Work order intelligent decision support method and device based on retrieval enhancement generation and medium

Abstract

The invention belongs to the field of artificial intelligence, and particularly relates to a work order intelligent decision support method and device based on retrieval enhancement generation and a medium. The method comprises the following steps: receiving a work order of a user, and performing intention recognition on the work order through a large language model; generating a plurality of sub-queries focusing on different dimensions according to an intention recognition result; converting the sub-queries into vectors through the same embedding model; performing approximate nearest neighbor retrieval in a pre-constructed vector knowledge base through the vector to obtain an initial candidate knowledge set containing a first number of text blocks; the initial candidate knowledge set is rearranged by calculating relevancy scores of all the text blocks in the initial candidate knowledge set and the work order, and N text blocks most relevant to the work order are screened out to serve as evidences; and constructing a cue word input large language model, and generating a work order processing guidance report. Therefore, the problem of low efficiency of manually making a work order processing guidance report in a traditional mode is solved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, specifically relating to a method, device, and medium for intelligent decision support of work orders based on retrieval enhancement generation. Background Technology

[0002] Currently, the main technical solutions applied to social affairs work order processing include traditional keyword retrieval technology and general large language model technology. Traditional keyword search is the most commonly used solution. Staff input keywords such as "water leakage," "ownership," and "noise" into the system, and the system returns a list of documents containing these words. This method is simple to implement and fast in response, but it has significant drawbacks: First, its semantic understanding capability is insufficient; it cannot understand the semantic connection between "clanging" and "noise disturbance" based on literal matching. Second, its retrieval accuracy is limited; a single search often returns a large number of documents with low relevance, requiring staff to invest significant effort in secondary manual screening. Third, the cost of knowledge updating and maintenance is high; when new policies are introduced or documents are revised, the index needs to be manually updated, resulting in high maintenance costs and potential update delays.

[0003] While artificial intelligence technologies, exemplified by large language models, possess powerful natural language understanding and generation capabilities, directly applying general-purpose models to social affairs work order processing scenarios presents several shortcomings: First, there's the issue of knowledge timeliness. The knowledge of general-purpose models is limited to the last date of their training data, failing to keep abreast of the latest policy and regulatory changes. This is particularly problematic in the social affairs field, where policies are frequently updated, leading to insufficient timeliness in decision-making recommendations. Second, there's a lack of regional knowledge. General-purpose models lack "local knowledge" and are unaware of the latest policies in specific regions. Third, professional accuracy is difficult to guarantee. Social affairs processing demands a high degree of professionalism and accuracy, while general-purpose large language models suffer from the "illusion" problem, potentially generating seemingly reasonable but actually erroneous recommendations. Fourth, there's a lack of traceability. The process of generating recommendations is like a "black box," unable to provide clear sources of evidence, failing to meet the traceability requirements of social affairs processing. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a method, device and medium for intelligent decision support of work orders based on retrieval enhancement generation.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] In a first aspect, the present invention provides a method for intelligent decision support of work orders based on retrieval enhancement generation, comprising:

[0007] Receive user work orders and perform intent recognition on the work orders through a large language model. The intent recognition includes extracting time, location, matters involved and related subject information to obtain the intent recognition result.

[0008] Based on the results of intent recognition, multiple sub-queries focusing on different dimensions are generated. The sub-queries include: source dimension, subject responsibility dimension, handling process dimension, and historical case reference identification dimension.

[0009] Each subquery is converted into a vector using the same embedding model;

[0010] By performing an approximate nearest neighbor search on the pre-built vector knowledge base using the vector, an initial candidate knowledge set containing a first number of text blocks is obtained;

[0011] The initial candidate knowledge set is rearranged by calculating the relevance score between each text block in the initial candidate knowledge set and the work order, and the N text blocks most relevant to the work order are selected as evidence, wherein the N text blocks satisfy the second quantity and at least two different dimensions of knowledge.

[0012] Inputting prompts containing role definitions, task descriptions, the evidence, the work order, and output format requirements into a large language model generates a structured work order processing guidance report. The work order processing guidance report includes the source of the basis, and the source of the basis is accompanied by source information.

[0013] Furthermore, the method for constructing the pre-built vector knowledge base includes:

[0014] Collect multi-source heterogeneous knowledge in the social affairs domain, including laws and regulations, service standards, historical case handling records, lists of the rights and responsibilities of the main entities, cross-entity collaboration standards, and historical work order processing records;

[0015] The knowledge in the social affairs domain is cleaned to obtain cleaned text. The cleaning includes removing irrelevant information, standardizing the layout, correcting semantic conflicts, and supplementing missing key elements.

[0016] A hybrid segmentation strategy is used to segment the cleaned text according to predefined text segmentation rules to obtain multiple original text blocks with complete and independent semantics. The predefined text segmentation rules are to segment according to the natural structure of the document, and to segment text without obvious structure according to semantics and set overlapping areas for text blocks.

[0017] Each of the original text blocks is converted into a vector using an embedding model;

[0018] The vector and its corresponding original text block and metadata are associated and stored in a vector database to construct a vector knowledge base. The metadata includes the publication time, publishing entity, and validity level of the source document.

[0019] Furthermore, the correction of semantic conflicts includes:

[0020] The system utilizes explicit conflict verification based on the rule base of the social affairs domain and implicit conflict resolution based on contextual semantic reasoning. The explicit conflict verification is triggered by inconsistencies in the time limits for the same matter, while the implicit conflict resolution employs a fine-tuned BERT model to analyze the relevance of conflict scenarios.

[0021] The key information to be supplemented for missing elements includes:

[0022] Supplementing timeliness information, which is completed by associating the document number with the official database; and

[0023] The entities responsible for rights and responsibilities are supplemented, and these entities are determined based on the responsible entities corresponding to the items matched by the knowledge graph.

[0024] Furthermore, the segmentation based on the document's natural structure includes:

[0025] The chapter title, clause number, and paragraph mark are used as dividing points;

[0026] The semantic segmentation of text without obvious structure and the setting of overlapping text block regions include:

[0027] A semantic similarity calculation model is used to determine the segmentation boundary, and the similarity threshold between adjacent text segments is set as the first threshold.

[0028] The length of the overlapping area of ​​the text blocks is a first percentage of the average length of the text blocks.

[0029] Furthermore, the step of converting each of the original text blocks into vectors through an embedding model includes:

[0030] The embedding model is trained using labeled corpora in the social affairs domain, and a contrastive learning loss function is introduced during training to optimize the vector space distribution.

[0031] The original text blocks are converted into vectors using the trained embedding model.

[0032] Further, the step of rearranging the initial candidate knowledge set by calculating the relevance score between each text block in the initial candidate knowledge set and the work order includes:

[0033] Obtain the relevance score between each text block in the initial candidate knowledge set and the work order;

[0034] A time decay factor is added to text blocks that involve timeliness, and the time decay factor makes the text blocks that are older than the current time have a lower weight;

[0035] The text blocks adjusted by the time decay factor are rearranged in the initial candidate knowledge set.

[0036] Furthermore, the task description includes compliance verification and feasibility analysis based on the evidence; the source information of the basis includes the document's unique identifier, specific clause number, and the full name of the publishing entity.

[0037] Furthermore, it also includes:

[0038] The system receives feedback from staff regarding the work order processing guidance report and updates the vector knowledge base or marks knowledge blind spots based on this feedback. Updating the vector knowledge base includes adding text blocks, deleting invalid text blocks, and correcting metadata. Marking knowledge blind spots triggers a knowledge collection task, which supplements the vector knowledge base with uncollected social affairs domain knowledge according to a predetermined cycle.

[0039] In a second aspect, the present invention provides an electronic device including a processor and a memory, the processor being configured to execute a computer program stored in the memory to implement the intelligent decision support method for work orders based on retrieval enhancement generation as described in any of the preceding claims.

[0040] In a third aspect, the present invention provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores at least one instruction, which, when executed by a processor, implements the intelligent decision support method for work orders based on retrieval enhancement as described above.

[0041] Compared with existing technologies, the beneficial effects of this invention are as follows: It solves the problems of low efficiency and long processing time in traditional manual processing of user work orders. It also overcomes the problem that it is difficult to generate accurate and compliant work order processing reports using a general large language model. Attached Figure Description

[0042] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0043] In the attached diagram:

[0044] Figure 1 This is a flowchart of the intelligent decision support method for work orders based on retrieval enhancement according to an embodiment of the present invention;

[0045] Figure 2 This is a flowchart of the method for constructing a vector knowledge base according to an embodiment of the present invention;

[0046] Figure 3 This is a schematic diagram of the intelligent decision support system for work orders based on retrieval enhancement according to an embodiment of the present invention;

[0047] Figure 4 This is a structural block diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0048] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0049] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0050] Example 1

[0051] To address the issue that existing work order decision-making relies heavily on human experience, resulting in a time-consuming and highly specialized process, this method is difficult to replicate. Existing large language models cannot be directly applied to the creation of work orders that demand high levels of expertise, accuracy, and timeliness. Therefore, this invention proposes a work order intelligent decision support method based on retrieval-enhanced generation.

[0052] Figure 1 This is a flowchart of the intelligent decision support method for work orders based on retrieval enhancement in an embodiment of the present invention.

[0053] Reference Figure 1 The intelligent decision support method for work orders based on retrieval enhancement of the present invention includes steps 100-600.

[0054] Step 100: Receive user work orders and perform intent recognition on the work orders through a large language model. The intent recognition includes extracting time, location, matters involved, and related subject information to obtain the intent recognition result.

[0055] Step 200: Generate multiple sub-queries focusing on different dimensions based on the results of intent recognition. The sub-queries include: source dimension, subject responsibility dimension, handling process dimension, and historical case reference dimension.

[0056] Step 300: Convert each subquery into a vector using the same embedding model;

[0057] Step 400: Perform an approximate nearest neighbor search on the pre-built vector knowledge base using the vector to obtain an initial candidate knowledge set containing a first number of text blocks;

[0058] Step 500: The initial candidate knowledge set is rearranged by calculating the relevance score between each text block in the initial candidate knowledge set and the work order, and the N text blocks most relevant to the work order are selected as evidence, wherein the N text blocks satisfy the second quantity and at least two different dimensions of knowledge.

[0059] Step 600: Input the prompt words containing role definition, task description, evidence, work order and output format requirements into the large language model to generate a structured work order processing guidance report. The work order processing guidance report includes the source of the basis, and the source of the basis is accompanied by source information.

[0060] The following sections will detail steps 100 through 600.

[0061] Step 100: Receive user work orders and perform intent recognition on the work orders using a large language model. Intent recognition includes extracting time, location, matters involved, and related subject information. Step 100 of this invention is used to receive social affairs work orders submitted by users. By using a large language model to perform deep parsing of the work order text, the true intent of the submitted work order can be identified. This is because the content of user work orders is often colloquial; directly vectorizing the content of user work orders would result in too many information points, which would cause the retrieval topic to deviate from the actual intent of the user work order. Therefore, in order to identify the true intent of the work order submitter, step 100 parses the key information of the user-submitted work order, which includes the time, location, matters involved, and related subject information. Specifically, time includes the time of the event and the processing time limit. Location includes the location of the incident and the jurisdiction. Matters involved include Item 1, Item 2, and Item 3. Related subjects include the applicant, the involved unit, and the responsible department. Step 100 forms a structured intent recognition result, which can provide accurate basis for the generation of subsequent subqueries.

[0062] This invention performs intent recognition on user-submitted work orders in step 100. This is a crucial prerequisite for generating subqueries based on the intent recognition results, as it helps to understand the true needs within the user's work order. The intent recognition in this invention uses a large language model to deeply analyze the information in the user-input work order, understanding the semantics, context, and potential user needs. By using a large language model to recognize the intent in the user's work order, this invention can identify ambiguous, implicit, or polysemous expressions in the submitted work order, transforming the text into structured intent recognition results. Only by parsing the true user intent in the work order can subqueries be generated that focus on different dimensions and meet the user's needs.

[0063] In step 200 of the present invention, multiple sub-queries focusing on different dimensions are generated based on the results of intent recognition. The sub-queries include: source dimension, subject responsibility dimension, handling process dimension, and historical case reference dimension.

[0064] In step 200 of this invention, a subquery of the source dimension is generated based on the result of intent recognition. Its function is to obtain knowledge of legitimacy and authority based on the result of intent recognition of user work orders, thereby ensuring that the processing conclusion of user work orders is based on official or authoritative source documents such as current laws and regulations, service specifications, and industry-specific management regulations. This ensures that the subsequently generated work order processing guidance report has clear and traceable policy or standard support, and avoids referencing unofficial, invalid, or ineffective knowledge content.

[0065] In step 200 of this invention, a sub-query of the subject's rights and responsibilities dimension is generated based on the result of intent recognition. Its function is to clarify the responsibilities and rights of different subjects in the events reflected in the user's work order.

[0066] In step 200 of this invention, a sub-query of the handling process dimension is generated based on the result of intent recognition. Its purpose is to help relevant personnel clearly understand the steps and processes of handling an event from start to finish.

[0067] In step 200 of this invention, a subquery of the historical case reference dimension is generated based on the result of intent recognition. Its function is to extract valuable information from relevant events or cases that have occurred in the past, and to provide reference and guidance for solving the current user ticket problem.

[0068] Through step 200 of the present invention, the ambiguous, multi-focused issues in the user work order are decomposed into multiple clear, single-focused issues, thereby providing accurate knowledge at different levels for subsequent resolution of the user work order.

[0069] Step 300: Convert each subquery into a vector using the same embedding model. Specifically, convert the text into a high-dimensional floating-point vector to represent the semantic features of the text, facilitating semantic understanding in subsequent steps.

[0070] In one optional implementation, before each subquery is input into the model, the embedding model of this invention is a domain fine-tuning model based on the BERT architecture. Specifically, the model is fine-tuned using labeled corpora. The labeled corpora of this invention include: legal and regulatory provisions, text content of user work orders, and each subquery. During the model fine-tuning process, a contrastive learning loss function is introduced. This allows for better optimization of the vector distance between semantically similar and semantically different subqueries. Consequently, the model of this invention can accurately understand the semantics of professional terms in the field of social affairs processing. After fine-tuning, the model of this invention can understand the specific meanings of "subject responsibilities" and "processing procedures," avoiding semantic misjudgments of domain descriptions by general models.

[0071] Step 300 includes two sub-steps, namely step 310 and step 320.

[0072] Step 310: Before inputting each subquery of the present invention into the same embedding model, preprocess each subquery. This is to ensure that each subquery is consistent with the representation style of the vector knowledge base in step 400. The advantage of doing so is to eliminate vector mapping differences caused by representational differences.

[0073] Step 320: The same embedding model converts each subquery into a vector. This includes the following three sub-steps: Step 321, Step 322, and Step 323.

[0074] Step 321: Perform word segmentation on the preprocessed subquery to obtain the word segmentation results.

[0075] Step 322: Semantically encode the word segmentation results, transforming the lexical information, grammatical structure, and latent semantics such as the authoritative basis and demand implied in the source dimension of the subquery into hidden layer features.

[0076] Step 323 transforms the hidden layer features into fixed-dimensional vectors. Specifically, a pooling strategy combining mean pooling and attention weights is used to compress the hidden layer features. The attention weights assign higher weights to key semantic segments representing core needs based on the semantic importance of the subquery. For example, higher weights are given to statements related to "legal basis" in the source dimension subquery and to statements related to "approval department" in the subject responsibility dimension subquery. Meanwhile, lower weights are given to non-core semantic segments. This aims to enhance the vector's ability to represent key needs. After processing with this pooling strategy, the hidden layer features are compressed into fixed-dimensional vectors, with the vector dimension consistent with the preset output dimension of the embedding model; optionally, the dimension is set to 768 dimensions. Step 323 yields subquery vectors that can be used for subsequent retrieval. Each dimension of the subquery vector corresponds to a specific semantic feature of the subquery. For example, one dimension specifically represents the semantics related to "legal source", and another dimension specifically represents the semantics related to "process steps". This allows the output subquery vector to accurately map the complete semantics of the subquery, providing a basis for the similarity ratio with the text block vector in the vector knowledge base in the subsequent step 400.

[0077] Step 400: Perform an approximate nearest neighbor search on the pre-built vector knowledge base using the vector to obtain an initial candidate knowledge set containing a first number of text blocks;

[0078] Figure 2 This is a flowchart of the method for constructing a vector knowledge base according to an embodiment of the present invention.

[0079] See Figure 2 In one specific implementation of the present invention, the method for pre-constructing a vector knowledge base includes five sub-steps, namely steps 410-450.

[0080] Step 410: Collect multi-source heterogeneous knowledge of social affairs, including laws and regulations, service standards, historical case handling records, lists of the rights and responsibilities of the main entities, cross-entity collaboration standards, and historical work order processing records.

[0081] Step 420: Clean the knowledge in the social affairs domain to obtain cleaned text. The cleaning includes removing irrelevant information, standardizing the layout, correcting semantic conflicts, and supplementing missing key element information.

[0082] In one alternative implementation, semantic conflict correction includes: explicit conflict verification based on the rule base of the social affairs domain and implicit conflict resolution based on contextual semantic reasoning; wherein, the explicit conflict verification is triggered by inconsistencies in the timeliness regulations for the same matter, and the implicit conflict resolution uses a fine-tuned BERT model to analyze the relevance of conflict scenarios.

[0083] In one alternative implementation, supplementing the missing key information elements includes: supplementing timeliness information, which is completed by associating the document number with the official database; and supplementing the responsible parties, which are determined by matching the responsible parties corresponding to the matters in the knowledge graph.

[0084] Step 430: A hybrid segmentation strategy is used to segment the cleaned text according to predefined text segmentation rules to obtain multiple original text blocks with complete and independent semantics. The predefined text segmentation rules are: segmenting according to the natural structure of the document; segmenting text without obvious structure according to semantics; and setting overlapping regions for text blocks. The social affairs domain has special requirements because segmenting the cleaned text too long introduces excessive noise, affecting retrieval accuracy; segmenting it too short may destroy semantic integrity, leading to information loss. Therefore, this invention employs a hybrid segmentation strategy.

[0085] In one specific implementation, the document is segmented according to its natural structure, including using chapter titles, clause numbers, and paragraph marks as segmentation nodes.

[0086] In one specific implementation, text without obvious structure is semantically segmented and text block overlap regions are set, including: using a semantic similarity calculation model to determine the segmentation boundary, and setting the similarity threshold of adjacent text segments as a first threshold; wherein, the length of the text block overlap region is set as a first percentage of the average length of the text blocks.

[0087] The hybrid segmentation strategy of this invention ensures that each text block can represent complete and independent semantics.

[0088] Step 440: Convert each of the original text blocks into vectors using an embedding model;

[0089] Step 450: The vector, its corresponding original text block, and metadata are associated and stored in a vector database to construct a vector knowledge base. In other words, step 450 of this invention not only stores the vector in the vector database but also associates and stores each vector with its corresponding original text block (obtained in step 430) and metadata. The metadata includes the publication time, publishing entity, and validity level of the source document. This ensures that the work order processing guidance report provided by this invention includes a source of information. These sources of information originate from the metadata.

[0090] Through steps 410-450, multi-source heterogeneous social affairs domain knowledge is constructed into a structured, semantic, and rapidly searchable vector knowledge base. Subsequently, in step 400, an approximate nearest neighbor search is performed on the pre-constructed vector knowledge base using the vectors to obtain an initial candidate knowledge set containing a first number of text blocks. Specifically, the first number can be 200; for example, an initial candidate knowledge set containing 200 text blocks is obtained.

[0091] Step 500: The initial candidate knowledge set is rearranged by calculating the relevance score between each text block in the initial candidate knowledge set and the work order, and the N text blocks most relevant to the work order are selected as evidence, wherein the N text blocks satisfy the second quantity and at least two different dimensions of knowledge.

[0092] Specifically, in step 500, the second quantity can be 8-12. This ensures that the selected evidence text blocks have sufficient information density, such as covering legal basis, handling steps, and historical cases, thus supporting the completeness of the work order processing guidance report, while avoiding redundancy due to an excessive number of texts. Furthermore, the requirement of "at least two different dimensions of knowledge," such as including both source-dimensional legal text and handling process-dimensional step text, ensures that the evidence covers the three key decision dimensions required for the work order processing guidance report: source dimension (providing compliance basis), subject responsibility dimension (clarifying the implementing entity), and handling process dimension (guiding operational steps). These three dimensions are the core basis for forming the work order processing conclusion, while the historical case reference dimension is a supplementary dimension. By covering at least two of these three key decision dimensions, the one-sidedness of the work order processing guidance report conclusion caused by relying solely on a single dimension of knowledge (such as only referring to the historical case reference dimension) can be avoided.

[0093] In one alternative implementation, rearranging the initial candidate knowledge set by calculating the relevance scores between each text block in the initial candidate knowledge set and the work order includes three sub-steps: step 510, step 520, and step 530.

[0094] Step 510: Obtain the relevance score between each text block in the initial candidate knowledge set and the work order. Specifically, obtaining the relevance score between each text block in the initial candidate knowledge set and the work order in step 510 aims to measure the semantic matching degree between each text block in the initial candidate knowledge set and the current work order, providing a basic scoring basis for subsequent reordering.

[0095] First, the same embedding model as the subquery vector transformation is used to convert the current work order text into a work order vector, thus ensuring that the work order vector and the text block vectors in the initial candidate knowledge set are in the same semantic space. Then, the cosine similarity algorithm is used to calculate the similarity value between each text block vector and the work order vector. This similarity value is the relevance score between the corresponding text block and the work order.

[0096] The relevance score ranges from [0,1]. The closer the score is to 1, the higher the degree of matching between the semantic content of the text block and the work order requirements.

[0097] For example, for work orders related to "renewal of business licenses for individual businesses," the relevance score of the "text block of clauses regarding renewal materials in the 'Regulations on the Registration and Management of Individual Businesses'" in the initial candidate knowledge set will be significantly higher than the score of the "text block explaining the social security transfer process." This sub-step can initially filter out text blocks that are semantically relevant to the work order, but it needs to be combined with the timeliness processing sub-step in subsequent step 520 to further process the reasonableness of the scores.

[0098] Step 520 involves adding a time decay factor to text blocks related to timeliness. This time decay factor reduces the weight of text blocks older than the current time. Specifically, the purpose of step 520 is to process time-sensitive knowledge in the social affairs domain to avoid work order processing guidance reports from failing to comply with current policy norms due to outdated or invalid knowledge text blocks. This is related to the metadata associated with the text blocks in the vector knowledge base in step 400. The metadata includes timeliness information such as the publication time and update time of the knowledge corresponding to the text block.

[0099] Specifically, firstly, all text blocks in the initial candidate knowledge set are traversed, and their timeliness information in the metadata is used to determine whether they belong to "text blocks involving timeliness". If the text block corresponds to legal provisions, policy norms, service process details, etc., such as "list of materials for business license renewal" or "time limit requirements for cross-departmental collaborative approval", this type of knowledge will be adjusted with policy updates, so it is determined to be a text block involving timeliness. If the text block is a general definition or basic concept, such as "definition of individual business owners", this type of knowledge is stable in the long term and has no timeliness restrictions, so it is not included in the scope of timeliness processing.

[0100] Secondly, for text blocks determined to be time-sensitive, a time decay factor is introduced to correct their relevance scores obtained in step 510. Specifically, the formula for calculating the time decay factor is: Decayed Score = Initial Relevance Score × [1 - α × (Current Time - Knowledge Publication Time) / T], where α is the decay coefficient, with a value range of [0.1, 0.3], which can be dynamically adjusted according to the knowledge update frequency in the social affairs field, with a default value of 0.2, and T is the effective period of the knowledge, set according to the knowledge type, such as setting the T value for legal knowledge to 3 years and the T value for service process knowledge to 2 years. This formula shows that the further back in time a text block is from the current time, the lower its decayed score. For example, the initial relevance score of the "Business License Renewal Materials" block published in 2021 was 0.9. When calculated in 2024, the decayed score = 0.9 × [1 - 0.2 × (2024-2021) / 3] = 0.9 × 0.8 = 0.72. However, the initial relevance score of the same type of text block published in 2023 was 0.85, and the decayed score = 0.85 × [1 - 0.2 × (2024-2023) / 3] ≈ 0.85 × 0.93 = 0.79. The latter, due to its stronger timeliness, still has a higher decayed score than the former, ensuring the scoring logic of prioritizing timeliness.

[0101] Step 530: The text blocks adjusted by the time decay factor are rearranged in the initial candidate knowledge set.

[0102] By reordering the adjusted scores, the priority of each text block in the initial candidate knowledge set is determined, providing a clear order for subsequent screening of evidence text blocks. Specifically, firstly, the final scores of all text blocks in step 520 are collected, where the scores of time-sensitive text blocks are attenuated, while the initial relevance scores of non-time-sensitive text blocks remain unchanged from step 510. Then, all text blocks in the initial candidate knowledge set are sorted in descending order of their final scores, forming a reordered sequence of knowledge text blocks. After reordering, the text block with the highest score will be given priority as candidate evidence for the work order processing guidance report. Through step 530, it is ensured that the N text blocks subsequently selected not only have a high semantic match with the work order requirements but also meet the timeliness requirements of the work order processing guidance report, avoiding decision-making biases caused by outdated knowledge.

[0103] Step 600: Input the prompt words containing role definition, task description, evidence, work order and output format requirements into the large language model to generate a structured work order processing guidance report. The work order processing guidance report includes the source of the basis, and the source of the basis is accompanied by source information.

[0104] By constructing the prompt words in step 600 and inputting them into the large language model, a structured work order processing guidance report with compliance, completeness, and traceability is generated, providing clear operational guidelines for staff to handle work orders.

[0105] The prompts include five requirements: role definition, task description, evidence, work order, and output format. Specifically, the evidence module must fully import the N text blocks selected in step 500. N must meet the second-highest quantity requirement and cover at least two key dimensions. Each piece of evidence must be labeled with its corresponding dimension attributes, such as the legal text for the source dimension and the step text for the handling process dimension, to facilitate model association and invocation.

[0106] The work order module needs to reproduce the complete information of the user work order in step 100, including the work order submission time, user request, and involved parties, to ensure that the model generates report evidence based on the actual work order scenario. The source information corresponding to each piece of evidence must be listed separately to provide the possibility for subsequent traceability.

[0107] Steps 100-600 of this invention solve the problems of low efficiency and long processing time in traditional manual processing of user work orders. It also overcomes the difficulty of generating accurate and compliant work order processing reports using general-purpose large language models.

[0108] Example 2

[0109] Figure 3 This is a schematic diagram of an intelligent decision support system for work orders based on retrieval enhancement provided by the present invention.

[0110] Reference Figure 3 The intelligent decision support system for work orders based on retrieval enhancement of the present invention includes: an intelligent service module 10, a data processing module 20, and a knowledge base module 30.

[0111] The knowledge base module 30 is used to acquire and store laws and regulations, policy documents, historical case databases, departmental responsibility lists, and other unstructured data. In one optional implementation, laws and regulations are in TXT format, policy documents in PDF format, historical case databases in DOCX format, and departmental responsibility lists in XLSX format. The data processing module 20 is used to receive the aforementioned multi-source heterogeneous data from the knowledge base module 30 and construct a vector database 24.

[0112] The data processing module 20 includes: a data loading and cleaning module 21, a text segmentation module 22, a vectorization module 23, and a vector database 24. Specifically, the data processing module 20 is used to execute steps 410-450 of Embodiment 1 of the present invention. Steps 410-450 have been described in Embodiment 1 of the present invention and will not be repeated here.

[0113] The intelligent service module 10 includes: a work order input module 11, a query conversion module 12, a vector retrieval module 13, a rearrangement and filtering module 14, a prompt word module 15, a large language model module 16, and an answer generation and tracing module 17. In this embodiment of the invention, the intelligent service module 10 sends the work order processing guidance report generated by the answer generation and tracing module 17 to the client, providing the client with results such as document correction, processing guidance, transfer suggestions, and acceptance analysis. This solves the problems of low efficiency and long processing time in traditional manual processing of user work orders. It also overcomes the difficulty of providing accurate and compliant work order processing reports using a general large language model.

[0114] Example 3

[0115] like Figure 4 As shown, the present invention also provides an electronic device 100 for implementing a work order intelligent decision support method based on retrieval enhancement generation;

[0116] The electronic device 100 includes a memory 101, at least one processor 102, a computer program 103 stored in the memory 101 and executable on at least one processor 102, and at least one communication bus 104.

[0117] The memory 101 can be used to store the computer program 103. The processor 102 implements the steps of the intelligent decision support method for work orders based on retrieval enhancement generation in Embodiment 1 by running or executing the computer program stored in the memory 101 and calling the data stored in the memory 101.

[0118] The memory 101 may primarily include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created based on the use of the electronic device 100 (such as audio data), etc. In addition, the memory 101 may include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other non-volatile solid-state storage device.

[0119] At least one processor 102 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Processor 102 may be a microprocessor or any conventional processor. Processor 102 is the control center of electronic device 100, connecting various parts of electronic device 100 via various interfaces and lines.

[0120] The memory 101 in the electronic device 100 stores multiple instructions to implement a work order intelligent decision support method based on retrieval enhancement generation, and the processor 102 can execute multiple instructions to implement the method from steps 100 to 600.

[0121] Example 3

[0122] If the modules / units integrated in the electronic device 100 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, and read-only memory (ROM).

[0123] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied 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.

[0124] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0125] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0126] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0127] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0128] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for intelligent decision support of work orders based on retrieval enhancement, characterized in that, include: Receive user work orders and perform intent recognition on the work orders through a large language model. The intent recognition includes extracting time, location, matters involved and related subject information to obtain the intent recognition result. Based on the results of intent recognition, multiple sub-queries focusing on different dimensions are generated. The sub-queries include: source dimension, subject responsibility dimension, handling process dimension, and historical case reference identification dimension. Each subquery is converted into a vector using the same embedding model; By performing an approximate nearest neighbor search on the pre-built vector knowledge base using the vector, an initial candidate knowledge set containing a first number of text blocks is obtained; The initial candidate knowledge set is rearranged by calculating the relevance score between each text block in the initial candidate knowledge set and the work order, and the N text blocks most relevant to the work order are selected as evidence, wherein the N text blocks satisfy the second quantity and at least two different dimensions of knowledge. Inputting prompts containing role definitions, task descriptions, the evidence, the work order, and output format requirements into a large language model generates a structured work order processing guidance report. The work order processing guidance report includes the source of the basis, and the source of the basis is accompanied by source information.

2. The method according to claim 1, characterized in that, The method for constructing the pre-built vector knowledge base includes: Collect multi-source heterogeneous knowledge in the social affairs domain, including laws and regulations, service standards, historical case handling records, lists of the rights and responsibilities of the main entities, cross-entity collaboration standards, and historical work order processing records; The knowledge in the social affairs domain is cleaned to obtain cleaned text. The cleaning includes removing irrelevant information, standardizing the layout, correcting semantic conflicts, and supplementing missing key elements. A hybrid segmentation strategy is used to segment the cleaned text according to predefined text segmentation rules to obtain multiple original text blocks with complete and independent semantics. The predefined text segmentation rules are to segment according to the natural structure of the document, and to segment text without obvious structure according to semantics and set overlapping areas for text blocks. Each of the original text blocks is converted into a vector using an embedding model; The vector and its corresponding original text block and metadata are associated and stored in a vector database to construct a vector knowledge base. The metadata includes the publication time, publishing entity, and validity level of the source document.

3. The method according to claim 2, characterized in that, The correction of semantic conflicts includes: The system utilizes explicit conflict verification based on the rule base of the social affairs domain and implicit conflict resolution based on contextual semantic reasoning. The explicit conflict verification is triggered by inconsistencies in the time limits for the same matter, while the implicit conflict resolution employs a fine-tuned BERT model to analyze the relevance of conflict scenarios. The key information to be supplemented for missing elements includes: Supplementing timeliness information, which is completed by associating the document number with the official database; and The entities responsible for rights and responsibilities are supplemented, and these entities are determined based on the responsible entities corresponding to the items matched by the knowledge graph.

4. The method according to claim 2, characterized in that, The segmentation based on the document's natural structure includes: The chapter title, clause number, and paragraph mark are used as dividing points; The semantic segmentation of text without obvious structure and the setting of overlapping text block regions include: A semantic similarity calculation model is used to determine the segmentation boundary, and the similarity threshold between adjacent text segments is set as the first threshold. The length of the overlapping area of ​​the text blocks is a first percentage of the average length of the text blocks.

5. The method according to claim 2, characterized in that, The step of converting each of the original text blocks into vectors through an embedding model includes: The embedding model is trained using labeled corpora in the social affairs domain, and a contrastive learning loss function is introduced during training to optimize the vector space distribution. The original text blocks are converted into vectors using the trained embedding model.

6. The method according to claim 1, characterized in that, The step of rearranging the initial candidate knowledge set by calculating the relevance score between each text block in the initial candidate knowledge set and the work order includes: Obtain the relevance score between each text block in the initial candidate knowledge set and the work order; A time decay factor is added to text blocks that involve timeliness, and the time decay factor makes the text blocks that are older than the current time have a lower weight; The text blocks adjusted by the time decay factor are rearranged in the initial candidate knowledge set.

7. The method according to claim 1, characterized in that, The task description includes compliance verification and feasibility analysis based on the evidence; the source information of the evidence includes the unique document identifier, specific clause number, and full name of the publishing entity.

8. The method according to claim 1, characterized in that, Also includes: Receive feedback from staff regarding the work order processing guidance report, and update the vector knowledge base or mark knowledge blind spots based on the feedback. The updated vector knowledge base includes adding text blocks, deleting invalid text blocks, and correcting metadata; Once the knowledge blind spot is marked, a knowledge collection task is triggered to supplement the vector knowledge base with uncollected social affairs domain knowledge according to a predetermined cycle.

9. An electronic device, characterized in that, It includes a processor and a memory, the processor being used to execute a computer program stored in the memory to implement the intelligent decision support method for work orders based on retrieval enhancement generation as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one instruction, which, when executed by a processor, implements the intelligent decision support method for work orders based on retrieval enhancement as described in any one of claims 1 to 8.

Images