Community life service agent system

Abstract

The invention belongs to the field of smart communities and artificial intelligence, and particularly relates to a community life service agent system, which comprises a data acquisition module used for acquiring multi-modal original data in a community, and an edge decision module comprising a data fusion unit, a workflow engine unit, an RAG enhancement unit and a code generation unit, the intelligent workflow generating module is used for performing fusion analysis and matching on the sensing data or dynamically generating an intelligent workflow and generating a decision instruction; and the terminal execution module is used for receiving and executing the decision instruction and feeding back a result. According to the method, the existing sensing equipment of the community is reused, and a'sensing-decision-execution 'closed loop is constructed on the edge side, so that active and intelligent services of community management are realized, and the deployment cost and the data security risk are remarkably reduced at the same time.

Description

A community life service intelligent system Technical Field

[0001] This invention belongs to the field of smart community and artificial intelligence technology, specifically relating to a smart system for community life services. Background Technology

[0002] With the acceleration of urbanization, security-related equipment such as access control systems, video surveillance systems, and fire sensors have been widely deployed in various residential communities to meet residents' basic safety needs. This has formed a basic sensing network covering scenarios such as personnel entry and exit, public area monitoring, and security early warning, providing crucial support for community security management. However, current smart community solutions on the market still have several significant shortcomings, hindering their evolution towards a higher and more universally accessible intelligent stage. Furthermore, most existing systems suffer from single-function architectures and severe data silos. Various sensing devices within the community (such as surveillance cameras, access control systems, and sensors) typically only complete data collection and simple responses within their limited scope of responsibility, such as video recording and identity verification, lacking the ability to integrate and analyze multi-source data and make collaborative decisions. Property management systems also primarily focus on basic functions such as recording, repair requests, and payment, failing to deeply integrate security data and equipment status with property service processes, thus hindering the shift from "passive response" to "proactive service."

[0003] Secondly, the cost of achieving deep intelligence remains high. Although embodied intelligence systems with "perception-decision-execution" capabilities are theoretically achievable, their implementation typically relies on deploying specialized hardware such as intelligent robots and high-end sensors, as well as complex cloud-based AI analysis platforms. These solutions require significant initial investment and demanding maintenance requirements, making them suitable only for high-end residential or commercial projects and difficult to scale up in ordinary residential communities. Furthermore, because intelligent systems depend on cloud services for data processing and model iteration, system downtime or data breaches can occur if the vendor's servers malfunction or are attacked.

[0004] In addition, existing systems mostly remain at the perception layer, lacking data-based intelligent decision-making mechanisms, and are unable to automatically convert decision results into execution instructions for property staff or smart devices, resulting in low management efficiency and redundant investment of human resources.

[0005] Therefore, how to build a community service intelligent system with autonomous decision-making and execution capabilities by carrying out lightweight intelligent transformation of existing systems without replacing existing sensing devices or significantly increasing costs has become a technical problem that the industry urgently needs to solve. Summary of the Invention

[0006] To address the aforementioned problems in existing technologies, this invention provides a community life service intelligent agent system. The system includes a data acquisition module, an edge decision-making module, and a terminal execution module connected via communication. The data acquisition module includes multiple heterogeneous sensing devices deployed within the community for real-time acquisition of multimodal raw data of the community environment. The edge decision-making module, privately deployed within the community and communicatively connected to the data acquisition module, receives the multimodal raw data and performs calculations. The decision-making module includes a data fusion unit, a workflow engine unit, an RGA enhancement unit, and a code generation unit. The data fusion unit preprocesses the multimodal data and performs feature extraction and spatiotemporal fusion to generate structured feature vectors. The workflow engine unit… The RGA enhancement unit, in collaboration with the data fusion unit, RGA enhancement unit, and code generation unit, is used to match, execute, and optimize the corresponding workflow based on the structured feature vector, and to generate decision instructions to be sent to the terminal execution module based on the workflow. The RGA enhancement unit retrieves domain knowledge from a multi-level knowledge base based on query requests from the workflow engine unit during the decision-making process, and returns the retrieval results to the workflow engine unit to assist in decision optimization. The code generation unit parses and converts the natural language instructions transmitted by the terminal execution module into executable workflow code for dynamic loading and execution by the workflow engine unit. The terminal execution module receives and executes the decision instructions from the edge decision module and feeds back the execution results to the edge decision module.

[0007] Preferably, the terminal execution module has a built-in AI voice interaction unit, which is used by the user to interact with the workflow engine and code generation unit by voice.

[0008] Preferably, the data fusion unit includes: a time synchronization subunit, used to timestamp align the multimodal raw data from different sensing devices based on network time protocol and adaptive filtering algorithm; a multimodal feature extraction subunit, which has built-in trained YOLOv5 model, LSTM anomaly detection model and graph neural network model, wherein the YOLOv5 model is used for video analysis target feature extraction, the LSTM anomaly detection model is used for sensor data processing, and the graph neural network model is used for access control recording; and a feature fusion subunit, which uses an attention mechanism to weightedly fuse the extracted multimodal features to obtain a structured feature vector.

[0009] Preferably, the workflow engine unit is built on the LangChain framework and includes: a workflow template library storing standardized workflow templates predefined for community management scenarios; a dynamic workflow generator that automatically generates a workflow DAG based on event type and context when template matching is not required; a workflow optimizer that optimizes the execution path of the workflow in real time based on reinforcement learning algorithms to minimize execution latency or resource consumption; and a workflow execution engine that instantiates and drives the step-by-step execution of the workflow and generates the decision instruction when execution reaches a preset node.

[0010] Preferably, the workflow templates stored in the workflow template library include at least the following types: standard workflows triggered based on fixed rules, template workflows for quick invocation, data workflows for data querying, and Plan workflows that can automatically generate execution plans based on temporary tasks. The automatic generation of the Plan workflow is achieved by calling the code generation unit, which decomposes the user's natural language task description based on a large language model and transforms it into executable workflow task code.

[0011] Preferably, the RAG enhancement unit includes: a multi-level vector knowledge base and an open-source knowledge base management subunit, including a first knowledge base storing property management rules, a second knowledge base storing historical disposal cases, and a third knowledge base storing device failure modes. The open-source knowledge base management subunit is used to manage and maintain the multi-level vector knowledge base and provide a streaming application interface compatible with standard protocols. A hybrid retrieval subunit is used to simultaneously perform vector-based dense retrieval and keyword-based sparse retrieval on query requests and re-rank the retrieval results. A knowledge enhancement subunit is used to integrate the retrieved contextual information with the current decision problem to form enhanced prompts, which are then fed back to the workflow engine unit.

[0012] Preferably, the RAG enhancement unit further includes a decision review subunit, which is used to determine whether the content to be output by the workflow engine unit is legal and compliant by calling the locally stored review rules and value knowledge base. If compliant, the output is allowed; if not compliant, the output is not allowed.

[0013] Preferably, the RAG enhancement unit further includes a model proxy and routing subunit, which is used to route the request to a dedicated large language model, the multi-level vector knowledge base, or a pre-trained computer vision model for processing, based on the type of the query request.

[0014] Preferably, the code generation unit, built on the QWen3Coder model, is configured to perform the following functions: receiving and parsing natural language instructions from the user or the workflow engine unit; decomposing the natural language instructions into a sequence of tasks that can be executed sequentially or in parallel based on a pre-trained large language model; generating corresponding executable code segments for each task in the task sequence based on the code generation model, wherein the executable code segments conform to the execution specifications of the workflow engine unit; assembling the generated code segments into complete workflow code, and outputting it to the workflow engine unit for dynamic loading and execution.

[0015] The beneficial effects of this invention are as follows: By reusing various sensing devices (such as surveillance, access control, and sensors) already widely deployed in the community and utilizing edge computing nodes for localized intelligent processing, this invention eliminates the need for large-scale hardware replacement or the deployment of expensive robots, greatly reducing the initial investment and barriers to smart community transformation, and enabling the popularization and promotion of advanced intelligent capabilities in ordinary residential communities; through data fusion and intelligent workflow engines, the originally isolated security monitoring data is transformed into automatic decision-making and execution, effectively reducing the repetitive inspection burden on property staff and improving management efficiency; adopting a private edge deployment architecture, all data collection, processing, and decision-making are completed in a closed loop within the community's internal network, and sensitive data (such as videos and access control records) do not need to be uploaded to the public cloud, fundamentally eliminating the risk of data leakage due to cloud server security issues, and protecting community data sovereignty and residents' privacy. Attached Figure Description

[0016] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0017] Figure 1 is a schematic diagram of the system structure of the present invention; Figure 2 is a flowchart of the system execution of the present invention. Detailed Implementation

[0018] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0019] Please refer to Figures 1-2. This embodiment provides a community life service intelligent agent system. The system includes a data acquisition module, an edge decision-making module, and a terminal execution module connected by communication. The data acquisition module includes multiple heterogeneous sensing devices deployed within the community for real-time collection of multimodal raw data of the community environment, including video surveillance, access control gates, elevator control, license plate recognition, and various IoT sensors already deployed within the community. These devices aggregate real-time data to privately deployed edge computing nodes via standard protocols (such as RTSP, ONVIF, MQTT). An edge computing node is a high-performance, high-reliability server that hosts an edge decision-making module. After analyzing and processing data, the edge decision-making module sends decision instructions to the execution terminal module. The execution terminal module includes property management personnel's PC / mobile office system, a resident's app, and controllable smart devices (such as broadcasting systems and gate barriers). The built-in units or execution processes of other functional modules of the system specifically include: an edge decision-making module, privately deployed within the community, communicating with the data acquisition module to receive and perform calculations on the multimodal raw data; the decision-making module includes a data fusion unit, a workflow engine unit, an RGA enhancement unit, and a code generation unit; the data fusion unit preprocesses the multimodal data and performs feature extraction and spatiotemporal fusion to generate structured feature vectors; the data fusion unit includes a time synchronization subunit, used for network time protocols and adaptive filtering algorithms. The system performs timestamp alignment on the multimodal raw data from different sensing devices, including using PTS / DTS timestamps on the video stream (H.264 / H.265) and adaptive Kalman filtering for time synchronization of data from each sensor. A multimodal feature extraction subunit includes an improved YOLOv5 model for target feature extraction in video analysis, an LSTM anomaly detection model for sensor data processing, and a graph neural network model for access control recording. An SE attention module is embedded in the existing YOLOv5 model to detect people, vehicles, objects, and their behaviors in real time from the video stream. The graph neural network model analyzes access control card swipe records to construct personnel flow patterns. The LSTM network analyzes sensor data sequences to detect anomalies such as temperature, humidity, and overflowing garbage. A feature fusion subunit uses a multi-head self-attention mechanism to weightedly fuse the extracted multimodal features to obtain a structured feature vector.

[0020] The workflow engine unit, built on the LangChain framework, collaborates with the data fusion unit, RAG enhancement unit, and code generation unit to match, execute, and optimize corresponding workflows based on the structured feature vectors. The execution of the workflow generates decision instructions sent to the terminal execution module. The workflow engine unit, built on the LangChain framework, includes workflow templates, a dynamic workflow generator, a workflow optimizer, and a workflow execution engine. A workflow template library stores standardized workflow templates predefined for community management scenarios. The workflow templates stored in the library include at least the following types: standard workflows triggered by fixed rules, template workflows for quick invocation, data workflows for data querying, and Plan workflows that can automatically generate execution plans based on temporary tasks. The standard workflow, based on the LangChain-developed standard workflow engine, is a standardized and repeatable workflow customized by engineers for clients during the project implementation phase, specifically for high-frequency, fixed-process community management scenarios. It is triggered directly by matching the structured feature vectors output by the data fusion unit. Each standard workflow has its corresponding feature vector pattern; Template Workflow: a configurable template based on the abstraction and encapsulation of standard workflows. It strips away specific building, device ID, and other parameters, allowing engineers or trained property managers to quickly configure and enable it during implementation or maintenance without redeveloping code. Through a graphical interface, administrators can select cameras from dropdown menus, enter building numbers, set alarm thresholds (such as the judgment duration for fire lane obstruction), select notification groups, etc. After configuration, the set trigger engine automatically binds the template to specific physical devices and organizational personnel, generating a directly runnable workflow instance; Data Workflow: a data query workflow provided by MCP Server (Model Context Protocol Server) and code generator, enabling workflow engine units to directly query data from various databases. Users can customize the data query during the interaction process. It is a special workflow designed based on interactive data query and analysis.It is not triggered by an event, but rather driven by a data exploration request initiated proactively by the user (property management staff) through natural language or forms. The process includes: request parsing: the user inputs "Check all records and screenshots of parking space number 3 being occupied by an unknown vehicle in the past week"; data connection and query generation: based on MCP Server. The system encapsulates a unified query interface for connecting to various databases within the community (license plate recognition record database, video metadata database, user vehicle whitelist database); a code generation engine (based on QWen3Coder) understands user semantics and automatically generates a series of SQL queries and API call codes, assembling them into a data workflow. For example, it queries the license plate database for all parking records of "parking space number 3" over the past 7 days; compares the license plate numbers in the records with the "owner whitelist database" to filter out "unknown vehicles"; the automatic generation of the Plan workflow is achieved by calling the code generation unit, which decomposes the user's natural language task description based on multiple large language models and transforms it into executable workflow task code. An example of this is breaking down a task into multiple thought steps based on an Agent: Step 1: Extract all video metadata from the East Gate camera between 7-9 AM and 5-7 PM; Step 2: Call the traffic flow statistics model to analyze the number of vehicles entering and exiting; Step 3: Call the event detection model to analyze whether there are abnormal events such as traffic accidents or temporary parking obstructing traffic; Step 4: ... Step 4: Link access control system data to analyze the impact of pedestrians and non-motorized vehicles on lanes during peak hours; Step 5: Based on the above data, generate an analysis report and improvement suggestions, such as "adjusting the gate response speed" and "increasing traffic control personnel during peak hours"; and transform these steps into an ordered plan diagram containing conditional judgments, and then transform the plan diagram into specific, executable LangChain workflow code, which is then executed by the execution engine step by step; The dynamic workflow generator automatically generates a workflow DAG based on event type and context when no template matching is required. The specific implementation process includes: receiving unmatched feature vectors or complex user requests; packaging the current context (feature vectors, user request text, available device list, permission information) and sending it to the code generation unit; receiving the workflow code draft returned by the code generation unit; calling the sandbox environment to simulate and test the draft to verify its logical rationality and resource access security; after the test is passed, compiling and encapsulating it into a temporary workflow instance, submitting it to the execution engine, and optionally storing it as a new template.

[0021] The workflow optimizer employs multi-objective reinforcement learning. Its state space includes workflow node states, system load, network latency, etc.; its action space includes adjusting node execution order, enabling caching, switching to backup models, etc.; and its reward function comprehensively considers task completion time, CPU / memory consumption, and decision accuracy. It monitors and dynamically adjusts the execution process of each workflow in real time to optimize execution efficiency, resource consumption, and result accuracy.

[0022] The workflow execution engine is used to instantiate and drive the step-by-step execution of the workflow, and generate the decision instructions when it reaches the preset nodes. This includes loading the workflow definition, scheduling the execution of each node sequentially or in parallel, managing data transfer between nodes, handling exceptions, and generating and issuing the final decision instructions at key nodes. The generation process of the decision instructions includes: specific nodes in the workflow are designed as action nodes. When the execution engine runs to such a node, the node will instantiate a specific, structured decision instruction object based on the input data of the upstream node. After generation, the instruction object is not issued immediately, but is sent to the subsequent decision review stage for compliance filtering. After the review is passed, the execution engine accurately sends it to the target terminal specified in the terminal execution module through a message queue or API call.

[0023] The RGA enhancement unit is used to retrieve domain knowledge from the multi-level knowledge base based on query requests from the workflow engine unit during the decision-making process, and return the retrieval results to the workflow engine unit to assist in decision optimization. The RGA enhancement unit includes a multi-level vector knowledge base and an open-source knowledge base management subunit. This includes a first knowledge base storing property management rules, a second knowledge base storing historical disposal cases, and a third knowledge base storing device failure modes. The open-source knowledge base management subunit manages and maintains the multi-level vector knowledge base and provides a streaming application interface compatible with standard protocols. The first knowledge base uses documents such as the "Property Management Regulations," the community's "Management Rules," and fire / security standards... The first knowledge base is divided into three parts: a first knowledge base and a second knowledge base. The second knowledge base is stored after slicing and vectorizing the data. The vector model is fine-tuned by text in the community management domain to improve semantic matching accuracy. The third knowledge base extracts and vectorizes successfully handled work orders and event reports in a structured (problem-decision-result) manner to form a searchable experience memory. The fourth knowledge base transforms fault patterns and abnormal patterns (such as elevator shaking features and abnormal power consumption curves of water pumps) mined from equipment logs and sensor time series data into descriptive text and quantifies them. The open-source knowledge base management subunit uses MaxKB as its core, providing a graphical interface for adding, deleting, modifying, and querying knowledge bases, replacing and optimizing vector models, and encapsulating the three knowledge bases into a unified OpenAI-compatible streaming API.This allows the workflow engine or code generation unit to query the local knowledge base via standard HTTP requests, just like calling ChatGPT, greatly simplifying integration complexity. The hybrid retrieval subunit performs both vector-based dense retrieval and keyword-based sparse retrieval on the query request simultaneously, and re-ranks the search results. The retrieval ranking process specifically includes the following first step: initiating dense and sparse retrieval in parallel. Dense retrieval uses vector models such as Contriever to encode the query statement into vectors, performs similarity searches in the vector spaces of the three knowledge bases, and recalls semantically related fragments. Sparse retrieval includes using the BM25 algorithm to analyze keywords in the query. ("Fire lane", "occupancy", "bicycle") is matched against the original text in the knowledge base to retrieve fragments with high keyword matching. The second step uses a more sophisticated but slower model, such as Cross-Encoder, to re-score and rank the relevance of the initially retrieved candidate fragments (e.g., Top-20) to the query, selecting the 3-5 most relevant fragments. The knowledge enhancement subunit integrates the retrieved contextual information (TOP fragments, which may come from different knowledge bases, such as a rule + two similar cases) with the current decision problem to form enhanced suggestions, which are then fed back to the workflow engine unit's decision review subunit. The decision review subunit uses... The system comprises several sub-units: a decision-making review sub-unit, a model proxy and routing sub-unit, and a code generation unit. The first sub-unit uses locally stored review rules and value knowledge bases to determine the legality and compliance of the content to be output by the workflow engine unit, and decides whether to allow the output. The second sub-unit uses locally stored review rules and value knowledge bases to determine the legality and compliance of the content to be output by the workflow engine unit, and decides whether to allow the output. The third sub-unit uses locally stored review rules and value knowledge bases to determine whether to allow the output. The fourth sub-unit uses locally stored model proxy and routing sub-units to route requests to a dedicated large language model, the multi-level vector knowledge base, or a pre-trained computer vision model for processing, based on the type of query request. The fifth sub-unit uses locally stored model proxy and routing sub-units to route requests to a dedicated large language model, the multi-level vector knowledge base, or a pre-trained computer vision model for processing. The sixth sub-unit uses locally stored model proxy and routing sub-units to parse and convert natural language instructions transmitted by the terminal execution module into executable workflow code for the workflow engine unit. Meta-dynamic loading and execution; built on the QWen3Coder model, configured to perform the following functions: receiving and parsing natural language instructions from the user (AI voice interaction unit of the terminal execution module) or workflow engine unit; based on a pre-trained large language model, decomposing the natural language instructions into a sequence of tasks that can be executed sequentially or in parallel, performing basic cleaning and intent classification to determine whether they belong to "data query", "process creation" or "analysis task"; utilizing the natural language understanding capabilities of the QWen3Coder model, decomposing complex instructions into a sequence of sequential or parallel atomic tasks, and generating a corresponding LangChain tool call code segment for each atomic task.These code segments conform to the LangChain Tool interface specification and can call the database API, model API, etc. encapsulated by the MCP Server. Based on the dependencies between tasks, LangChain's SequentialChain or ParallelChain primitives are used to assemble the code segments into a complete, executable LangChain workflow (Chain object or code segment). The executable code segment conforms to the execution specification of the workflow engine unit. The generated multiple code segments are assembled into complete workflow code and output to the workflow engine unit for dynamic loading and execution.

[0024] The terminal execution module receives and executes decision commands from the edge decision module and feeds back the execution results to the edge decision module. It includes a built-in AI voice interaction unit, which allows users to interact with the workflow engine and code generation unit via voice commands. The specific implementation process includes: the message center in the terminal execution module continuously listens for structured decision commands issued by the workflow engine unit through a message queue (such as Redis Streams); based on the `target` field in the command, it accurately distributes the command to the corresponding terminal: the property management PC backend, security PDA, owner's APP, or smart device controller (such as a broadcast system); through the built-in AI voice interaction unit: property personnel input voice commands through a worn smart terminal or fixed microphone; the lightweight wake word detection module on the terminal is first activated, and then the subsequent voice stream is uploaded to the speech recognition (ASR) service on the edge computing node and converted into text; the converted text is sent to a lightweight intent recognition model, which determines it to be a data query command. Subsequently, the text command is sent to the code generation unit; after the data query workflow generated by the code generation unit is executed, the result is returned to the AI ​​voice interaction unit. The text-to-speech (TTS) service converts the resulting text into speech, which is then broadcast to property management staff via the terminal, completing the closed loop. Whether the property management staff manually process the work order and click "complete," or the smart device automatically executes the instruction and returns a status code, these execution results are collected by the terminal execution module and structured into feedback messages. These feedback messages are sent back to the workflow engine unit through the uplink channel, thereby triggering subsequent closed-loop learning processes such as data archiving, case study, and model optimization.

[0025] The system's execution process includes: Multimodal data acquisition: Existing sensing devices (monitoring, access control, sensors, etc.) distributed throughout the community continuously collect multimodal raw data. Once an anomaly or pre-set event is detected (such as intrusion or obstruction of fire lanes), the relevant data is aggregated in real time to privately deployed edge computing nodes via standard protocols; Multimodal data fusion and feature extraction: The data fusion unit at the edge nodes performs timestamp alignment, cleaning, and formatting on the aggregated heterogeneous data. Subsequently, improved visual models, graph neural networks, and temporal models are invoked in parallel to extract multi-dimensional features such as video targets, behavioral patterns, and sensor anomalies, and an attention mechanism is used to fuse and generate a structured feature vector that comprehensively describes the current situation; Intelligent workflow matching and decision instruction generation: The workflow engine unit receives the feature vector and matches the most suitable pre-set workflow (such as "roadside parking handling process") in its template library. If no perfect match is found, a dynamic generator coordinates the code generation unit to create a new workflow in real time. The engine executes a workflow, during which it can call the RAG enhancement unit to retrieve domain knowledge to assist in judgment and ultimately generate structured decision instructions (such as generating alarm work orders). The decision instructions undergo review and secure issuance; all instructions must pass a compliance review by the decision review subunit before being issued. This unit calls a local review knowledge base to ensure that the instructions comply with regulations and values. After approval, the instructions are precisely issued to the target specified in the terminal execution module (such as property management staff APP, community broadcast system) via a message queue. Multi-terminal collaboration and execution result feedback are implemented: terminals receive and execute instructions, property management staff handle them on-site according to guidance, or smart devices respond automatically. After processing, the execution results (success, failure, process data) are fed back to the edge decision module through the same channel for closed-loop learning and system optimization. The feedback results, along with the original event data, are stored in the case library and used to trigger incremental training of the model and optimization of workflow parameters. For example, feedback on a false alarm can be used to reduce the trigger sensitivity of similar events, thereby enabling continuous self-learning and evolution of the system.

[0026] This invention constructs a collaborative decision-making framework with a "workflow engine" as the core scheduling hub, integrating a "RAG enhanced knowledge base" and "dynamic code generation" capabilities. This framework transforms fragmented perception data into executable intelligent workflows, further ensuring professional decision-making by injecting domain knowledge through RAG units, and achieving dynamic responses to non-preset tasks through code generation units. Thus, at the software level, it fully endows traditional communities with a closed-loop capability of "perception-decision-execution," solving the problems of single-function systems and lack of proactive decision-making and execution layers in existing technologies. Secondly, by reusing existing equipment, it avoids expensive hardware modifications; by utilizing private edge deployment, it eliminates the security risks of data migration to the cloud; and by adopting a template-based, natural language interaction design, it greatly reduces operational complexity, lowers maintenance costs and complexity, and addresses industry pain points such as cloud security vulnerabilities present in existing technical solutions.

[0027] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A community life service intelligent agent system, characterized in that: The system includes a data acquisition module, an edge decision-making module, and a terminal execution module with communication connections: the data acquisition module includes multiple heterogeneous sensing devices deployed in the community for real-time acquisition of multimodal raw data of the community environment; An edge decision-making module, privately deployed within the community, communicates with the data acquisition module to receive and perform calculations on the multimodal raw data. The decision-making module includes a data fusion unit, a workflow engine unit, an RGA enhancement unit, and a code generation unit. The data fusion unit preprocesses the multimodal data, extracts features, and performs spatiotemporal fusion to generate structured feature vectors. The workflow engine unit collaborates with the data fusion unit, RGA enhancement unit, and code generation unit to match, execute, and optimize corresponding workflows based on the structured feature vectors, and generates decision instructions to be sent to the terminal execution module based on the workflows. The RGA enhancement unit is used to retrieve domain knowledge from the multi-level knowledge base according to the query request of the workflow engine unit during the decision-making process, and return the retrieval results to the workflow engine unit to assist in decision optimization; the code generation unit is used to parse the natural language instructions transmitted by the terminal execution module and convert them into executable workflow code for the workflow engine unit to dynamically load and execute. The terminal execution module is used to receive and execute the decision instructions of the edge decision module, and to feed back the execution results to the edge decision module.

2. The community life service intelligent agent system according to claim 1, characterized in that: The terminal execution module has a built-in AI voice interaction unit, which is used by the user to interact with the workflow engine and code generation unit by voice.

3. The community life service intelligent agent system according to claim 1, characterized in that: The data fusion unit includes: a time synchronization subunit, used to timestamp-align the original multimodal data from different sensing devices based on network time protocol and adaptive filtering algorithm; a multimodal feature extraction subunit, which has built-in trained YOLOv5 model, LSTM anomaly detection model and graph neural network model, wherein the YOLOv5 model is used for video analysis target feature extraction, the LSTM anomaly detection model is used for sensor data processing, and the graph neural network model is used for access control recording; and a feature fusion subunit, which uses an attention mechanism to weightedly fuse the extracted multimodal features to obtain a structured feature vector.

4. The community life service intelligent agent system according to claim 1, characterized in that: The workflow engine unit is built on the LangChain framework and includes: a workflow template library storing standardized workflow templates predefined for community management scenarios; a dynamic workflow generator that automatically generates a workflow DAG based on event type and context when template matching is not required; a workflow optimizer that optimizes the execution path of the workflow in real time based on reinforcement learning algorithms to minimize execution latency or resource consumption; and a workflow execution engine that instantiates and drives the step-by-step execution of the workflow and generates the decision instructions when execution reaches a preset node.

5. The community life service intelligent agent system according to claim 4, characterized in that: The workflow template library stores at least the following types of workflow templates: standard workflows triggered based on fixed rules, template workflows for quick invocation, data workflows for data querying, and Plan workflows that can automatically generate execution plans based on temporary tasks. The automatic generation of Plan workflows is achieved by calling the code generation unit, which decomposes the user's natural language task description based on a large language model and transforms it into executable workflow task code.

6. The community life service intelligent agent system according to claim 1, characterized in that: The RAG enhancement unit includes: a multi-level vector knowledge base and an open-source knowledge base management subunit, comprising a first knowledge base storing property management rules, a second knowledge base storing historical disposal cases, and a third knowledge base storing device failure modes. The open-source knowledge base management subunit is used to manage and maintain the multi-level vector knowledge base and provides a streaming application interface compatible with standard protocols. A hybrid retrieval subunit is used to simultaneously perform vector-based dense retrieval and keyword-based sparse retrieval on query requests and re-rank the retrieval results. A knowledge enhancement subunit is used to fuse the retrieved contextual information with the current decision problem to form enhanced suggestions, which are then fed back to the workflow engine unit.

7. A community life service intelligent agent system according to claim 6, characterized in that: The RAG enhancement unit also includes a decision review subunit, which is used to determine whether the content to be output by the workflow engine unit is legal and compliant by calling the locally stored review rules and value knowledge base. If it is compliant, the output is allowed; if it is not compliant, the output is not allowed.

8. A community life service intelligent agent system according to claim 6, characterized in that: The RAG enhancement unit further includes a model proxy and routing subunit, which is used to route the query request to a dedicated large language model, the multi-level vector knowledge base, or a pre-trained computer vision model for processing, based on the type of the query request.

9. A community life service intelligent agent system according to claim 1, characterized in that: The code generation unit, built on the QWen3Coder model, is configured to perform the following functions: receive and parse natural language instructions from the user or the workflow engine unit; and decompose the natural language instructions into a sequence of tasks that can be executed sequentially or in parallel based on a pre-trained large language model. Based on the code generation model, an executable code segment is generated for each task in the task sequence. The executable code segment conforms to the execution specifications of the workflow engine unit. The generated code segments are assembled into complete workflow code and output to the workflow engine unit for dynamic loading and execution.

Images