not be construed as requiring these operations to be performed in the specific order shown or sequentially, or requiring all illustrated operations to be performed to achieve the desired result. In some cases, multitasking and parallel processing may be advantageous. Furthermore, the separation of various system modules and components in the above embodiments should not be construed as requiring such separation in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

[0122] Thus, specific embodiments of the subject matter have been described. Other embodiments are within the scope of the appended claims. In some cases, the actions recited in the claims may be performed in a different order and still achieve the desired result. Furthermore, the processes depicted in the drawings are not necessarily shown in a specific order or sequence to achieve the desired result. In some implementations, multitasking and parallel processing may be advantageous.

[0123] The above description is merely a preferred embodiment of one or more embodiments of this specification and is not intended to limit the scope of one or more embodiments of this specification. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments of this specification should be included within the protection scope of one or more embodiments of this specification.

Claims

1. A data analysis method based on multi-agent cooperation, characterized in that, The method comprises: a workflow orchestration agent receives a user's natural language data analysis request; the workflow orchestration agent calls a data exploration agent to obtain metadata information of a data source related to the request; the workflow orchestration agent calls an integrated large language model, based on pre-defined atomic data operation subtask types and the obtained metadata information, to decompose and orchestrate the request into a workflow framework comprising subtasks with logical dependencies; the workflow orchestration agent calls a data engine understanding agent to determine a task processing unit corresponding to each subtask based on the capability profile of the task processing unit, and generates a data processing execution plan in combination with the workflow framework; the workflow orchestration agent drives the execution of the subtasks according to the execution plan, wherein, for each subtask to be executed, the data exploration agent or the data engine understanding agent matched with the subtask converts the subtask into executable instructions supported by the corresponding task processing unit, and the data engine scheduling agent calls and coordinates each task processing unit to execute the converted executable instructions, and reports the execution status to the workflow orchestration agent; the workflow orchestration agent receives the execution status of the subtasks, and drives the completion of all subtasks in the execution plan to obtain the final data analysis result.

2. The method of claim 1, wherein, When the workflow orchestration agent orchestrates the workflow framework, the obtained metadata information is provided as context to the large language model to assist in generating a workflow framework matched with the data source.

3. The method of claim 2, wherein, The data exploration agent obtains structured metadata information by accessing a unified data semantic directory, and provides a unified view for heterogeneous data using data weaving technology.

4. The method of claim 1, wherein, The workflow orchestration agent calls the data engine understanding agent to determine a task processing unit corresponding to each subtask based on the capability profile of the task processing unit, comprising: The workflow orchestration agent provides the characteristics of the subtasks to the data engine understanding agent, and the data engine understanding agent matches the characteristics of the subtasks with the capability profiles of various task processing units to determine the task processing unit corresponding to the subtasks, wherein the subtask characteristics include part or all of the subtask type, data size, data distribution, and data type.

5. The method of claim 4, wherein, The data engine understanding agent obtains the capability profiles of various task processing units by accessing a unified data engine directory, and the capability profiles are used to record the capability characteristics, applicable scenarios, and connection information of various task processing units.

6. The method of claim 1, wherein, The method further comprises: The workflow orchestration agent calls dynamic optimization for the part of the execution plan that has not been executed based on the actual runtime information in the execution status; and / or, the workflow orchestration agent calls the large language model to re-decompose and orchestrate the request to generate a new execution plan based on the execution status indicating that the expectation is not met or user feedback.

7. The method of claim 1, wherein, The workflow orchestration agent, the data exploration agent, the data engine understanding agent, and / or the data engine scheduling agent improve themselves through continuous learning, and the continuous learning comprises at least one of self-reflection, reinforcement learning, or a reward model.

8. The method of claim 1, wherein, For each subtask to be executed, a data exploration agent or a data engine understanding agent matched with each subtask converts the subtask into executable instructions supported by the corresponding task processing unit, including: If the subtask is an exploration subtask, a data exploration agent matched with the subtask converts the subtask into executable instructions supported by the corresponding task processing unit; If the subtask is a calculation / verification subtask, a data engine understanding agent matched with the subtask converts the subtask into executable instructions supported by the corresponding task processing unit.

9. The method of claim 1, wherein, The method further includes: providing a bottom perception component for the plurality of agents for understanding the environment, data, agents and tools, wherein the perception component aligns each agent through offline fine-tuning or preparing offline prompt templates.

10. The method of claim 1, wherein, The method further includes: providing a bottom memory component for the plurality of agents for managing long-term memory and short-term memory, wherein the long-term memory includes domain knowledge and environmental knowledge, and the short-term memory includes user context and reflection context.

11. The method of claim 1, wherein, The method further includes: providing a bottom agent-agent component for the plurality of agents for coordinating and cooperating multiple agents to complete decomposed subtasks, the agent-agent interaction component further includes: an agent portrait and selection subcomponent for constructing a capability portrait for each agent and determining the agent matched with the subtask; an agent interaction and collaboration subcomponent for communicating and state synchronization between agents through an agent-to-agent interaction protocol (A2A), wherein subtasks with direct dependency relationship transfer information through the agent interaction and collaboration subcomponent; an agent execution subcomponent for executing multiple agents in a serial, parallel or pipeline manner.

12. The method of claim 1, wherein, The method further includes: providing a bottom agent-tool component for the plurality of agents for selecting and invoking a task processing unit, wherein the agent-tool component constructs a capability portrait for each task processing unit and supports the integration of a new task processing unit through a model context protocol (MCP) to support the invocation and interaction between the data engine scheduling agent and the task processing unit.

13. An electronic device, comprising: The device includes a memory for storing computer instructions executable on a processor, and a processor for implementing the method of any one of claims 1 to 12 when executing the computer instructions.

14. A computer readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the method of any one of claims 1 to 12. The program is executed by the processor to implement the method of any one of claims 1 to 12.

Images