Workflow decision-making system based on LLM, multi-agent and vectorization retrieval
By constructing a workflow decision-making system based on LLM, multi-agent, and vectorized retrieval, the problems of low data processing efficiency, high risk of missing clues, and high costs in public safety and judicial investigations have been solved, achieving efficient and accurate data analysis and flexible case handling.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SICHUAN SHENGBANG RUNDA TECH CO LTD
- Filing Date
- 2026-04-13
- Publication Date
- 2026-05-12
AI Technical Summary
Existing technologies are inefficient and costly when processing massive, multi-source, and unstructured data in public safety and judicial investigations. They also pose a high risk of missing clues, lack in-depth analysis capabilities, and have rigid workflows, making it difficult to adapt to the needs of different types of cases.
We construct a workflow decision system based on LLM, multi-agent, and vectorized retrieval, including a data source layer, a data processing and filtering layer, an agent engine layer, and an application layer. We employ intelligent preprocessing, multi-feature fusion, vectorized retrieval, and multi-agent collaborative reasoning to achieve efficient data filtering, association, and analysis.
It achieves more than ten times the improvement in analysis efficiency, increases the recall and accuracy of clues, reduces system costs, breaks through the context window limitation of mainstream large models in terms of processing capabilities, and has the flexibility and scalability to adapt to the rapidly changing needs of different business scenarios.
Smart Images
Figure CN122022437A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of big data processing technology, and more specifically, to a workflow decision system based on LLM, multi-agent and vectorized retrieval. Background Technology
[0002] In the fields of public safety and judicial investigations, analysts face massive amounts of unstructured data from multiple sources, including WeChat chat logs, bank statements, and call logs. Traditional analysis methods primarily rely on manual line-by-line review and rule-based keyword matching, which have the following significant drawbacks: Inefficient and costly: Manual processing of massive amounts of data is slow, prone to fatigue, and difficult to cope with the ever-increasing scale of data.
[0003] The risk of missing clues is relatively high: the keyword-based search method cannot understand the semantic context and is prone to missing case-related information that uses code words, pseudonyms or has complex logical relationships.
[0004] Lack of in-depth analysis capabilities: Traditional methods struggle to achieve automatic correlation analysis across data sources (such as linking mentions in chat logs with transfer records) and crime pattern reasoning.
[0005] Applying large models directly is costly and ineffective: If massive amounts of raw data are directly input into a large model, it will exceed its context window, resulting in information loss, incomplete analysis, and extremely high computational costs and token consumption.
[0006] Rigid workflow: Existing analysis tools have fixed workflows, making it difficult to dynamically adjust and personalize them according to different case types (such as drug crimes and duty-related crimes) and users' immediate intentions.
[0007] Therefore, there is an urgent need for an analysis system capable of intelligently, automatically, and deeply processing multi-source case data, possessing high precision, high efficiency, and high flexibility. In this regard, we propose a workflow decision-making system based on LLM, multi-agent systems, and vectorized retrieval. Summary of the Invention
[0008] The purpose of this invention is to provide a workflow decision system based on LLM, multi-agent and vectorized retrieval to solve the problems mentioned in the background art.
[0009] To achieve the above objectives, the present invention provides the following technical solution: A workflow decision-making system based on LLM, multi-agent, and vectorized retrieval comprises a data source layer, a data processing and filtering layer, an agent engine layer, and an application layer that work in sequence. It constructs a three-layer collaborative computing architecture of "data purification - intelligent scheduling - deep mining," as detailed below: The data source layer is used to provide multi-source heterogeneous data, including unstructured data and semi-structured data, and specifically includes at least one of chat logs, bank statements, and call logs. The data processing and filtering layer is used to perform intelligent preprocessing and hybrid filtering operations on the data input from the data source layer and output highly suspicious data. The operations include semantic compression preprocessing based on entity recognition, scoring and filtering by multi-feature fusion, and vectorized retrieval. The intelligent agent engine layer is a multi-agent collaborative execution engine, including a three-layer reusable intelligent agent architecture of "Function Agent→WorkFlow→Task", a dynamic workflow and knowledge management module, and a context storage module, which is used to persistently store intermediate clues and the status of analyzed data, perform intent parsing, dynamic workflow orchestration, multi-agent collaborative reasoning, and reflective decision-making on highly suspicious data. The application layer receives the analysis results from the agent engine layer and provides functions such as clue visualization, relationship graph display, analysis report output, and continuous follow-up question response based on session memory. The system employs a context-layered, isolated massive data processing mechanism, including layered data storage and task-isolated analysis, to enable incremental analysis and parallel multitasking. It also adapts to the professional knowledge requirements of different business scenarios through a dynamic modular prompting mechanism.
[0010] Preferably, the intelligent preprocessing and hybrid filtering process of the data processing and filtering layer includes the following steps: S101: Based on multi-source data input from the data source layer, construct a thematic knowledge base with "personnel, location, communication, and funds" as the core dimensions. Each knowledge base is linked through a unique identifier key to form a three-dimensional data network. The unique identifier key includes ID card number and mobile phone number. S102: For conversational raw data, including but not limited to chat logs, intelligent segmentation is performed according to time and parties involved in the conversation. Semantic compression is performed on each segment: key entities are extracted, redundant words are filtered, and after deduplication, the data is recombined into semantically condensed text. Key entities include names, locations, and amounts, and redundant words include modal particles, adverbs, and interjections. S103: Extract and score hybrid features: embed semantically condensed text into vectors, calculate the similarity score (Faiss score) with the preset case-related sample vector library, the proportion of key entities and the weighted score of the dynamic positive and negative word library, and the context reordering score, and generate a comprehensive case-related probability score through the preset fusion model. S104: Filter highly suspicious paragraphs according to preset thresholds, push them to the intelligent agent engine layer, and associate them with the corresponding topic library information.
[0011] Preferably, in the dynamic positive and negative lexicon, the negative lexicon is derived from high-frequency non-involved paragraphs, and the positive lexicon is derived from high-frequency entities in historically involved paragraphs.
[0012] Preferably, in the three-layer reusable intelligent agent architecture of "Function Agent→WorkFlow→Task": the function Agent is a dedicated functional module, the WorkFlow is a preset standardized execution chain, and the Task layer is used to perform secondary integration and orchestration of at least one WorkFlow to realize the reuse of functional modules.
[0013] Preferably, the context-layered isolation mechanism for massive data processing is as follows: it performs layered storage and analysis isolation for tens of millions of words of data, persistently stores intermediate clues and the status of analyzed data, marks "processed" data blocks, realizes incremental analysis to avoid duplicate processing, and supports parallel execution of multiple tasks and data security isolation.
[0014] Preferably, the dynamic modular prompt word mechanism splits the prompt words into "basic reasoning logic" and "expandable professional knowledge base", which supports dynamic loading and combination according to business scenarios; the reflective decision-making mechanism is implemented through code generation and logic re-evaluation. When there is ambiguity in the decision, it automatically calls supplementary information for verification and corrects the wrong decision.
[0015] Preferably, the multi-agent collaborative reasoning process of the agent engine layer includes: S201: Intent parsing and task routing: Identify the user's intent for natural language commands by fine-tuning the language model or a lightweight BERT classification model, and route them to the corresponding Task; S202: Dynamic Workflow Orchestration: Based on the requirements of the Task, the corresponding WorkFlow and Agent function are invoked to build a personalized execution process; S203: Chained execution and persistence: The function Agent executes sequentially, and uses a "divide and conquer-persistence" strategy to input highly suspicious data into LLM analysis in batches. The analysis results are stored in real time and can be called by other Agents across data sources. S204: Reflection and Verification: When the judgment output by the LLM is uncertain or contradictory, the reflection mechanism is triggered to call the professional knowledge base and related data for secondary reasoning and verification.
[0016] Preferably, the WorkFlow includes a chat history analysis WorkFlow, which consists of an information query agent, a transaction analysis agent, and a chat depth analysis agent in sequence.
[0017] Preferably, the application layer's clue visualization includes structured clue display, relationship graph drawing, and continuous follow-up questioning response based on the system's built-in session memory module, supporting users to conduct subsequent targeted follow-up questions and detailed queries on the analysis results.
[0018] Preferably, the data processing and filtering layers use the Faiss high-efficiency vector retrieval library to calculate vector similarity, reducing vector computation costs; semantic compression preprocessing technology can reduce token consumption for large model processing, saving at least 30% of computation costs.
[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) Revolutionary improvement in analysis efficiency: Through intelligent filtering, the amount of data that needs to be examined manually and by large models is reduced by 1-2 orders of magnitude. Combined with multi-agent parallel pipeline operation, the overall clue discovery efficiency is improved by more than ten times.
[0020] (2) The clue recall rate and accuracy are high: the hybrid filtering model combines semantic and statistical features to effectively reduce false negatives; the agent's reflection mechanism and professional knowledge base are enhanced, which greatly reduces false positives and logical errors.
[0021] (3) The system cost has been significantly optimized: the "semantic compression" technology directly reduces the API call cost of large models by about 30%; the use of efficient vector retrieval such as Faiss to replace complex classification models trained with a large amount of labeled data reduces training costs and data dependence.
[0022] (4) Breakthrough in processing capacity: Through the “context hierarchical isolation” and “incremental persistence” mechanisms, the system is able to handle single-task data of tens of millions of words with ease, breaking through the context window limitation of mainstream large models.
[0023] (5) Strong flexibility and scalability: The three-tier architecture supports flexible assembly and reuse of workflows like building blocks; modular prompts and knowledge bases support low-code updates and cross-domain adaptation of business knowledge, meeting the rapidly changing needs of reconnaissance business. Attached Figure Description
[0024] Figure 1 This is a schematic diagram of the system structure of the present invention; Figure 2 This is a schematic diagram of the intelligent data preprocessing and hybrid filtering process of the present invention; Figure 3 This is a schematic diagram of the workflow of the multi-agent collaborative execution engine of the present invention. Detailed Implementation
[0025] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0026] Example: Please see Figure 1-3The workflow decision-making system based on LLM, multi-agent, and vectorized retrieval includes a data source layer, a data processing and filtering layer, an agent engine layer, and an application layer that work in sequence. It constructs a three-layer collaborative computing architecture of "data purification - intelligent scheduling - deep mining," as detailed below: The data source layer is used to provide multi-source heterogeneous data, including unstructured data and semi-structured data, and specifically includes at least one of chat logs, bank statements, and call logs.
[0027] The data processing and filtering layer performs intelligent preprocessing and hybrid filtering operations on the data input from the data source layer, outputting highly suspicious data. The operations include semantic compression preprocessing based on entity recognition, scoring and filtering based on multi-feature fusion, and vectorized retrieval. Specifically, the data processing and filtering layer uses the Faiss high-efficiency vector retrieval library to calculate vector similarity, reducing vector computation costs. Semantic compression preprocessing technology can reduce token consumption in large model processing, saving at least 30% of the computational cost.
[0028] The intelligent agent engine layer is a multi-agent collaborative execution engine, comprising a three-layer reusable intelligent agent architecture of "Agent → Workflow → Task", a dynamic workflow and knowledge management module, and a context storage module. This module is used to persistently store intermediate clues and the status of analyzed data, perform intent parsing on highly suspicious data, dynamically orchestrate workflows, conduct multi-agent collaborative reasoning, and make reflective decisions. Within the three-layer reusable intelligent agent architecture of "Agent → Workflow → Task": the Agent layer is a dedicated functional module, the Workflow layer is a pre-defined standardized execution chain, and the Task layer is used to perform secondary integration and orchestration of at least one Workflow, enabling the reuse of functional modules and reducing process development and maintenance costs.
[0029] The application layer receives the analysis results from the agent engine layer and provides functions such as clue visualization, relationship graph display, analysis report output, and continuous follow-up question response based on the session memory. Specifically, the application layer's clue visualization includes structured clue display and relationship graph drawing. The continuous follow-up question response is implemented based on the system's built-in session memory module, which supports users to conduct subsequent targeted follow-up questions and detailed queries on the analysis results.
[0030] The system employs a context-layered, isolated massive data processing mechanism, including layered data storage and task-isolated analysis, to achieve incremental analysis and multi-task parallelism. A dynamic, modular prompting mechanism adapts to the professional knowledge requirements of different business scenarios. Specifically, the context-layered, isolated massive data processing mechanism performs layered storage and analysis isolation for tens of millions of words of data, persistently stores intermediate clues and the status of analyzed data, marks "processed" data blocks, enables incremental analysis to avoid redundant processing, and supports parallel execution of multiple tasks and secure data isolation.
[0031] In this application, the intelligent preprocessing and hybrid filtering process of the data processing and filtering layer includes the following steps: S101: Based on multi-source data input from the data source layer, construct a thematic knowledge base with "personnel, location, communication, and funds" as the core dimensions. Each knowledge base is linked through a unique identifier key to form a three-dimensional data network. The unique identifier key includes ID card number and mobile phone number. S102: For conversational raw data, including but not limited to chat logs, intelligent segmentation is performed according to time and parties involved in the conversation. Semantic compression is performed on each segment: key entities are extracted, redundant words are filtered, and after deduplication, the data is recombined into semantically condensed text. Key entities include names, locations, and amounts, and redundant words include modal particles, adverbs, and interjections. S103: Extract and score hybrid features: Vector embedding is performed on the semantically condensed text. The similarity score (Faiss score) with the preset case-related sample vector library, the proportion of key entities, the weighted score of the dynamic positive and negative word libraries, and the contextual reordering score are calculated. A comprehensive case-related probability score is generated through a preset fusion model. In the dynamic positive and negative word libraries, the negative word library comes from high-frequency non-case-related paragraphs, and the positive word library comes from high-frequency entities in historical case-related paragraphs. Specifically, the dynamic modular prompt word mechanism splits prompt words into "basic reasoning logic" and "expandable professional knowledge base," supporting dynamic loading and combination according to business scenarios. The reflective decision-making mechanism is implemented through code generation and logic re-evaluation. When there is ambiguity in the decision, supplementary information is automatically called for verification to correct erroneous decisions.
[0032] S104: Filter highly suspicious paragraphs according to preset thresholds, push them to the intelligent agent engine layer, and associate them with the corresponding topic library information.
[0033] In this application, the multi-agent collaborative reasoning process of the agent engine layer includes: S201: Intent parsing and task routing: Identify the user's intent for natural language commands by fine-tuning the language model or a lightweight BERT classification model, and route them to the corresponding Task; S202: Dynamic Workflow Orchestration: Based on the requirements of the Task, the corresponding WorkFlow and function Agent are called to build a personalized execution process; among them, the WorkFlow includes the Chat Log Analysis WorkFlow, which is composed of the Information Query Agent, Transaction Analysis Agent, and Chat Depth Analysis Agent in sequence.
[0034] S203: Chained execution and persistence: The function Agent executes sequentially, and uses a "divide and conquer-persistence" strategy to input highly suspicious data into LLM analysis in batches. The analysis results are stored in real time and can be called by other Agents across data sources. S204: Reflection and Verification: When the judgment output by the LLM is uncertain or contradictory, the reflection mechanism is triggered to call the professional knowledge base and related data for secondary reasoning and verification.
[0035] I. Overview of Implementation Examples: Using the application scenario of "uncovering clues about Li Si's suspected job-related crimes," this paper details the specific implementation process of a workflow decision-making system based on LLM, multi-agent, and vectorized retrieval. In this embodiment, the system needs to process multi-source case-related data related to Li Si. Through a three-layer collaborative computing architecture of "data purification - intelligent scheduling - deep mining," it achieves automatic screening, correlation reasoning, and accurate output of case-related clues, verifying the technical feasibility and practical application value of the invention.
[0036] II. Implementation Environment Description Hardware environment: The server is configured with an Intel Xeon Platinum 8375C CPU, 128GB of memory, and a 2TB SSD, supporting multi-task parallel computing; the client is a regular office computer used to receive the system's output visualization results and interactive operations.
[0037] Software environment: The operating system is Linux CentOS 8.4; the LLM uses the GPT-4 Turbo model, and the API calls are made via HTTP requests; the vector retrieval library uses Faiss version 1.7.4; the database uses MySQL 8.0 to store the topic knowledge base, intermediate clues, and the status of analyzed data; the front-end visualization uses ECharts to draw the relationship graph and display clues.
[0038] Data preparation: The multi-source heterogeneous data provided by the data source layer includes: 100,000 WeChat chat records between Li Si and related persons (unstructured data), 30,000 bank statements of Li Si and suspected related parties (semi-structured data), and 20,000 call logs (semi-structured data); the professional knowledge base includes "Compilation of Laws and Regulations on Duty-Related Crimes", "Guidelines for Compliance of Duty-Related Behaviors", and "Database of Slang Terms Related to Cases".
[0039] III. Specific Implementation Process (a) Data provision at the data source layer According to the preset interface specifications, the data source layer synchronizes the above-mentioned heterogeneous data from multiple sources, such as WeChat chat records, bank statements, and call logs, to the data processing and filtering layer. Among them, chat records are unstructured text data, bank statements contain semi-structured fields such as transaction time, transaction amount, counterparty, and account information, and call logs contain semi-structured fields such as caller ID, called number, call duration, and call time.
[0040] (II) Data Processing and Filtering Layer: Intelligent Preprocessing and Hybrid Filtering Execution Step S101: Construction and Association of Multi-Dimensional Topic Knowledge Base Based on synchronized multi-source data, the system automatically constructs a thematic knowledge base with "personnel, location, communication, and funds" as its core dimensions. Personnel database: Extracted information such as names, ID numbers, mobile phone numbers, and work units of 12 related individuals, including Li Si and Wang Wu (contacts that frequently appeared in the chat history); Funds Theme Database: Integrates account information and transaction records from bank statements, linking personal ID numbers with bank accounts; Communication Topic Database: Extracts number information and call relationships from call logs, and associates mobile phone numbers with corresponding personnel; Location-based database: Extracts three key location information points mentioned in chat logs and call logs, including office locations and project sites.
[0041] Each subject knowledge base is linked by unique identifiers such as ID card number and mobile phone number, forming a three-dimensional data network of "personnel-funds-communication-location".
[0042] Step S102: Session data segmentation and semantic compression The system intelligently segments 100,000 WeChat chat records by "single day + conversation parties", resulting in 2,300 data segments. Semantic compression is then performed on each segment. Entity recognition: Extract key entities such as names (e.g., Li Si, Wang Wu), locations (e.g., the site of the XX project), and amounts (e.g., 50,000 yuan, 100,000 yuan) from each segment; Redundancy filtering: Remove interjections such as "ah," "oh," and "ne," adverbs such as "very" and "extremely," interjections such as "!" and "?", and meaningless small talk; Deduplication and Recombination: After deduplicating the extracted key entities and core words, the text is recombined into semantically condensed text according to the logic of "person-event-amount-location". For example, the original chat record "Li Si: Wang Wu, I have already asked the finance department to transfer the kickback for the XX project to your card. It's 50,000 yuan in total. Please check it." After semantic compression, it becomes "Li Si and Wang Wu transferred 50,000 yuan in kickback for the XX project to the finance department".
[0043] This step reduced the total amount of original text in 100,000 chat records from 8 million words to 2.4 million words, and reduced token consumption by about 70%, thus lowering the computational cost for subsequent processing.
[0044] Step S103: Hybrid Feature Extraction and Scoring Vector embedding: The Sentence-BERT model was used to embed vectors into 2300 semantically condensed texts, generating 768-dimensional vectors; Feature 1 Calculation: Using the Faiss high-efficiency vector retrieval library, calculate the similarity score (Faiss score) between each semantically condensed text vector and the preset sample vector library of duty-related crimes. The sample library contains 1,000 highly suspicious text vectors from historical duty-related crime cases. Feature 2 Calculation: Calculate the percentage of key entities in each segment (e.g., a segment has 15 words, 10 of which are key entities, accounting for 66.7%), and combine this with a dynamic positive and negative word library weighted score—the positive word library contains 300 historically high-frequency entities involved in cases such as "kickbacks," "rebates," and "project revenue sharing," while the negative word library contains 500 high-frequency non-involved words such as "advertisements," "weather," and "daily meals." This segment has a 20% higher weighted score because it contains the positive word "kickbacks." Feature 3 calculation: The paragraph importance score is calculated by combining the semantic relevance of the three related segments before and after the segment using a lightweight BERT reordering model. Overall score: The scores of the above three features are input into the fusion model with a weight of 4:3:3 to generate an overall probability score of involvement in the case for each segment (out of 100).
[0045] Step S104: Screening of highly suspicious paragraphs With a comprehensive scoring threshold of 60 points, the system automatically filters out 2195 low-scoring segments (mostly casual conversations and work-irrelevant communications), selects 105 highly suspicious segments, and simultaneously associates them with the corresponding personnel, funds, and communication topic database information, pushing them to the intelligent agent engine layer.
[0046] (III) Agent Engine Layer: Multi-agent Cooperative Reasoning Execution Step S201: Intent Resolution and Task Routing The user inputs the natural language command "Mining clues about Li Si's job-related crimes" through the client. The system uses a finely tuned lightweight BERT classification model to identify the intent, determine the core intent as "mining clues about job-related crimes", and route it to the corresponding "Task for Mining Clues about Job-Related Crimes".
[0047] Step S202: Dynamic orchestration of workflow The "Task for Uncovering Clues to Duty-Related Crimes" automatically calls two standardized Workflows for secondary integration and arrangement: Chat log analysis Workflow: The execution chain is constructed in the order of "Information Query Agent → Transaction Analysis Agent → Chat In-Depth Analysis Agent"; Cross-data source association Workflow: Construct the execution chain in the order of "Fund Flow Analysis Agent → Personnel Relationship Analysis Agent → Clue Composite Agent"; The two Workflows work together to form a personalized execution process, covering the entire process of "text analysis - fund verification - personnel association - clue integration".
[0048] Step S203: Chained Execution and Persistence Information Query Agent: Calls the personnel subject database to obtain Li Si's job information (project manager of a certain unit), his relatives and the identity information of Wang Wu (a certain engineering contractor), and persists the results to the context storage module; Chat depth analysis agent: Receives 105 highly suspicious segments and uses a "divide and conquer-persistence" strategy, inputting them into the GPT-4 Turbo model in batches of 30,000 words / batch (approximately 30 segments / batch) for analysis. The first batch of analysis revealed "Li Si mentioned paying Wang Wu a kickback of 50,000 yuan for project XX," which was immediately persisted as clue 1; the second batch of analysis revealed "Li Si and Wang Wu frequently discussed the bidding details of project XX, mentioning 'circumventing review,'" which was persisted as clue 2; subsequent batches successively analyzed 3 related clues, all of which were stored in real time and marked as "processed." Transaction Analysis Agent: Calls the fund subject database, extracts the bank statement information of Li Si and Wang Wu, and finds that the "50,000 transfer" mentioned in Clue 1 corresponds to an interbank transfer on March 15, 2024, with the counterparty being Wang Wu's personal account. The transfer record is persisted after being associated with Clue 1. Funds Flow Analysis Agent: Further investigation into the subsequent fund flows of Wang Wu's account revealed that the funds were used to purchase wealth management products without any legitimate business transactions to support them, generating abnormal fund clues; Personnel Relationship Analysis Agent: By calling the communication topic database, it was found that the frequency of calls between Li Si and Wang Wu during the bidding period of the XX project increased by 3 times compared to usual, and the total call duration exceeded 10 hours, generating abnormal communication clues; The composite clue agent calls all the above persistent clues, performs cross-data source association, and generates a composite clue: "Li Si, as the person in charge of the XX project, had frequent communication with contractor Wang Wu during the project bidding period and paid Wang Wu a bribe of 50,000 yuan, which is suspected to be a money-for-power transaction."
[0049] Step S204: Reflection and Verification When the chat depth analysis agent was analyzing the nature of "kickbacks" in a certain segment, it triggered a reflection mechanism because the text did not explicitly state whether it was directly related to the project. The system automatically retrieves the "Guidelines for Compliance of Job Conduct" professional knowledge base to obtain the compliance clauses regarding "project leaders paying contractors fees without legal basis, suspected of committing job-related crimes"; The fund subject database was consulted, confirming that the 50,000 yuan transfer lacked corresponding legal contracts, invoices, or other supporting documentation. Use the personnel relationship analysis agent to verify that Wang Wu's company is the winning bidder for the XX project; Based on the supplementary information above, the large model re-reasoned and confirmed that the "kickbacks" were directly related to official conduct, corrected the initial vague judgment, and finally clarified the legality and relevance of the clues.
[0050] (iv) Application Layer: Result Output and Interaction Clue visualization: The system displays 5 core clues and 3 related clues in a structured list, labeling the source of each clue (chat logs / bank statements / call logs), related personnel, time points, and credibility score (all ≥85 points). Relationship Graph: Using ECharts, a relationship graph of "Li Si - Wang Wu - XX Project - 50,000 Transfer" is drawn to visually demonstrate the connection paths between people, projects, and funds; Analysis Report: Automatically generates "Analysis Report on Clues of Li Si's Suspected Job-Related Crimes", which includes a summary of clues, data sources, reasoning process, legal basis, and suggestions for further investigation; Response to follow-up inquiries: When the user subsequently enters "What other details are there regarding the XX project?", the system quickly responds based on the analyzed data stored in the session memory module: "Li Si and Wang Wu mentioned 'adjustment of bidding parameters' in their chat on March 8, 2024. The corresponding call record shows that the call lasted 45 minutes that day, and there is no publicly available information on the filing of adjustments to bidding parameters."
[0051] IV. Implementation Results Verification Efficiency Improvement: The system processed 8 million words of multi-source data in just 40 minutes. The data filtering process reduced the amount of data that needed to be analyzed by the large model from 8 million words to 75,000 words (a reduction of 99.06%), and the overall efficiency of clue discovery was 15 times higher than that of traditional manual analysis. Cost optimization: Semantic compression technology reduces token consumption for large models by 70%, and combined with Faiss vector retrieval to replace training of complex classification models, the overall computational cost is reduced by 38%; Accuracy verification: All 5 core clues selected were verified by manual inspection and found to be valid leads related to the case, with no omissions; the reflection mechanism corrected 2 initial vague judgments, resulting in a false alarm rate of 0. Breakthrough in processing capabilities: The system successfully processed 8 million words of data (far exceeding the single context window limit of mainstream large models). Through the incremental persistence mechanism, the subsequent analysis of 20,000 new chat records of Li Si was completed in only 5 minutes, with no duplicate processing.
[0052] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.
Claims
1. A workflow decision-making system based on LLM, multi-agent, and vectorized retrieval, characterized in that, The architecture comprises a data source layer, a data processing and filtering layer, an intelligent agent engine layer, and an application layer that work in sequence, constructing a three-layer collaborative computing architecture of "data purification - intelligent scheduling - deep mining," as detailed below: The data source layer is used to provide multi-source heterogeneous data, including unstructured data and semi-structured data, and specifically includes at least one of chat logs, bank statements, and call logs. The data processing and filtering layer is used to perform intelligent preprocessing and hybrid filtering operations on the data input from the data source layer and output highly suspicious data. The operations include semantic compression preprocessing based on entity recognition, scoring screening based on multi-feature fusion, and vectorized retrieval. The intelligent agent engine layer is a multi-agent collaborative execution engine, including a three-layer reusable intelligent agent architecture of "Function Agent→WorkFlow→Task", a dynamic workflow and knowledge management module, and a context storage module, which is used to persistently store intermediate clues and the state of analyzed data, perform intent parsing, dynamic workflow orchestration, multi-agent collaborative reasoning, and reflective decision-making on highly suspicious data. The application layer receives the analysis results from the agent engine layer and provides functions such as clue visualization, relationship graph display, analysis report output, and continuous follow-up question response based on session memory. The system employs a context-layered, isolated massive data processing mechanism, including layered data storage and task-isolated analysis, to achieve incremental analysis and multi-task parallelism. It also adapts to the professional knowledge requirements of different business scenarios through a dynamic modular prompt word mechanism.
2. The workflow decision-making system based on LLM, multi-agent, and vectorized retrieval according to claim 1, characterized in that: The intelligent preprocessing and hybrid filtering process of the data processing and filtering layer includes the following steps: S101: Based on the multi-source data input from the data source layer, construct a thematic knowledge base with "personnel, location, communication, and funds" as the core dimensions. Each knowledge base is linked through a unique identifier key to form a three-dimensional data network. The unique identifier key includes ID card number and mobile phone number. S102: For conversational raw data, including but not limited to chat logs, intelligent segmentation is performed according to time and parties involved in the conversation. Semantic compression is performed on each segment: key entities are extracted, redundant words are filtered, and after deduplication, the data is recombined into semantically condensed text. The key entities include names, locations, and amounts. The redundant words include modal particles, adverbs, and interjections. S103: Extract and score hybrid features: embed semantically condensed text into vectors, calculate the similarity score (Faiss score) with the preset case-related sample vector library, the proportion of key entities and the weighted score of the dynamic positive and negative word library, and the context reordering score, and generate a comprehensive case-related probability score through the preset fusion model. S104: Filter highly suspicious paragraphs according to preset thresholds, push them to the intelligent agent engine layer, and associate them with the corresponding topic library information.
3. The workflow decision-making system based on LLM, multi-agent, and vectorized retrieval according to claim 2, characterized in that: In the dynamic positive and negative lexicon, the negative lexicon comes from high-frequency non-involved paragraphs, and the positive lexicon comes from high-frequency entities in historically involved paragraphs.
4. The workflow decision-making system based on LLM, multi-agent, and vectorized retrieval according to claim 1, characterized in that: In the three-layer reusable intelligent agent architecture of "Agent→WorkFlow→Task", the Agent is a dedicated functional module, the WorkFlow is a pre-defined standardized execution chain, and the Task layer is used to perform secondary integration and orchestration of at least one WorkFlow to realize the reuse of functional modules.
5. The workflow decision-making system based on LLM, multi-agent, and vectorized retrieval according to claim 1, characterized in that: The context-layered isolation mechanism for massive data processing specifically involves: performing layered storage and analysis isolation on tens of millions of words of data, persistently storing intermediate clues and the status of analyzed data, marking "processed" data blocks, enabling incremental analysis to avoid redundant processing, and supporting parallel execution of multiple tasks and secure data isolation.
6. The workflow decision system based on LLM, multi-agent, and vectorized retrieval according to claim 1, characterized in that: The dynamic modular prompt word mechanism breaks down prompt words into "basic reasoning logic" and "expandable professional knowledge base", supporting dynamic loading and combination according to business scenarios; the reflective decision-making mechanism is implemented through code generation and logic re-evaluation. When there is ambiguity in the decision, it automatically calls supplementary information for verification and corrects the wrong decision.
7. The workflow decision system based on LLM, multi-agent and vectorized retrieval according to claim 1, characterized in that: The multi-agent collaborative reasoning process of the agent engine layer includes: S201: Intent parsing and task routing: Identify the user's intent for natural language commands by fine-tuning the language model or a lightweight BERT classification model, and route them to the corresponding Task; S202: Dynamic Workflow Orchestration: Based on the requirements of the Task, the corresponding WorkFlow and Agent function are invoked to build a personalized execution process; S203: Chained execution and persistence: Function agents execute sequentially, and use a "divide and conquer-persistence" strategy to input highly suspicious data into LLM analysis in batches. The analysis results are stored in real time and can be called by other agents across data sources. S204: Reflection and Verification: When the judgment output by the LLM is uncertain or contradictory, the reflection mechanism is triggered to call the professional knowledge base and related data for secondary reasoning and verification.
8. The workflow decision-making system based on LLM, multi-agent, and vectorized retrieval according to claim 7, characterized in that: The Workflow includes a chat history analysis Workflow, which consists of an information query agent, a transaction analysis agent, and a chat depth analysis agent in sequence.
9. The workflow decision system based on LLM, multi-agent and vectorized retrieval according to claim 1, characterized in that: The application layer's clue visualization includes structured clue display and relationship graph drawing. The continuous follow-up questioning response is implemented based on the system's built-in session memory module, supporting users to conduct subsequent targeted follow-up questions and detailed queries on the analysis results.
10. The workflow decision system based on LLM, multi-agent and vectorized retrieval according to claim 1, characterized in that: The data processing and filtering layer uses the Faiss high-efficiency vector retrieval library to calculate vector similarity, reducing vector computation costs; the semantic compression preprocessing technology can reduce token consumption in large model processing, saving at least 30% of computation costs.