Engineering design report review method, system and device based on agent arrangement

Abstract

The application discloses an engineering design report review method, system and equipment based on an intelligent agent arrangement, 1, obtaining an engineering design report to be reviewed; 2, loading applicable review rules from a structured review rule library according to the report stage and the engineering type; 3, dynamically generating a review task sequence by a review intelligent agent according to the report stage, the engineering type and the review range; 4, arranging rules for each chapter; 5, for the chapter arrangement conclusion, the content requirement rules associated with the matched chapter are executed by the deterministic rule engine in terms of each review item, including parameter extraction, synonym mapping, unit normalization and tolerance comparison; 6, for each chapter arrangement conclusion and content review conclusion, evidence records are forcedly bound; 7, according to the review type and the review conclusion, a judgment word caliber is selected from a judgment word template set, interactive review opinion data carrying a stable identifier is generated and output. The application is suitable for the engineering design report with heterogeneous chapter arrangement through five mode matching of chapter semantics.

Description

Technical Field

[0001] This invention relates to the field of computer data processing technology, and in particular to a method, system and device for reviewing engineering design reports based on intelligent agent orchestration. Background Technology

[0002] With the continuous advancement of infrastructure construction in water conservancy, hydropower, transportation, and municipal engineering, the workload of preparing and reviewing engineering design reports has increased dramatically. Taking water conservancy and hydropower projects as an example, project proposals, feasibility study reports, and preliminary design reports need to be reviewed in accordance with relevant preparation procedures and design specifications. The review content covers multiple aspects, including the completeness of chapter arrangement, the sufficiency of content coverage, the consistency of design parameters, and the compliance of technical solutions. This type of review heavily relies on manual review by senior experts, resulting in problems such as long cycles, high costs, and inconsistent review standards. Therefore, automated and intelligent review of reports has become an important requirement in the field of engineering digitalization. The core technical challenge lies in the significant differences in chapter arrangement and deliverable presentation methods among reports prepared by different design units, as well as the diverse types of engineering structures, making it difficult to definitively implement review rules.

[0003] Currently, existing systems for automated document review typically employ the following approaches: First, they use hard-coded rules or big data statistics to write review requirements as fixed keyword rules or template matching logic. However, the chapter arrangement and presentation of engineering design reports vary from design unit to design unit, resulting in numerous variations in chapter names and parameter titles for the same review requirement. Fixed rules cannot be exhaustively applied, and the cost of rule maintenance increases dramatically with the source of the report and the type of project. Second, they directly use large language models to review the entire report. However, large language models suffer from illusion problems, potentially fabricating standard clause numbers and content as review basis. Furthermore, the review conclusions lack traceable original text evidence, making judgment difficult in long document scenarios. The methods suffer from several drawbacks: First, some methods exhibit poor stability, failing to meet the reliability requirements of engineering reviews. Second, while retrieval-enhanced generation methods supplement the large language model with evidence from a retrieval knowledge base, these methods only address the "finding evidence" problem and do not solve the closed-loop judgment problem of "rule elements—report evidence—review process—conclusion." For review items requiring deterministic comparison, such as parameter consistency and calculation verification, the correctness of the judgment cannot be guaranteed. Third, methods that fine-tune the model based on review rules require retraining when review rules or procedures change, failing to take effect immediately with procedure updates, and incurring high training costs and poor interpretability. Fourth, methods that combine intelligent agent architecture with model fine-tuning and retrieval enhancement rely on intelligent... The approach involves dynamically arranging metadata perception and review tasks for reports, followed by formal reviews such as terminology, calculation, and contractual consistency by a large language model fine-tuned for specific review tasks, combined with a domain knowledge base. While this alleviates the efficiency bottleneck of multi-task review of long documents, it still relies on task-specific fine-tuning. When procedures or rules change, the dataset needs to be rebuilt and retrained. Its hybrid mechanism involves the large language model extracting data first and predefined rules being verified afterward. All review items in the report need to call the large language model, resulting in high computational overhead and insufficient reliability for review items requiring deterministic comparison. Its review conclusions do not carry traceable original text evidence and lack awareness and governance of the hierarchy and usability of review basis. Furthermore, its structural analysis is report-oriented. The method of segmenting and aggregating reports itself does not solve the problem of semantic alignment between the report chapter arrangement and the requirements of the compilation procedure; sixth, the method of scoring the overall compliance of the report structure according to the standard template library and scheduling the review strategy accordingly, evaluates the compliance of the report structure and determines the review threshold through statistical measures such as chapter coverage and chapter order consistency. However, this type of method achieves the gating of the review strategy by scoring the overall score, but does not establish a semantic mapping and arrangement judgment between each requirement of the procedure chapter and the report chapter, and cannot output the missing conclusions at the rule granularity; its review execution relies on a large language model combined with vector retrieval, and there are also reliability and computational overhead problems for deterministic comparison review items, and it does not involve the governance of the availability of review basis.

[0004] For example, a certain drafting procedure requires a "water-retaining structure" chapter in the preliminary design report. However, different design units may categorize this content under specific headings such as "roller-compacted concrete gravity dam" or "earth-rock dam design," or combine it with a chapter on spillway structures. A fixed rule might directly deem this a "missing chapter," leading to misjudgment. Another example is when using a large language model for review. The model might output a conclusion that "complies with Article XXX of a certain standard" without providing supporting evidence in the report. However, the article number and content are fabricated by the model, making verification impossible for reviewers.

[0005] In summary, existing automated review methods for engineering design reports have the following main problems: 1. Difficulty in implementing rule determinism: The report chapter arrangement and expression are heterogeneous, and hard-coded rules and template matching cannot exhaustively list variations, resulting in a high misjudgment rate and high maintenance costs; even if an intelligent agent is introduced to adaptively parse the report structure and arrange tasks, it only completes the segmentation and aggregation of the report's own structure, without solving the alignment determination between the report chapters and the chapters required by the procedure; the method of gating by scoring the overall compliance of the standard template library can only provide an overall measure of the degree of structural compliance, and cannot locate the specific missing arrangement requirements of the procedure. 2. Unreliable review conclusions: Direct review using large language models carries the risk of fabricated evidence, and the conclusions do not carry traceable evidence, failing to meet the reliability requirements of engineering review; existing methods using intelligent agent architecture also fail to bind traceable evidence to review conclusions, and lack governance over the hierarchy and availability of review evidence. 3. The judgment link is not closed: The retrieval enhancement method only solves the problem of obtaining evidence and lacks a closed judgment mechanism of "rule elements - report evidence - review process - conclusion". Even if a hybrid mechanism of first extracting with a large language model and then verifying predefined rules is adopted, the judgment of the deterministic comparison review items still depends on the output of the large language model, which is not reliable enough. Moreover, the large language model is called for all review items, resulting in high computational cost. 4. High cost of rule updates: Fine-tuning routes (including task-specific fine-tuning under the agent architecture) require rebuilding training data and retraining the model when the procedure version changes, and cannot take effect immediately. Summary of the Invention

[0006] In view of this, the purpose of this invention is to provide an engineering design report review method, system and device based on intelligent agent orchestration, which can adapt to heterogeneous report structures, make review conclusions traceable and controllable and make rule updates effective immediately, so as to meet the needs of actual engineering review scenarios.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: The intelligent review method for engineering design reports based on intelligent agent orchestration described in this invention includes the following steps: S1. Obtain the engineering design report to be reviewed, parse the report into a chapter title tree and node index. The chapter title tree records the text of each level of title and their hierarchical relationship. The node index records the node range of the main text paragraphs and tables under each title. Identify the report stage and project type of the report to be reviewed based on preset multi-level criteria. The multi-level criteria are sorted in descending order of priority according to user-specified information, cover and compilation instructions information, chapter system characteristics, and main text citation information. Among them, the previous stage information cited in the main text is not used as the basis for determining the current report stage. S2. Based on the report stage and project type, load applicable review rules from the structured review rule base; each review rule in the structured review rule base includes at least a rule identifier, rule type, parent-child relationship, and conditional strength field, and the rule type distinguishes between chapter arrangement rules and content requirement rules; for review rules with conditional strength, first determine the applicability of the rule based on the project type, building type, and construction nature of the report to be reviewed, and filter out inapplicable review rules; S3. The review agent dynamically generates a review task sequence based on the report stage, project type, and review scope, and binds corresponding review skills and deterministic review tools to each review task; the review skills are reusable execution units that encapsulate review methods and processes, and the deterministic review tools include at least document parsing tools, specification retrieval tools, and calculation and review scripts; S4. For each of the chapter arrangement rules, perform exact matching, synonym matching, specific matching, merge matching, and split matching sequentially on the chapter title tree, and do not execute subsequent modes if the previous mode matches successfully; wherein, exact matching is when the title text is the same or only the number is different, synonym matching is when the semantic similarity of the title reaches a preset threshold or hits a preset thesaurus, specific matching is when the chapter title of the report to be reviewed is a subordinate concept of the chapter name of the chapter arrangement rule, merge matching is when multiple chapter arrangement rules are mapped to the same chapter of the report to be reviewed, and split matching is when a chapter arrangement rule is covered by multiple sub-chaps of the report to be reviewed; obtain the chapter arrangement conclusion based on the matching results, and the chapter arrangement conclusion is taken from the conclusion set including matching, partial matching, missing, inapplicable, and undeterminable; when the chapter arrangement conclusion of a certain chapter arrangement rule is missing, skip the content judgment of all content requirement rules associated with it through the parent-child relationship field; S5. For the content requirement rules associated with chapters whose chapter arrangement conclusions are matching or partially matching, the deterministic rule engine first performs parameter extraction, synonym mapping, unit normalization, and tolerance comparison for each review item. For review items that are determined, the content review conclusion is directly obtained. Only the residual review items that the deterministic rule engine cannot determine, together with the original text fragments and rule elements located by the node index, are submitted to the large language model for review to obtain the content review conclusion of the residual review items. The content review conclusion is taken from the conclusion set including compliant, basically compliant, non-compliant, insufficient evidence, and inapplicable. S6. For each chapter's arrangement conclusion and content review conclusion, an evidence record must be forcibly bound. This evidence record must include at least the chapter path and an excerpt of the original text. The standard provisions upon which the review is based are actually obtained by the standard retrieval tool and used as the basis for review. The review basis is categorized into availability levels based on the completeness of the obtained provisions: Level 1 includes provisions whose full text has been obtained, and performs provision-level compliance review; Level 2 includes provisions whose partial text has been obtained, and performs review only on the obtained provisions, declaring in the availability statement that it does not cover the full standard text; Level 3 includes provisions whose text has not been obtained, downgraded to a completeness and self-consistency auxiliary review, and does not output provision-level review conclusions. The availability statement must include at least the required standard identifier, search results, available provisions, the type of this review, and the scope of reviews that cannot be covered. For the review results returned by the large language model, it is verified whether they reference valid node indexes and whether the referenced content is consistent with the original text at the node index. Review results that fail the verification are deemed invalid, and the review conclusion for the corresponding residual review item is reverted to insufficient evidence. For review items that fail to bind evidence records, their review conclusion is determined to be insufficient evidence. S7. Select the judgment criteria from the judgment template set according to the examination type and examination conclusion, perform controlled polishing on the examination opinion text while keeping the data structure unchanged, and generate and output interactive examination opinion data with stability indicators.

[0008] Optionally, in step S1, identifying the reporting stage of the report to be reviewed includes: when the report to be reviewed contains a reference to the results of the previous stage, and the reference is used to respond to, implement, or deepen the results of the previous stage, the current reporting stage is determined to be a subsequent stage of the previous stage; when the report name in the cover, title page, or compilation instructions conflicts with the reference information in the main text, the determination result of the report name shall prevail, and a stage consistency risk warning shall be output in the review results.

[0009] Optionally, in step S2, each review rule in the structured review rule base also includes a standard number, clause source, requirement text, and sorting key fields; during loading, it is sorted according to the sorting key, and a parent-child relationship is established between chapter arrangement rules and content requirement rules through the parent-child relationship field.

[0010] Optionally, in step S3, when the review task execution is abnormal, the review agent processes it according to a preset degradation path: when the document parsing is abnormal, the parsing granularity is reduced and the process is retried; when the standard retrieval has no results, the availability level of the basis for the corresponding review item is set to the third level and subsequent review tasks are continued; when the large language model review does not return a valid result, the review conclusion of the corresponding residual review item is reverted to insufficient evidence.

[0011] Optionally, in step S4, the semantic similarity of the title is calculated using the cosine similarity between the semantic vectors of the title text; the subordinate concept is determined by a preset sub-lexicon or semantic determination model.

[0012] Optionally, in step S5, the tolerance comparison includes: letting v_r be the parameter value extracted from the report to be reviewed, v_e be the benchmark parameter value, ε be the preset relative deviation threshold, and the formula for determining the comparison status m is: The determination of m is as follows: when |v_r − v_e| / |v_e| ≤ ε, m=1, that is, the values ​​are the same; when |v_r − v_e| / |v_e|>ε, m=0, that is, the values ​​are not the same. Where m=1 indicates that the values ​​are the same, and m=0 indicates that the values ​​are different; v_r and v_e are normalized to the same unit according to the preset unit conversion table before comparison; The remaining review items include at least one of the following: parameter alias pending judgment, items with the same parameter appearing in multiple places, table remarks column reference items, and cross-chapter contradiction items.

[0013] Furthermore, in step S5, the determination of parameter consistency review rules also includes: using the parameter set in the engineering characteristic table as the benchmark parameter set; establishing comparison objects for parameters with the same name according to the engineering object, region, or design scheme to which they belong, and performing comparison only between parameters belonging to the same category; when the report to be reviewed contains a recommended scheme and a comparison scheme, using the parameter value of the recommended scheme as the comparison object; when the same parameter takes different values ​​in different positions in the report to be reviewed, outputting the internal contradiction conclusion of the report, and binding evidence records to each position respectively.

[0014] Optionally, in step S6, when a supplementary excerpt of the clause provided by the user is received, the availability level of the basis for the corresponding review is updated from level three to level two, and the review of the corresponding review item is re-executed.

[0015] Optionally, in step S5, the determination performed by the deterministic rule engine further includes calculation review: for content requirement rules containing calculation relationships, input parameters are extracted from the report to be reviewed, the calculation review script is called to recalculate the calculation results, the recalculated value is compared with the result value recorded in the report to be reviewed, and the comparison result and deviation are written into the evidence record; when input parameters are missing, the inferred value is not substituted, the missing parameters are listed and the conclusion of insufficient evidence is output.

[0016] Optionally, in step S7, the selection of the judgment criteria satisfies the following constraints: acknowledgment-type judgments are only used when the review type is professional compliance review and the content review conclusion is compliant; factual statement judgments are used for opinion items with a conclusion of compliance in editorial compliance review; supplementary and improvement-type judgments are used for opinion items with a conclusion of insufficient evidence or needing supplementation; the controlled polishing maintains the review conclusion, maintains the parameters other than the review report and evidence record, and maintains the stable identifier and data structure of the opinion items unchanged.

[0017] Optionally, in step S7, the interactive review opinion data includes an array of opinion items, each opinion item containing at least a stability identifier, review topic, opinion text, and evidence citation; in response to an adoption instruction, rejection instruction, or modification instruction for an opinion item, the review result is updated and the original opinion text and modification record are retained.

[0018] The intelligent review system for engineering design reports based on intelligent agent orchestration described in this invention includes: The document parsing module is used to perform the parsing and recognition operations in step S1 of claim 1; The rule management module is used to store the structured review rule base and perform the rule loading and applicability screening operations in step S2 of claim 1; The intelligent agent orchestration module is used to perform the review task sequence generation and skill and tool binding operations in step S3 of claim 1; The chapter matching module is used to perform the chapter semantic matching operation in step S4 of claim 1; The hierarchical determination module is used to perform the deterministic determination and large language model verification operation in step S5 of claim 1; The evidence management module is used to perform the evidence binding, availability-based grading and downgrading operations in step S6 of claim 1; The opinion generation module is used to perform the controlled polishing and interactive examination opinion data generation operation in step S7 of claim 1.

[0019] The intelligent review device for engineering design reports based on intelligent agent orchestration of the present invention includes a processor and a readable storage medium. The readable storage medium stores a computer program. When the computer program is executed by the processor, it implements each step of the intelligent review method for engineering design reports based on intelligent agent orchestration as described in any one of claims 1 to 11.

[0020] The readable storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0021] The beneficial effects of this invention are: 1. By using the five-pattern matching of chapter semantics and the parent-child association mechanism of the structured review rule base, the same rule base can adapt to engineering design reports with heterogeneous chapter arrangement without customizing rules or templates according to the report source. It can output the arrangement conclusion and missing location at the rule granularity level, rather than the overall compliance score, which solves the problem of difficult implementation of rule determinism and significantly reduces the rule maintenance cost. 2. By mandating the binding of evidence records, reviewing the actual acquisition of constraints, classifying and downgrading based on availability, and verifying the validity of the review results of the large language model, each review conclusion can be traced back to the report chapter path and original text excerpt, eliminating the fabrication of normative clauses and unfounded conclusions, and solving the problem of the unreliability of the review conclusions of the large language model; 3. Through a hierarchical judgment mechanism that prioritizes the deterministic rule engine and only submits residual review items to the large language model for review, as well as a closed loop for rule element decomposition and calculation review, a closed judgment of "rule element - report evidence - review process - conclusion" is achieved, while significantly reducing the amount of large language model calls and computing power consumption. 4. The review rules are stored in the rule base in the form of structured data. When the procedures or review requirements change, only the rule base data needs to be updated to take effect immediately without retraining the model, which solves the problem of high cost of updating the fine-tuning route rules. Attached Figure Description

[0022] Figure 1 is an overall flowchart of an intelligent review method for engineering design reports based on intelligent agent orchestration in an embodiment of the present invention.

[0023] Figure 2 is an architecture diagram of an intelligent review system for engineering design reports based on intelligent agent orchestration in an embodiment of the present invention.

[0024] Figure 3 is a flowchart of the implementation of chapter semantic matching in an embodiment of the present invention.

[0025] Figure 4 is a flowchart illustrating the implementation of layered content determination and residual review item routing in an embodiment of the present invention.

[0026] Figure 5 is a state transition diagram based on availability grading in an embodiment of the present invention.

[0027] Figure 6 is a timing diagram of the dynamic orchestration of the review agent task in an embodiment of the present invention.

[0028] Figure 7 is a schematic diagram of the data structure for evidence recording and based on availability statements in an embodiment of the present invention. Detailed Implementation

[0029] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings. These embodiments are implemented based on the technical solution of the present invention, and detailed implementation methods and specific operation processes are given. However, the scope of protection of the present invention is not limited to the following embodiments.

[0030] In this invention, the review agent refers to a program entity that can dynamically select and invoke review skills and deterministic review tools based on the review request and task execution status, such as an agent program built based on a large language model.

[0031] Example 1: Compliance review of preliminary design report preparation procedures: Referring to Figure 1, this embodiment provides an intelligent review method for engineering design reports based on intelligent agent orchestration, including steps S1 to S7. This embodiment takes the compliance review of the preparation procedure of a preliminary design report for a reservoir project as an example for explanation, and the preparation procedure used is the Preliminary Design Report Preparation Procedure for Water Conservancy and Hydropower Projects.

[0032] S1. Structured parsing and stage identification: The document parsing tool parses the report under review into a chapter title tree and node index. The chapter title tree records the text of each level of headings, the heading level, and numbered headings; the node index records the start and end ranges of nodes for the main text paragraphs and tables under each heading. For example, a report parsed yields 13 first-level heading nodes, 164 main text paragraphs, and 1 table, with the content node range of a certain level of chapter being [20750, 52448]. All subsequent evidence location is performed based on this node index, ensuring that the original text excerpt can be traced back.

[0033] The report stage is identified using a multi-level criterion: the first level is user-specified information; the second level is the report name in the cover, title page, compilation instructions, and file name; the third level is the consistency between the table of contents and the first-level chapter system and the chapter system of a certain stage's compilation procedures; and the fourth level is the information cited in the main text. Among these, "feasibility study report review opinions" and "project proposal approval" appearing in the main text are cited materials from previous stages and are not used as the basis for determining the current report stage. When the main text responds to, implements, or deepens the design of previous stage results, the current report stage is determined to be a subsequent stage of that previous stage. For example, if the report cover is titled "Preliminary Design Report" and the main text cites "feasibility study stage review opinions" in multiple places and explains the implementation status of each item, then the current stage is determined to be the preliminary design stage. If the cover name conflicts with the inference from the main text citations, the cover name prevails, and a stage consistency risk warning is output in the review results. Project types (reservoirs, sluices, pumping stations, irrigation areas, dikes, etc.) are identified from chapters such as comprehensive descriptions, project tasks, and scale.

[0034] S2, rule loading and applicability filtering: The structured review rule base embeds review rules for the three-stage procedure preparation process in the form of structured data tables. Each review rule includes the following fields: rule identifier (rule_id), standard number (standard_code), source (source_ref), requirement text (requirement_text), rule type (requirement_type, which can be either heading or content), parent-child relationship (parent_rule_id), requirement_strength (mandatory or conditional), and sort key (sort_key). During loading, the rule set of the unique master procedure is filtered according to the reporting stage, sorted by the sort key, and the parent-child relationship is established between the heading rules and their content requirement rules through the parent-child relationship field. The field structure and loading process of the structured review rule base are shown in Figure 2.

[0035] For rules with conditional strength, applicability is determined first: for example, the "Irrigation and Drainage" section requires that the rule only applies to irrigation district projects, so it is determined that the rule does not apply to reservoir projects and is filtered out; the applicability of the rules related to "Mechanical and Electrical and Metal Structures" is determined based on whether the project includes drainage or power generation structures. If the applicability conditions cannot be determined, an "unable to determine" conclusion is output, and the rule is not considered missing.

[0036] S3, Dynamic scheduling of review tasks: Referring to Figure 6, after receiving a review request, the review agent dynamically generates a review task sequence based on the report stage, project type, and user-specified review scope (all chapters or specified chapters) identified by S1, and binds review skills and deterministic review tools to each task. Review skills are reusable execution units that encapsulate review methodologies and processes; in this embodiment, they include skills for drafting procedural compliance reviews, parameter consistency reviews, technical compliance reviews, and review opinion polishing. Deterministic review tools include document parsing tools, specification retrieval tools, and calculation and review scripts. In one implementation, the review agent performs dynamic orchestration in the following manner: (1) Task sequence generation: The review agent parses the review type set and review scope from the review request, searches and matches the skill registry according to the capability description, required input and output of each review skill declaration, and selects the matching review skill; each review skill declares its input dependency (for example, the parameter consistency review skill declaration requires the engineering characteristic table as input, and the technical compliance review skill declaration requires the specification clause as input), and the review agent establishes a task dependency graph among the selected skills accordingly, marks the review tasks without input dependency as parallel tasks, marks the review tasks with input dependency or output dependency as serial tasks, and arranges the execution order according to the dependency order of "document parsing → basis acquisition → item-by-item review → opinion generation"; (2) Skill and tool binding: Bind a primary review skill and the deterministic review tools required for each review task. For example, the procedure compliance review task is bound to the procedure compliance review skill, document parsing tool and structured review rule base; the technical compliance review task is bound to the technical compliance review skill, specification retrieval tool and calculation review script. (3) Failure retry and downgrade path: When a review task is executed abnormally, the review agent will handle it according to the preset strategy without interrupting the overall process - when the document parsing is abnormal, it will retry once and reduce the parsing granularity (such as reducing the parsing from the whole chapter to the parsing by node range); when the standard retrieval has no results, the availability level of the basis of the review item will be reduced to the third level and the subsequent tasks will continue; when the large language model review does not return a valid result, the corresponding residual review item will be rolled back to the conclusion of insufficient evidence; (4) Parallel scheduling: Independent review tasks (e.g., parameter consistency review task and procedure compliance review task) share the same document parsing result and are executed in parallel. The opinion generation task depends on the output of all review tasks and is placed at the end of the task sequence.

[0037] For example, the task sequence in this embodiment is: document parsing task (bound to document parsing tool) → parallel review of compilation compliance and parameter consistency (bound to corresponding review skills and rule base, engineering characteristic table respectively) → comment generation task (bound to review comment polishing skill).

[0038] S4, Chapter Semantic Matching: Referring to Figure 3, for each chapter arrangement rule, five patterns are matched in the chapter title tree in the following order: the subsequent patterns are terminated when the previous pattern is successfully matched: (1) Exact match: The title text is the same or only the number is different, for example, the rule section "Engineering Geology" and the report section "3 Engineering Geology"; (2) Synonym matching: The semantic similarity of the titles reaches a preset threshold or matches a preset synonym list, for example, the rule section "Engineering Layout and Buildings" and the report section "Overall Engineering Layout and Main Buildings". In one implementation, the semantic similarity is calculated using the cosine similarity of the semantic vectors of the title text, and the threshold is taken as an empirical value; in another implementation, the candidate title pairs are determined to be synonyms by a large language model; (3) Specific matching: The report chapter title is a subordinate concept of the rule chapter name, such as the rule chapter "water-retaining structure" and the report chapter "roller-compacted concrete gravity dam"; the subordinate concept can be determined based on the preset superordinate and subordinate term list, or by the semantic determination model; (4) Merging and matching: Multiple rule chapters are mapped to the same chapter of the report. For example, the contents of the two chapters of "Construction Diversion" and "Material Yard Planning" of the rule are explained in the "Construction Organization Design" chapter of the report. (5) Split matching: A rule chapter is covered by multiple sub-chapters in the report.

[0039] The chapter arrangement conclusions are derived from {matching, partial matching, missing, inapplicable, and cannot be determined}. For chapter arrangement rules with missing conclusions, the judgment of all its child content requirement rules is skipped based on parent-child relationships to avoid invalid calculations and forced inferences.

[0040] S5, Layered content determination: (See Example 2 for details) S6, Evidence Chain Binding: Each chapter arrangement conclusion and content review conclusion is forcibly linked to evidence records, which include the chapter path (e.g., "6 Engineering Layout and Buildings > 6.3 Water Retaining Structures") and excerpts from the original text. Externally output review bases are uniformly expressed as the standard number, code name, and clause number of the adopted code; internal data identifiers such as rule base file paths are not used as review base outputs.

[0041] S7, Controlled Opinion Generation: The review comments text selects the appropriate terminology from the terminology template set according to the review type and conclusion: For comments with a conclusion of compliance in the compilation and regulation compliance review, factual statements such as "The report already includes…" or "The report's content meets… requirements" are used, without using approval-type terms; for comments with a conclusion of needing supplementation, terms such as "Needs to be supplemented…" or "Needs to be supplemented and improved…" are used, specifying the chapters to be added or adjusted, and the content points to be improved. Controlled polishing only replaces the main text of the comments, without changing the review conclusion or adding parameters beyond the report and evidence records, maintaining the stability identifiers (e.g., op-001, op-002), subject, evidence citations, and other data structures unchanged. The final output is interactive review comment data (combination of comment items) carrying stability identifiers. Reviewers perform adoption, rejection, or modification operations on each comment, and the system updates the review results accordingly while retaining the original comments and modification records; the generated interactive review comment data structure is shown in Figure 7.

[0042] Example 2, Design Report Parameter Consistency Review: This embodiment takes the consistency review of the information technology-specific description parameters of a large-scale water conservancy project (Class I large (1) type, whose main structures include roller-compacted concrete gravity dam, flood discharge and sediment flushing structures and water diversion and power generation structures) as an example to illustrate the layered content judgment mechanism of step S5. The review benchmark is the engineering characteristic table of the project (3 data tables, 212 parameter records), and the report to be reviewed is the information technology-specific chapter of the preliminary design report of the project.

[0043] Referring to Figure 4, the layered content determination is performed according to the following levels: First layer: Deterministic rule engine judgment.

[0044] The deterministic rule engine performs parameter extraction, synonym mapping, unit normalization, and tolerance comparison for each baseline parameter: (1) Synonym mapping: Match parameter name variants according to the preset synonym list, such as "installed capacity", "total installed capacity", and "total installed capacity" are mapped to the same parameter; (2) Unit normalization: Normalize the reported value and the benchmark value to the same unit according to the preset unit conversion table. For example, "2100MW" and "2.1 million kilowatts" are equal after normalization. (3) Tolerance Comparison: Let v_r be the parameter value extracted from the report, v_e be the benchmark parameter value in the engineering characteristic table, and ε be the preset relative deviation threshold. The formula for determining the comparison status m is: The determination of m is as follows: when |v_r − v_e| / |v_e| ≤ ε, m=1 (consistent value); when |v_r − v_e| / |v_e|>ε, m=0 (inconsistent value).

[0045] Here, m=1 indicates that the values ​​are consistent, and m=0 indicates that the values ​​are inconsistent. For example, if ε is set to 0.01%, the reported value of 489,900 km² has a relative deviation of approximately 0.001% after normalization compared to the baseline value of 489,948 km², so m=1 is determined. If a certain section of the report records a dam crest length of 900m, while the baseline value is 960m, the relative deviation is 6.3%, so m=0 is determined, and the conclusion of inconsistency is output.

[0046] Parameter attribution disambiguation: For parameters with the same name, establish comparison pairs according to the engineering object, area, or design scheme to which they belong, and only compare parameters belonging to the same scheme. For example, if two sluice gates each have a "gate top elevation" parameter, compare them separately with the corresponding sluice gate records in the characteristic table, without overlap; when the report contains recommended and alternative schemes, use the parameter value of the recommended scheme as the comparison object. When the same parameter has different values ​​in different locations in the report (e.g., "maximum dam height" is recorded as 219.4m in the project background section and 219.5m in the overall project layout section), output the conclusion of internal contradiction in the report, and bind evidence records to the two locations respectively.

[0047] The second layer involves submitting residual review items to the large language model for verification. Only the following four types of residual review items that the deterministic rule engine cannot determine should be submitted to the large language model for review: parameter alias items to be judged (e.g., whether "bottom hole" in the report corresponds to "sand discharge bottom hole" in the characteristic table), items with the same parameter appearing in multiple places (checking each instance to ensure consistency with the baseline), items referenced in the table remarks column (informal parameter references in the equipment table remarks column, such as the remarks "normal reservoir water level 626 meters" being inconsistent with the baseline value of 627 meters), and cross-chapter contradictions. When submitting, the prompt message should include the original text fragment located by the node index and the rule element, and require the review result to cite the node index on which it is based. Review results that do not cite a valid node index or whose cited content is inconsistent with the original text at the node index will be deemed invalid, and the corresponding residual review item will be reverted to a conclusion of insufficient evidence.

[0048] For example, of the 212 baseline parameters in this embodiment, approximately 16 actually appeared in the pending IT special report, while the remaining approximately 196 were not mentioned in the report due to the nature of the document (IT special description). Of the approximately 16 parameters that actually appeared, approximately 11 were directly determined by the deterministic rule engine through synonym mapping, unit normalization, and tolerance comparison. Only about 5 parameters (involving parameter aliases to be determined, references in the device table remarks column, and verification of multiple occurrences of the same parameter) entered the large language model for review. Compared to submitting each parameter that actually appeared in the report to the large language model, this embodiment reduces the number of large language model calls by about two-thirds, and the approximately 196 unmentioned parameters are marked as unmentioned by deterministic determination without calling the large language model.

[0049] Third layer, summary of conclusions: The output of each parameter is consistent, inconsistent, unmentioned, or requires review. Unmentioned parameters that can be explained by the nature of the document (such as parameters related to construction period not covered in the information technology special instructions) are explained in the review results as normal non-coverage and are not judged as omissions.

[0050] Example 3: Technical Compliance Review and Degradation of Availability of Basis: This embodiment takes the review of the crest elevation of a roller-compacted concrete gravity dam as an example to illustrate the rule element decomposition, calculation review closed loop (the calculation review script bound in step S3) and the availability-based graded degradation mechanism of step S6.

[0051] Rule element decomposition and evidence location. The review rule is the dam crest elevation clause in the concrete gravity dam design code: the dam crest elevation should be higher than the highest static water level of the reservoir. The difference between the wave wall crest elevation and the normal storage water level or check flood level is calculated using Delta h = h_1% + h_z + h_c, and the higher of the two water level conditions is taken. The review skills decompose this rule into verifiable elements: applicable objects (dam level, water level conditions), mandatory actions (should be higher than), judgment indicators (dam crest elevation, wave wall crest elevation), calculation relationships (formula and safety height increase table lookup), and evidence requirements (parameter source, calculation process, result comparison). Subsequently, the report locates evidence according to keywords and chapter clues, records the chapter path and original text excerpts, and confirms that the evidence belongs to the recommended scheme.

[0052] Calculate and verify the closed loop. Extract input parameters from the report (normal water level, check flood level, wind speed, effective draft, dam class, safety height increase h_c, etc.), call the calculation and verification script to recalculate h_1%, h_z, Delta h and the control elevation of the wave wall top, compare the recalculated values ​​with the values ​​recorded in the report, and write the comparison results and deviations into the evidence record; when input parameters are missing, do not substitute the inferred values, list the missing parameters and output the conclusion of insufficient evidence.

[0053] Based on availability classification and downgrading. See Figure 5. The normative provisions upon which the review is based must be actually retrieved from the standard reference database using a normative retrieval tool before they can be cited. Three availability levels are determined based on the completeness of the retrieved information: (1) Level 1 (Article Level): The full text of the article has been obtained and read. Article-level compliance review is performed. The conclusion can be: compliant, basically compliant, non-compliant, insufficient evidence, or inapplicable. Each conclusion is marked with the standard number, article number, and key points of the original text. (2) Second level (partial clause level): Only a portion of the clauses or a clause excerpt provided by the user is obtained, and the review is performed only on the clauses that have been obtained, and the full text of the standard is declared not to be covered in the availability statement; (3) Level 3 (Auxiliary Review Level): If the standard is not in the library, there are no available slices, or the retrieval fails, and the user does not supplement the clause, it will be automatically downgraded to auxiliary review for completeness and self-consistency. No clause-level review conclusions will be output, and industry common sense or model memory will not be used as the basis for the clause. When the user provides supplementary clause extracts, the availability level of the basis for the corresponding review will be updated from Level 3 to Level 2, and the review of the corresponding review items will be re-executed.

[0054] Each review output includes a statement of availability, which includes: the required standard identifier, search results (obtained / partially obtained / not obtained), the scope of available provisions, the type of review (provisional review / partial review / auxiliary review), and the scope of review that cannot be covered. The data structure of the evidence record and the statement of availability is shown in Figure 7. This mechanism eliminates the possibility of fabricating normative provisions using large language models as the basis for review from the data link.

[0055] This embodiment provides an intelligent review system for engineering design reports based on intelligent agent orchestration, as shown in Figure 2. It includes: a document parsing module for performing parsing and recognition operations in step S1; a rule management module for storing a structured review rule base and performing rule loading and applicability screening operations in step S2; an intelligent agent orchestration module for performing review task sequence generation and skill / tool ​​binding operations in step S3; a chapter matching module for performing chapter semantic matching operations in step S4; a hierarchical judgment module for performing deterministic judgment and large language model verification operations in step S5; an evidence management module for performing evidence binding, usability-based grading, and downgrading operations in step S6; and an opinion generation module for performing controlled polishing and interactive review opinion data generation operations in step S7. The system is organized into a four-layer architecture: intelligent agent layer, skill layer, tool layer, and data layer. The intelligent agent layer is responsible for dynamic task orchestration; the skill layer carries various review skills; the tool layer provides deterministic capabilities such as document parsing, standard retrieval, and calculation review; and the data layer includes a structured review rule base, a thesaurus, a unit conversion table, a judgment term template set, and a standard specification base.

[0056] This embodiment also provides an intelligent review device for engineering design reports based on intelligent agent orchestration, including a processor and a readable storage medium. The readable storage medium stores a computer program, which, when executed by the processor, implements the steps of the intelligent review method for engineering design reports based on intelligent agent orchestration as described in any one of claims 1 to 11.

[0057] It is understood that if the method in this embodiment is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium and executes all or part of the steps of this embodiment.

[0058] Finally, it should be emphasized that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still make modifications to the technical solutions described in the foregoing embodiments without creative effort, or make equivalent substitutions for some of the technical features. Therefore, any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent review of engineering design reports based on intelligent agent orchestration, characterized in that, Includes the following steps: S1. Obtain the engineering design report to be reviewed, parse the report to be reviewed into a chapter title tree and node index, and identify the report stage and engineering type of the report to be reviewed according to the preset multi-level criteria. The multi-level criteria are sorted in descending order of priority according to user-specified information, cover and compilation instructions information, chapter system characteristics, and text citation information. The previous stage information cited in the text is not used as the basis for determining the current report stage. S2. Based on the report stage and project type, load applicable review rules from the structured review rule base; each review rule in the structured review rule base includes at least a rule identifier, rule type, parent-child relationship, and conditional strength field, and the rule type distinguishes between chapter arrangement rules and content requirement rules; for review rules with conditional strength, first determine the applicability of the rule based on the project type, building type, and construction nature of the report to be reviewed, and filter out inapplicable review rules; S3. The review agent dynamically generates a review task sequence based on the report stage, project type, and review scope, and binds corresponding review skills and deterministic review tools to each review task; the review skills are reusable execution units that encapsulate review methods and processes, and the deterministic review tools include at least document parsing tools, specification retrieval tools, and calculation and review scripts; S4. For each of the chapter arrangement rules, perform exact matching, synonym matching, specific matching, merge matching, and split matching sequentially on the chapter title tree, and do not execute subsequent modes if the previous mode matches successfully; obtain the chapter arrangement conclusion based on the matching results, and the chapter arrangement conclusion is taken from the conclusion set including matching, partial matching, missing, inapplicable, and undeterminable; when the chapter arrangement conclusion of a certain chapter arrangement rule is missing, skip the content determination of all content requirement rules associated with it through the parent-child relationship field; S5. For the content requirement rules associated with chapters whose chapter arrangement conclusion is a match or partial match, the deterministic rule engine first performs parameter extraction, synonym mapping, unit normalization and tolerance comparison for each review item. For the review items that are determined, the content review conclusion is directly obtained. Only the residual review items that the deterministic rule engine cannot determine, together with the original text fragments and rule elements located by the node index, are submitted to the large language model for review to obtain the content review conclusion of the residual review items. S6. For each chapter's arrangement conclusion and content review conclusion, an evidence record is forcibly bound. This evidence record includes at least the chapter path and an excerpt from the original text. The normative provisions upon which the review is based are actually obtained by the normative retrieval tool and used as the basis for review. The review basis is categorized into availability levels based on the completeness of the obtained provisions: Level 1 includes provisions whose full text has been obtained, and a provision-level compliance review is performed; Level 2 includes provisions whose only part has been obtained, and a review is performed only on the obtained provisions, with a declaration in the availability statement stating that the full text of the standard is not covered; Level 3 includes provisions whose not obtained, which are downgraded to a completeness and self-consistency auxiliary review and no provision-level review conclusion is output. For the review results returned by the large language model, it is verified whether they reference valid node indexes and whether the referenced content is consistent with the original text at the node index. Review results that fail the verification are deemed invalid, and the review conclusion for the corresponding residual review item is reverted to insufficient evidence. For review items that fail to be linked to evidence records, the review conclusion shall be determined as insufficient evidence; S7. Select the judgment criteria from the judgment template set according to the examination type and examination conclusion, perform controlled polishing on the examination opinion text while keeping the data structure unchanged, and generate and output interactive examination opinion data with stability indicators.

2. The intelligent review method for engineering design reports based on intelligent agent orchestration according to claim 1, characterized in that, In step S1, identifying the reporting stage of the report to be reviewed includes: when the report to be reviewed contains a reference to the results of the previous stage, and the reference is used to respond to, implement or deepen the results of the previous stage, the current reporting stage is determined to be the subsequent stage of the previous stage; when the report name in the cover, title page or compilation instructions conflicts with the reference information in the main text, the determination result of the report name shall prevail, and a stage consistency risk warning shall be output in the review results.

3. The intelligent review method for engineering design reports based on intelligent agent orchestration according to claim 1, characterized in that, In step S2, each review rule in the structured review rule base also includes a standard number, clause source, requirement text, and sorting key fields; during loading, it is sorted according to the sorting key, and a parent-child relationship is established between chapter arrangement rules and content requirement rules through the parent-child relationship field.

4. The intelligent review method for engineering design reports based on intelligent agent orchestration according to claim 1, characterized in that, In step S3, when the review task execution is abnormal, the review agent processes it according to a preset degradation path: when the document parsing is abnormal, the parsing granularity is reduced and the process is retried; when the standard retrieval has no results, the availability level of the basis for the corresponding review item is set to the third level and subsequent review tasks are continued; when the large language model review does not return a valid result, the review conclusion of the corresponding residual review item is reverted to insufficient evidence.

5. The intelligent review method for engineering design reports based on intelligent agent orchestration according to claim 1, characterized in that, In step S4, the semantic similarity of the title is calculated using the cosine similarity between the semantic vectors of the title text; the subordinate concept is determined by a preset subordinate and subordinate word list or semantic determination model.

6. The intelligent review method for engineering design reports based on intelligent agent orchestration according to claim 1, characterized in that, In step S5, the tolerance comparison includes: letting v_r be the parameter value extracted from the report to be reviewed, v_e be the benchmark parameter value, ε be the preset relative deviation threshold, and the formula for determining the comparison status m is: The determination of m is as follows: when |v_r − v_e| / |v_e| ≤ ε, m=1, that is, the values ​​are consistent; when |v_r − v_e| / |v_e| > ε, m=0, that is, the values ​​are inconsistent. Where m=1 indicates that the values ​​are the same, and m=0 indicates that the values ​​are different; v_r and v_e are normalized to the same unit according to the preset unit conversion table before comparison; The remaining review items include at least one of the following: parameter alias pending judgment, items with the same parameter appearing in multiple places, table remarks column reference items, and cross-chapter contradiction items.

7. The intelligent review method for engineering design reports based on intelligent agent orchestration according to claim 1 or 5, characterized in that, In step S5, the determination of parameter consistency review rules also includes: using the parameter set in the engineering characteristic table as the benchmark parameter set; establishing comparison objects for parameters with the same name according to the engineering object, region or design scheme to which they belong, and performing comparison only between parameters belonging to the same category; when the report to be reviewed contains a recommended scheme and a comparison scheme, the parameter value of the recommended scheme is used as the comparison object; when the same parameter takes different values ​​in different positions in the report to be reviewed, outputting the internal contradiction conclusion of the report, and binding evidence records to each position respectively.

8. The intelligent review method for engineering design reports based on intelligent agent orchestration according to claim 1, characterized in that, In step S6, when a supplementary excerpt of the clause is received from the user, the availability level of the corresponding review basis is updated from level three to level two, and the review of the corresponding review item is re-executed.

9. The intelligent review method for engineering design reports based on intelligent agent orchestration according to claim 1, characterized in that, In step S5, the determination performed by the deterministic rule engine also includes calculation review: for content requirement rules containing calculation relationships, input parameters are extracted from the report to be reviewed, the calculation review script is called to recalculate the calculation results, the recalculated values ​​are compared with the result values ​​recorded in the report to be reviewed, and the comparison results and deviations are written into the evidence record; when input parameters are missing, the inferred values ​​are not substituted, the missing parameters are listed and the conclusion of insufficient evidence is output.

10. The intelligent review method for engineering design reports based on intelligent agent orchestration according to claim 1, characterized in that, In step S7, the selection of the judgment criteria satisfies the following constraints: acknowledgment judgments are only used when the review type is professional compliance review and the content review conclusion is compliant; factual statement judgments are used for opinion items with a compliant conclusion in editorial compliance review; supplementary and improvement judgments are used for opinion items with insufficient evidence or requiring supplementation; the controlled polishing maintains the review conclusion, maintains parameters other than the review report and evidence record, and maintains the stable identifier and data structure of the opinion items unchanged.

11. The intelligent review method for engineering design reports based on intelligent agent orchestration according to claim 1, characterized in that, In step S7, the interactive review opinion data includes an array of opinion items, each opinion item containing at least a stability identifier, review topic, opinion text, and evidence citation; in response to an adoption instruction, rejection instruction, or modification instruction for an opinion item, the review result is updated and the original opinion text and modification record are retained.

12. An intelligent review system for engineering design reports based on intelligent agent orchestration, characterized in that, include: The document parsing module is used to perform the parsing and recognition operations in step S1 of claim 1; The rule management module is used to store the structured review rule base and perform the rule loading and applicability screening operations in step S2 of claim 1; The intelligent agent orchestration module is used to perform the review task sequence generation and skill and tool binding operations in step S3 of claim 1; The chapter matching module is used to perform the chapter semantic matching operation in step S4 of claim 1; The hierarchical determination module is used to perform the deterministic determination and large language model verification operation in step S5 of claim 1; The evidence management module is used to perform the evidence binding, availability-based grading and downgrading operations in step S6 of claim 1; The opinion generation module is used to perform the controlled polishing and interactive examination opinion data generation operation in step S7 of claim 1.

13. An intelligent review device for engineering design reports based on intelligent agent orchestration, characterized in that, The method includes a processor and a readable storage medium storing a computer program that, when executed by the processor, implements the steps of the intelligent review method for engineering design reports based on intelligent agent orchestration as described in any one of claims 1 to 11.

Images