Knowledge graph and GraphRAG fused national security policy credible question and answer method

Abstract

The invention provides a national security policy credible question and answer method fusing a knowledge graph and GraphRAG, and relates to the technical field of artificial intelligence, and the method comprises the steps: extracting policy keywords and relationships from document data, constructing a structured knowledge network containing nodes and edges, and carrying out the query of the structured knowledge network; establishing an immutable metadata field set with fields such as an original ownership anchor point and an intention category for each document node, and generating a semantic feedback connection edge based on a trigger statement; performing problem-driven subgraph retrieval based on GraphRAG, and performing hop-by-hop semantic closed-loop verification on candidate subgraphs so as to only allow closed loops conforming to semantic situation circulation to enter subsequent processing; performing independent fingerprint and complementarity screening on the blocked set to introduce a controlled secondary evidence and record a discount and audit chain; performing intention and time causal verification on the admission evidence, and determining to output or trigger manual recheck according to a binding state; and generating structured output and attaching a complete auditing pointer to ensure traceability.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a credible question-answering method for national security policy that integrates knowledge graphs and GraphRAG. Background Technology

[0002] With the rapid growth of information processing needs, various data systems for public governance are constantly expanding in scale and complexity; how to achieve stable and consistent knowledge acquisition and interpretation in the context of parallel flow of multi-source data has become an increasingly important research focus; related technologies are continuously evolving towards structured expression and traceable management, and the trend is becoming increasingly stronger. Existing information question answering systems mostly rely on keyword matching and topic clustering to integrate data sources; they typically organize text through index structures and use statistical relevance to filter candidate content; in complex policy scenarios, multi-level rule templates are also used to assist in judgment to improve consistency, but overall they still tend to be linear in processing to enhance scalability; The shortcomings of existing technologies are: they often fail to fully identify deep connections when dealing with the correlation logic of materials from multiple sources; their ability to capture semantic clues across documents is limited and they lack a unified discrimination mechanism at the temporal and intentional reasoning levels; they lack protection against AI illusions and lack an effective mechanism for tracing information sources. Summary of the Invention

[0003] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a trustworthy question-answering method for national security policies that integrates knowledge graphs and GraphRAG. This method addresses the problems mentioned in the background, such as the difficulty in fully identifying deep connections when processing the correlation logic of materials from multiple sources; limited ability to capture semantic clues across documents; lack of a unified discrimination mechanism at the temporal and intentional reasoning levels; lack of protection against AI illusions; and lack of an effective mechanism for tracing information sources.

[0004] (II) Technical Solution To achieve the above objectives, this invention provides the following technical solution: a trustworthy question-answering method for national security policy that integrates knowledge graphs and GraphRAG, comprising the following steps: S1. Construct a knowledge graph: Extract policy keywords and relationships between keywords from the collected document data; the policy keywords include policy subject, relevant fields, main content, and time nodes; the relationships between keywords include formulation relationships, implementation relationships, and constraint relationships; construct graph nodes based on the extracted policy keywords, and construct graph edges based on the extracted relationships between keywords to form a structured knowledge network; establish an immutable metadata field set for each document node in the structured knowledge network, containing the following fields: original ownership anchor, intention category, semantic posture label, version identifier, publication time, citation list, audit pointer, and binding status; and generate semantic feedback connection edges in the structured knowledge network based on rule-based annotation trigger statements; S2, Question-driven subgraph retrieval based on GraphRAG: Semantic parsing of user questions to extract search keywords; using the GraphRAG retrieval engine to match the search keywords with the structured knowledge network to locate relevant graph nodes and connecting edges; the retrieval engine returns candidate nodes containing the immutable metadata field set, sorts them according to the reference relationship between nodes, assembles them into candidate subgraphs, and sends them to the semantic loop closure verifier. S3, Perform semantic loop closure verification: In the semantic loop closure verifier, identify closed-loop paths from the original ownership anchor point and back through semantic feedback connection edges from the candidate subgraph; allow the closed-loop path to enter subsequent processing only if each hop of at least one of the closed-loop paths is a semantic feedback connection edge and the semantic situation label sequence on the path conforms to the preset allowed flow sequence; otherwise, reject the candidate subgraph, classify its nodes into the blocked node set, and record the blocking reason; S4, Implement controlled secondary corroboration introduction: When the candidate subgraph is rejected, trigger the independence fingerprint determination of the blocked node set, and select the minimum complementary combination from the blocked node set according to the reverse confidence loop closure rule; only when the minimum complementary combination satisfies all rule truth values ​​in the independence fingerprint flag and there is no circular dependency, it is introduced in the form of controlled secondary corroboration combination, and the discount label and audit chain are recorded; S5, Execute Verification: For any admitted closed-loop path or the introduced controlled secondary supporting evidence combination, perform an intention causal arrangement verification based on intention category and publication time series; if the binding status of the starting node of the closed-loop path is true and the publication time series is consistent with the intention causal arrangement, or if the binding status of all documents in the controlled secondary supporting evidence combination is true and the publication time series is consistent, then use it to generate the answer; otherwise, mark it as unbound or restricted, and trigger a manual review or calibration process; S6, Generate Trusted Output: Generate a structured output that includes the closed-loop path node reference, the independence fingerprint judgment summary, the intention causal arrangement verification result, and the complete audit pointer; the structured output is generated primarily based on authoritative documents with a true binding status; when the output contains the unbound or restricted status items, a calibration or manual review process is triggered.

[0005] Preferably, in step S1, the specific implementation of constructing the knowledge graph includes: extracting policy keywords and relationships between keywords from the collected document data; the policy keywords include policy subject, involved field, main content, and time node; the relationships between keywords include formulation relationship, implementation relationship, and constraint relationship. Graph nodes are constructed based on the extracted policy keywords, and graph edges are constructed based on the extracted relationships between keywords, forming a structured knowledge network; for the original document to be entered into the database, the validity statement at the beginning and in the body of the document is parsed, and the original ownership anchor field is determined and written according to the institutional list rules based on the list of authoritative institution names. Trigger phrases based on the policy keyword list are matched against the entire original document, and the intention category field is written accordingly. An immutable content hash is established for the original document, and the immutable content hash is stored together with the audit pointer to ensure the traceability of subsequent rule determinations. According to the binding rules based on the authoritative institution name and validity statement, the binding status of the document node is set and written to the binding status field. For each document node in the structured knowledge network, an immutable metadata field set is established, containing the following fields: original ownership anchor, intent category, semantic state tag, version identifier, publication time, citation list, audit pointer, and binding status. Semantic feedback connection edges are generated in the structured knowledge network based on rule-based annotation trigger statements. All field writes are completed using immutable transactions, and field change history and trigger records are recorded in the audit storage. This process ensures the integrity and traceability of metadata for subsequent subgraph retrieval and closed-loop verification.

[0006] Preferably, in step S1, a trigger phrase matching process is executed based on the policy keyword list. First, the original document text to be entered into the database is segmented and normalized, preserving the original word order and removing control characters and whitespace noise. Following the longest phrase priority and unambiguous phrase priority rules, a trigger statement record is generated for each match. The trigger statement record includes the start and end positions of the trigger text, the original trigger text, the intent category, the candidate priority, the trigger timestamp, the trigger source document identifier, and the trigger unique identifier. A trigger statement hash is generated for the trigger text, and the trigger statement hash is associated with the immutable content hash of the original document and the audit pointer. The trigger statement record is mapped to a semantic feedback connection edge, which consists of a connection unique identifier, a source original ownership anchor node, and... The target document node and connection type are labeled as semantic feedback and carry trigger statement references, semantic situation tags, version identifiers, release times, reference lists, and audit pointer fields. The semantic feedback connection edges and trigger statement records are written into the structured knowledge network using immutable transactions, and trigger record fragments are generated in the audit storage for subsequent closed-loop path regression verification. When multiple triggers occur at a single text location, an ordered feedback chain is established sequentially according to the text appearance order and candidate priority, and conflict resolution reasons and final selection criteria are written into the trigger record. All trigger writing steps generate immutable audit entries, and the trigger statement fragments, trigger rule chains, and trigger timestamps are recorded in the audit pointer to ensure that the semantic closed-loop verifier can perform path regression verification based on the original ownership anchor point.

[0007] Preferably, in step S2, the specific implementation of question-driven subgraph retrieval based on GraphRAG includes: performing semantic parsing on the user question; the parser extracts intent, entity, and predicate relationships according to word segmentation semantic units; generating a retrieval expression based on the intent entity; the retrieval expression includes a keyword vector sequence, entity identifier, and time constraint; the retrieval expression generation uses embedded indexes and Boolean constraints; and writing the retrieval metadata into a temporary retrieval session record for auditing; using the retrieval expression to drive the GraphRAG retrieval engine to retrieve candidate nodes in the structured knowledge network; the retrieval process includes two steps: vector similarity retrieval and hierarchical expansion based on the reference list field; firstly, obtaining an initial candidate set through vector similarity; and then expanding the set of mutually referenced nodes outward based on the reference list field of each candidate node. The process involves: recording the reference direction and location; extracting the immutable metadata fields for each candidate node returned by GraphRAG: original ownership anchor, intention category, semantic state label, version identifier, release time, reference list, audit pointer, and binding status; and performing a stable sorting of candidate nodes according to the reference relationship between nodes, i.e., the reference order and reference strength in the reference list. The sorting stability adopts a topological sorting and similarity weighting rule, and a sorting evidence fragment is written. The sorted nodes are then assembled into a candidate subgraph data structure according to the reference relationship. The candidate subgraph includes a node list, an edge list, node order, an immutable metadata snapshot, and a reference link proof fragment. The candidate subgraph and the immutable metadata are transmitted to the semantic loop closure verifier through an interface, and the transmission audit pointer is recorded for subsequent loop closure regression verification.

[0008] Preferably, in step S3, the specific implementation of semantic loop closure verification in the semantic loop closure verifier includes: performing a hop-by-hop judgment operation on the incoming candidate subgraph. First, the semantic loop closure verifier identifies the connection type of each side in the candidate subgraph and matches the semantic feedback connection rule. The semantic feedback connection rule is based on a precise match between the trigger statement field and the intention category field, and requires that the text trigger position be consistent with the intention category. When the hop-by-hop judgment passes, the semantic loop closure verifier extracts the semantic situation label sequence of each node according to the node order of the candidate subgraph, and verifies each item according to the state transition rule of the allowed flow sequence. The allowed flow sequence is composed of ordered state transition rules between semantic situation labels, and requires that the sequence completely covers the state transition. For each hop of the closed-loop path; when at least one closed-loop path is identified where each hop is a semantic feedback connection and the semantic situation label sequence conforms to the allowed flow sequence, the semantic closed-loop verifier outputs the closed-loop path, records the evidence fragments and trigger rule chains that form the closed loop, the evidence fragments include trigger statement fragments, original ownership anchors, reference positions, inter-node reference relationships, version identifiers, release times and audit pointers, and verifies the consistency of version identifiers; if the hop-by-hop judgment or semantic situation label sequence verification fails, the candidate subgraph is rejected, and the blocking reason and matching rule summary are written into the audit pointer for subsequent independence fingerprint judgment; all judgment and writing operations are saved in an immutable record manner, and audit entries are generated for subsequent traceability.

[0009] Preferably, in step S4, the specific implementation of the controlled secondary corroboration introduction includes: performing an independence fingerprint determination process on the blocked node set; firstly, for each document node in the set, the original document corresponding to it is read item by item according to the list of immutable metadata fields, including the original ownership anchor field, intention category field, semantic state tag field, version identifier field, publication time field, citation list field, immutable content hash field, audit pointer field, and binding status field, and each field is compared word by word with the corresponding position in the source document to generate a true / false flag. The immutable content hash comparison adopts byte-by-byte hash matching and hash check code consistency determination; the original ownership anchor comparison adopts the longest name in the organization list as the priority match and records the matching position and matching weight; the intention category comparison requires that the trigger phrase position and the intention category sequence completely correspond and records the candidate priority; the publication time comparison requires that the timestamp and the version identifier are consistent and written into the time check summary; secondly, the true / false flag set of each document is constructed into an independence fingerprint vector, and the field coverage and the reasons for inconsistencies are marked in the fingerprint vector. The process involves searching for the minimum complementary combination within the blocked node set again, based on the complementarity determination rule. This rule requires that the union of the true / false flags of the selected combination must cover eight essential fields: original ownership anchor, intention category, semantic state label, version identifier, publication time, citation list, audit pointer, and binding status. The combination search employs exhaustive pruning and heuristic prioritization to minimize the number of documents and records the search path. Subsequently, the citation cross-references between the minimum complementary combinations are evaluated to construct a directed citation graph, and directed graph loop detection is performed for circular dependency determination. If the minimum complementary combination satisfies all truth values ​​in the independence fingerprint flag and does not have circular dependencies, the combination is introduced in a controlled secondary verification manner, and an auditable trigger record and discount label are generated upon introduction. The discount label content, including the discount reason, discount amount, source document identifier, and timestamp, is written into the audit chain. In the case of circular dependencies or failure of the complementarity determination, no combination may be introduced, and the blocking reason and the search summary of the inability to generate the minimum complementary combination are written into the audit pointer for subsequent self-calibration or manual review.

[0010] Preferably, in step S4, under the condition that all independence fingerprint flags are true and the minimum complementary combination has no circular dependency, a controlled secondary evidence combination is introduced as controlled secondary evidence for subsequent processing. Before introduction, an introduction permission check is performed, which includes comparing the immutable content hash of each document node in the combination byte by byte to confirm hash consistency, matching the original ownership anchor point according to the longest name in the organization list and writing the matching position and matching weight, aligning the intention category item by item with the intention category sequence according to the position of the trigger statement and writing the alignment digest, comparing the version identifier and the release time to confirm time series alignment, and performing cross-reference integrity verification on the citation list to construct a citation directed graph and perform loop detection. When all the above checks pass item by item, a controlled secondary evidence introduction record is constructed. The introduction record fields include an introduction unique identifier, a list of introduced documents, and each document The discount label, discount reason, discount percentage, source document identifier, timestamp, independence fingerprint summary, and audit trigger pointer are included. The discount percentage is expressed as a percentage and calculated by the audit rule calculator based on the number of inconsistent fields and the importance weight of the fields. The discount reason consists of a list of inconsistent fields and the degree of inconsistency, and is written into the discount label field as structured text. The imported record and the discount label are written to the audit storage with immutable transactions. The writing process generates audit entries, which include a writing timestamp, a writing operator identifier, a metadata snapshot before the change, and a metadata snapshot after the change. After the audit entry is generated, the unique identifier of the audit entry is referenced in the audit pointer field of the output report to ensure traceability and verifiability. If any verification fails, the import cannot proceed, and the blocking reason and verification summary are written to the audit storage to trigger subsequent manual review or self-calibration processes.

[0011] Preferably, in step S4, the specific implementation of performing a reverse confidence loop rule on the blocked node set to filter the minimum complementary combination includes: firstly, reading the independence fingerprint vector of each document node from the blocked node set, wherein the independence fingerprint vector includes the authenticity flags of the original ownership anchor, the intention category, the semantic situation label, the version identifier, the publication time, the citation list, the immutable content hash, the audit pointer, and the binding status; then, searching for the minimum complementary combination within the blocked node set according to the complementarity determination rule, wherein the complementarity determination rule requires that the union of the authenticity flags of the selected document nodes covers eight required fields: original ownership anchor, intention category, semantic situation label, version identifier, publication time, citation list, audit pointer, and binding status; the combination search adopts a layer-by-layer enumeration parallel pruning strategy and prioritizes the combination with the fewest documents. The fingerprint authenticity coverage priority ranking aims at the minimum document set, recording the search path and pruning reasons at each step. A referenced directed graph is constructed for candidate combinations, and directed loop detection is performed. Loop detection uses depth-first search to determine the existence of directed loops. If a directed loop exists, the combination is rejected, and the reason for the circular dependency and the specific node sequence of the loop are written into the audit pointer for subsequent manual review. When the minimum complementary combination satisfies the complementarity criterion and does not have a circular dependency, reverse confidence closed-loop synthetic evidence is calculated, generating an introduction trigger record and discount label. The discount label includes the discount reason, discount amount, source document identifier, and timestamp, and is written to the audit storage using immutable transactions. The audit storage records the write operator identifier, write timestamp, metadata snapshot before the change, and metadata snapshot after the change during writing to ensure the introduction is traceable. The unique identifier of the introduction trigger record is referenced in the output audit pointer for subsequent review and evidence chain verification.

[0012] Preferably, in step S5, the specific implementation of the causal consistency verification includes: performing intention locking and time causal gate verification on the introduced closed-loop path and the introduced controlled secondary evidence combination: first, reading the intention category field of each node in the closed-loop path and verifying that the binding status field of the starting node of the closed-loop path is true; second, reading the binding status field of each document node in the controlled secondary evidence combination, recording inconsistencies, and simultaneously constructing a time series vector by sorting the timestamp sequence in ascending order according to the publication time field; mapping the intention category sequence to an allowed causal order according to the intention category to time causal arrangement rule, and comparing the time series vector with the allowed causal order item by item, requiring that the publication time of each node is earlier than or equal to its successor. The publication time of the causal node is consistent with the intention category arrangement and the causal arrangement; when the binding status field value of the starting node of the closed loop path is true and the time series is consistent with the intention causal arrangement, the closed loop path is marked as bound and allowed to be used to generate answers; when the binding status field value of all document nodes in the controlled secondary evidence combination is true and the time series is consistent, the controlled secondary evidence combination is marked as bound and allowed to be used to generate answers; when any comparison fails, the target closed loop path or controlled secondary evidence combination is marked as unbound or restricted, the failure rule matching summary and inconsistency field list are written into the audit pointer, triggering the manual review process or self-calibration suggestion process, and the trigger timestamp and trigger subject identifier are recorded in the output for traceability.

[0013] Preferably, in step S6, the specific implementation of generating trusted output includes: compiling a structured output report based on the closed-loop path of the admission or the introduced controlled secondary evidence combination, the output report including answer text, a list of referenced nodes of the closed-loop path, an independence fingerprint summary, intention category and time series verification results, a complete audit pointer and discount annotation description; when generating the output report, listing the unique identifier of each document, the start and end positions of the trigger statement fragment, the version identifier and the release time according to the document node identifier, and formatting this information into reference entries; writing each rule trigger chain and decision node into the audit storage as an immutable record, the audit storage entry includes the unique identifier of the audit entry, the writing timestamp, the writing operator identifier, the metadata snapshot before the change and the metadata snapshot after the change; the output report uses a key-value pair format to represent the field set, and generates a verifiable hash for each field for integrity verification and external verification; for fields marked as unbound or restricted... The closed-loop path or controlled secondary supporting evidence combination is prominently marked in the output report with its restricted status, accompanied by a discount explanation and review suggestion. The discount explanation includes the discount reason and discount percentage, referencing the independence fingerprint summary and the list of inconsistent fields. The discount percentage is calculated by the audit rule calculator according to the number of inconsistent fields and the importance weight of the fields, and is expressed as an integer percentage. When the output report contains unbound or restricted status items, a calibration suggestion or manual review process is automatically triggered, and a trigger record is written to the audit pointer for management query and post-event traceability. The output report is accompanied by an audit pointer format definition for management query. The audit pointer fields include a unique identifier for the audit item, a storage location pointer, a write timestamp, and a verifiable hash, so that external auditors can retrieve and verify snapshots before and after changes when the unique identifier of the audit item is provided. All output generation and audit write operations are processed with immutable transactions, and the unique identifier of the audit item is returned in the output audit pointer for external verification and audit trail.

[0014] Preferably, when the semantic loop closure verifier fails to identify any loop closure path and the independence fingerprint determination cannot generate the minimum complementary combination that can be introduced from the blocked node set, a self-calibration process is automatically triggered, with a suggested feedback connection statement as the output. The suggested feedback connection statement consists of a trigger statement fragment, a unique identifier of the original ownership anchor document, a trigger timestamp, and a suggested connection reason, along with the corresponding document fragment's original start and end positions, immutable content hash, and trigger statement hash. The suggested feedback connection statement and document fragment are packaged according to the compliance subject's submission format. The submission package fields include the submitter identifier, submission timestamp, unique identifier of the submission session, and associated audit pointer. The submission package is written to the audit storage to generate a submission audit entry. After receiving the submission package, the compliance subject performs rule verification according to the organization list rules and trigger statement matching rules. When the rule verification passes, the compliance subject writes the suggested feedback connection into the structured knowledge network in the form of a formal entry. The write record includes the writing subject identifier, write timestamp, entry unique identifier, and verifiable hash. The write process is then returned. The unique identifier of the entry and the unique identifier of the audit entry are confirmed and attached. After the knowledge graph is confirmed to be written, the retrieval engine uses the unique identifier of the entry in the confirmation as the retrieval extension parameter to re-execute the subgraph retrieval from step S2. The newly formed candidate subgraph and the immutable metadata snapshot are sent to the semantic loop validator to try the loop closure judgment again. The unique identifier of the compliance subject's write record is referenced in the retrieval session record. When the compliance subject refuses to write in the rule verification, the reason for refusal is recorded in the audit storage. A manual review instruction is generated in the output to trigger the manual review process of the management end. All self-calibration triggering, submission, verification, writing and manual review behaviors generate immutable audit entries. The fields of the audit entries include the unique identifier of the entry, the identifier of the writing operator, the writing timestamp, the metadata snapshot before the change, the metadata snapshot after the change and the verifiable hash. The audit pointer is disclosed in the structured output report for subsequent traceability and external verification. All output generation and audit writing operations are processed with immutable transactions. The unique identifier of the audit entry is returned in the output audit pointer for external verification and audit tracking.

[0015] (III) Beneficial Effects This invention provides a trusted question-answering method for national security policy that integrates knowledge graphs and GraphRAG. It has the following beneficial effects: 1. This invention achieves strict consistency between question-and-answer results and the semantic flow of national security policies by introducing semantic feedback connection and closed-loop path determination mechanism in the retrieval and verification process; and ensures the authenticity of the introduction of controlled secondary evidence by using the minimum complementary combination of independence fingerprint and reverse confidence closed-loop screening.

[0016] 2. This invention establishes immutable metadata for each document and generates semantic feedback connections by combining trigger statements, enabling the knowledge graph to have stable anchor points and structured intention information; through GraphRAG's subgraph retrieval and hop-by-hop parsing of semantic loop closure verification, the generation of answers strictly depends on verified reference chains and semantic situation label sequences. Attached Figure Description

[0017] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1 This is the overall flowchart of the present invention; Figure 2 This is the question-and-answer processing logic diagram described in this invention. Detailed Implementation

[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1: This invention provides a trusted question-and-answer method for national security policies that integrates knowledge graphs and GraphRAG, including a complete implementation process from document reception to output and audit closure. First, the document receiving module receives the policy text and metadata carrier, and assigns a unique identifier to each document for full traceability. Then, a document parser performs structured parsing on the document's beginning and body, accurately extracting the position and string of the validity statement text. The document parser performs item-by-item matching using a list of authoritative institution names, applying longest name priority matching and string exact matching rules to determine the original ownership anchor field value and matching position. The matching result and matching weight are written to the original ownership anchor field. Next, a trigger phrase matcher performs word segmentation and normalization on the entire document while preserving the original character order. This trigger phrase matcher applies longest phrase priority and unambiguous priority matching rules based on a policy keyword list. Whenever a trigger phrase is matched, a trigger statement record is generated. The trigger statement record includes the original trigger text, start and end positions, candidate intention categories, trigger timestamp, unique trigger identifier, and trigger statement hash. The immutable content hashing module calculates an immutable hash value byte-by-byte for the normalized document content and generates a hash value using the SHA256 hash algorithm. Then, the immutable hash value, along with the trigger statement hash and audit pointer, is written to the knowledge graph and audit storage using an immutable transaction, and the writing timestamp and operator identifier are recorded to ensure traceability. The binding determiner judges the validity statement and original ownership anchor field values ​​according to binding rules. These rules require that if both the original ownership anchor field value and the validity statement indication are met simultaneously, the binding status field is set to true and the version identifier and publication time are written; otherwise, the binding status field is set to false and the blocking reason explanation and timestamp are written. The trigger statement mapping module maps the trigger statement record to a semantic feedback connection edge. This semantic feedback connection edge includes a unique connection identifier, a source original ownership anchor node identifier, a target document node identifier, a connection type, a trigger statement reference field, a semantic state label field, a version identifier field, a publication time field, a reference list field, and an audit pointer field. The mapping result is written to the knowledge graph using an immutable transaction, and trigger record fragments are generated in the audit storage.

[0020] When a user's question enters the retrieval stage, the question parser performs semantic parsing and generates a retrieval expression. This retrieval expression consists of a keyword vector sequence, entity identifiers, and time constraints, with embedding vectors and Boolean constraints serving as the retrieval driving parameters for the GraphRAG retrieval engine. The GraphRAG retrieval engine first performs vector similarity matching to obtain an initial set of candidate nodes. Then, it performs hierarchical expansion based on the citation list field of the candidate nodes, recording the citation direction and position. The retrieval engine extracts immutable metadata fields for each candidate node, including: original ownership anchor, intent category, semantic state label, version identifier, publication time, citation list, audit pointer, and binding status. After extraction, the nodes are sorted using a topological and similarity-weighted mixed sorting based on citation order and intensity, and the sorting evidence fragment is written to the retrieval session record. The sorted nodes are assembled into a candidate subgraph data structure according to citation relationships. The candidate subgraph contains a node list, an edge list, a node order snapshot, an immutable metadata snapshot, and citation proof fragments. The candidate subgraph is transmitted to the semantic loop closure verifier via a secure interface, and the transmission audit pointer is recorded.

[0021] The semantic loop closure verifier performs hop-by-hop judgment and semantic situation label sequence verification on the incoming candidate subgraph. Hop-by-hop judgment requires that the connection type of the edge is semantic feedback, the trigger statement field and the intention category field on the edge are completely matched, and the trigger text position and the intention category trigger position correspond one-to-one. Semantic situation label sequence verification requires extracting semantic situation labels in node order and comparing them item by item with the state transition rules of the allowed flow sequence. The allowed flow sequence is composed of ordered state transition rules between semantic situation labels and requires that the sequence covers every hop of the closed loop path. When at least one closed loop path is identified in which all hops are semantic feedback and the semantic situation label sequence conforms to the allowed flow sequence, the closed loop path is output, and the evidence fragment forming the closed loop, the trigger rule chain, and the unique identifier of the closed loop output are recorded in the audit storage. When all candidate subgraphs are blocked, the independent fingerprint determiner reads each document in the blocked set item by item according to the immutable metadata field and compares it word by word with the source document to generate a true / false flag vector. The true / false flag vector field includes nine items: true / false of original ownership anchor point, true / false of intention category, true / false of semantic situation label, true / false of version identifier, true / false of publication time, true / false of citation list, true / false of immutable content hash, true / false of audit pointer, and true / false of binding status. The fingerprint vector is also marked with a summary of the reason for inconsistency.

[0022] The complementarity determiner searches for the minimum complementary combination in the blocked set according to complementarity determination rules. These rules require the union of truth and falsehood flags of the selected combination to cover eight essential fields: original ownership anchor, intent category, semantic state label, version identifier, publication time, citation list, audit pointer, and binding status. The combination search employs exhaustive search with pruning and priority heuristics, aiming to minimize the number of documents, and records the search path and pruning reasons. A directed citation graph is constructed for the candidate minimum complementary combinations, and directed cycle detection is performed to determine circular dependencies. Cycle detection uses depth-first search to determine the existence of directed cycles, and if a cycle exists, the sequence of cycle nodes is written into the audit pointer and blocking digest. When the minimum complementary combination satisfies all rule truth values ​​in the independence fingerprint and does not have circular dependencies, the controlled secondary corroborator verifies it item by item according to the introduction license checklist. Verification items include: immutable content hash consistency, original ownership anchor matching position, intent category alignment, version identifier and publication time alignment, and citation cross-completeness. After all verifications are passed, the audit rule calculator generates a discount label. This label includes the discount reason, a list of inconsistent fields, a discount percentage (%), the source document identifier, and a timestamp. The discount percentage is calculated by the audit rule calculator based on the number of inconsistent fields and their importance weights, and is expressed as an integer percentage (%). The controlled secondary corroboration import record fields include an import unique identifier, a list of imported documents, a discount label for each document, a discount reason, a discount percentage, the source document identifier, a timestamp, an independence fingerprint summary, and an audit trigger pointer. These fields are written to the audit storage using immutable transactions, generating audit entries.

[0023] After introduction, intent locking and temporal causality gate verification are performed. This process reads the intent category field of each node in the closed-loop path and verifies that the binding status field of the starting node of the closed-loop path is true. Simultaneously, it reads the binding status field of each document in the controlled secondary evidence combination and sorts the publication time field in ascending order to construct a time series vector. Subsequently, the intent category sequence is mapped to the allowed causal order, and the time series vector is compared item by item, requiring that the publication time of each node is earlier than or equal to the publication time of its subsequent causal node, and that the intent category arrangement is consistent with the allowed causal order. When the binding status of the starting node of the closed-loop path is true and the time series is consistent with the allowed causal order, or when the binding status fields of all documents in the controlled secondary evidence combination are true and the time series are consistent, the target item is marked as bound and allowed for answer generation. When the comparison fails, the target item is marked as unbound or restricted, and a failure rule matching summary, a list of inconsistent fields, and a trigger timestamp are written to the audit pointer. The target item is then placed in a manual review queue or a self-calibration suggestion queue.

[0024] Finally, the output generator compiles a structured output report, which includes the answer text, a list of closed-loop path node references, an independence fingerprint summary, intent category and time series verification results, a complete audit pointer, and discount annotations. The output report lists a unique identifier, the start and end positions of the trigger statement fragment, the version identifier, and the release time for each supporting document, and generates a verifiable hash for each field for external verification. Simultaneously with output generation, each rule trigger chain and decision node is written to the audit storage as an immutable record, and a unique identifier for the audit entry is generated for management queries. If the output contains unbound or restricted items, a manual review process is triggered, and the manual review instruction and review subject identifier are recorded in the audit pointer. All write operations throughout the implementation process are completed using immutable transactions, and snapshots of metadata before and after changes, the write operator identifier, and timestamps are saved in the audit storage for post-event traceability and external audit verification.

[0025] Example 2: This embodiment, based on Embodiment 1, continues the core steps of document entry, parsing, trigger phrase generation, GraphRAG retrieval, semantic loop verification, independence fingerprint determination, minimum complementary combination screening, controlled secondary evidence introduction, intent locking and time causality verification, and output auditing. It also introduces higher-granularity audit records and manual intervention strategies at key nodes to demonstrate the differences from Embodiment 1.

[0026] During the document parsing phase, a name variant mapping step is introduced. This step standardizes the organization name using a composite matching table of authoritative names and abbreviations, and records the standardization path and matching pattern. The standardization path is written to the audit pointer as a mapping path hash to enhance the determinism of the original ownership anchor field. In the triggered phrase matching phase, multiple rounds of context backtracking are added, with a maximum of three backtracking rounds. Backtracking expands the triggered phrase matching area through forward and backward windows to capture cross-sentence triggering scenarios, and the start and end positions of the backtracking window and the number of backtracking rounds are written into the triggered statement record for verification.

[0027] During the retrieval phase, a secondary re-ranking is performed on the vector similarity retrieval results. This secondary re-ranking uses a linear weighting of the citation strength score and the trigger statement priority score. The linear weighting coefficient is specified by the retrieval session record and written into the ranking evidence fragment. After the secondary re-ranking, the retrieval confidence summary is generated by the retrieval engine and written into the transmission audit pointer for reference by the closed-loop validator. In the semantic closed-loop validator, a temporal consistency assessment of the semantic situation label sequence is added. This temporal consistency assessment requires that the publication time interval of each hop in the closed-loop path meets the publication time difference threshold constraint. The publication time difference threshold is stored numerically in the audit storage and managed by the audit rule configuration. The comparison result between the publication time difference and the threshold is written into the closed-loop evidence fragment for auditing and review.

[0028] In the independence fingerprint determination and minimum complementary combination search stages, a priority-guided heuristic search engine is introduced. This heuristic search engine calculates priority scores based on document credibility scores, original ownership anchor matching strength, immutable content hash consistency, and citation coverage. Documents with higher priority scores are prioritized during the search process to reduce the exhaustive search scope, and the priority score trajectory is written into the search path. When the minimum complementary combination meets the introduction conditions and enters the controlled secondary evidence introduction stage, a tiered discount strategy is adopted during the discount annotation generation process. This tiered discount strategy divides the discount percentage into three levels: the first level has a discount range of 0% to 20%, the second level has a discount range of 21% to 50%, and the third level has a discount range of 51% to 100%. The discount level is determined by the discount suggestion value calculated by the audit rule calculator based on the number of inconsistent fields and the importance weight of the fields, and the discount level and calculation formula summary are explicitly written into the discount annotation.

[0029] In the intent locking and time causality verification phase, a manual threshold intervention path is added. This path is triggered when the comparison between the time series and the intent arrangement fails. The triggering conditions include two scenarios: inconsistency in the time series and incomplete true values ​​in the binding status field. When the manual threshold intervention path is triggered, the target closed-loop path or controlled secondary supporting evidence combination is pushed to the manual review queue. The complete audit pointer and priority score are attached to the review queue item so that the reviewer can calibrate or confirm whether it is allowed to be introduced in a controlled manner according to the audit record execution rules. The reviewer's identifier, review timestamp, and decision reason are recorded in the manual review decision, and the manual decision is written to the audit storage.

[0030] Regarding the self-calibration process, the interaction process between compliance entities is refined. After receiving the suggested feedback connection statement and before returning the write confirmation, the compliance entity marks the verification results item by item through the compliance verification form, and indicates in the verification results whether to perform text correction or supplementary explanation on the trigger statement or document fragment of the suggested feedback connection statement. The compliance entity's correction and supplementary operations are written into the knowledge graph as new immutable document fragments, and new immutable content hashes and audit entries are generated. The unique identifier of the new entry returned in the write confirmation will be used as the search extension parameter when the search engine re-searches.

[0031] In the output phase, a management-side query view field is added. This field includes a visual summary link to the output report, a timeline of audit entries, and a review process history. A pointer to this field is also exposed in the audit pointer of the output report, allowing the management end to directly retrieve and replay the entire audit event trajectory from receipt to final decision based on the unique identifier of each audit entry. Example 2 enhances audit transparency and human control by adding operational details such as name variant mapping, multi-round context backtracking, secondary reordering, time consistency assessment, priority heuristic search, tiered discount strategy, manual threshold intervention path, detailed interaction of compliance entities, and management-side query view fields at key nodes. This ensures that all operational elements remain within the scope defined in the specification and claims, guaranteeing the consistency and rigor of the implementation steps in terms of functionality and traceability.

[0032] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A credible question-answering method for national security policy that integrates knowledge graphs and GraphRAG, characterized in that, Includes the following steps: S1, Construct a knowledge graph: Extract policy keywords and relationships between keywords from the collected document data; the policy keywords include policy subject, involved field, main content, and time node; the relationships between keywords include formulation relationship, implementation relationship, and constraint relationship; construct graph nodes based on the extracted policy keywords, and construct graph edges based on the extracted relationships between keywords to form a structured knowledge network; establish an immutable metadata field set for each document node in the structured knowledge network, containing the following fields: original ownership anchor, intention category, semantic situation label, version identifier, publication time, citation list, audit pointer, and binding status; Based on the trigger statements of rule-based annotation, semantic feedback connection edges are generated in the structured knowledge network; S2, Question-driven subgraph retrieval based on GraphRAG: Semantic parsing of user questions to extract search keywords; using the GraphRAG retrieval engine to match the search keywords with the structured knowledge network to locate relevant graph nodes and connecting edges; The retrieval system returns candidate nodes containing the immutable metadata field set, sorts them according to the reference relationships between nodes, assembles them into a candidate subgraph, and sends it to the semantic loop closure verifier. S3, Perform semantic loop closure verification: In the semantic loop closure verifier, identify closed-loop paths from the original ownership anchor point and back through semantic feedback connection edges from the candidate subgraph; allow the closed-loop path to enter subsequent processing only if each hop of at least one of the closed-loop paths is a semantic feedback connection edge and the semantic situation label sequence on the path conforms to the preset allowed flow sequence; otherwise, reject the candidate subgraph, classify its nodes into the blocked node set, and record the blocking reason; S4, Implement controlled secondary corroboration introduction: When the candidate subgraph is rejected, trigger the independence fingerprint determination of the blocked node set, and select the minimum complementary combination from the blocked node set according to the reverse confidence loop closure rule; only when the minimum complementary combination satisfies all rule truth values ​​in the independence fingerprint flag and there is no circular dependency, it is introduced in the form of controlled secondary corroboration combination, and the discount label and audit chain are recorded; S5, Perform verification: For any admitted closed-loop path or the introduced controlled secondary evidence combination, perform an intention causal permutation verification based on intention category and publication time series; If the binding status of the starting node of the closed-loop path is true and the publication time series is consistent with the intended causal arrangement, or if the binding status of all documents in the controlled secondary corroboration combination is true and the publication time series is consistent, then they are used to generate the answer. Otherwise, it will be marked as unbound or restricted, and a manual review or calibration process will be triggered; S6, Generate trusted output: Generate a structured output including the closed-loop path node reference, the independence fingerprint determination summary, the intention causal arrangement verification result, and the complete audit pointer; The structured output is generated primarily based on authoritative documents with a true binding status; when the output contains unbound or restricted status items, a calibration or manual review process is triggered.

2. The national security policy credible question-answering method integrating knowledge graph and GraphRAG as described in claim 1, characterized in that, In step S1, establishing the immutable metadata field set for each document node in the structured knowledge network specifically includes: parsing the original document corresponding to the document node to be entered into the database, parsing the header and the validity statement in the text, and determining and writing the original ownership anchor field according to the institution list rule based on the authoritative institution name list matching; matching the entire original document with trigger phrases based on the policy keyword list, and writing them into the intention category field accordingly; establishing an immutable content hash for the original document, and storing the immutable content hash together with the audit pointer to ensure the traceability of subsequent rule determinations; setting the binding status for the document node according to the binding rule based on the authoritative institution name and validity statement, and writing it into the binding status field.

3. The national security policy credible question-answering method integrating knowledge graph and GraphRAG as described in claim 1, characterized in that, In step S2, question-driven subgraph retrieval based on GraphRAG specifically includes: semantic parsing of the user question and generating a retrieval expression based on the parsing results, driving the GraphRAG retrieval engine to retrieve candidate nodes in the structured knowledge network; for each candidate node returned by GraphRAG, extracting the immutable metadata field and sorting them according to the node reference relationships determined based on the reference list field to form the candidate subgraph; and transmitting the candidate subgraph and the immutable metadata to the semantic loop closure validator in a data structure for subsequent rule verification.

4. The national security policy credible question-answering method integrating knowledge graph and GraphRAG as described in claim 1, characterized in that, In step S3, semantic loop closure verification is performed in the semantic loop closure verifier, specifically including: determining whether each connection of the candidate subgraph satisfies the semantic feedback connection rule based on the matching of the trigger statement and the intention category; if the hop-by-hop determination is successful, parsing the semantic situation label in node order and verifying whether the semantic situation label sequence is consistent with the allowed flow sequence based on the semantic situation label state transition rule; when at least one closed-loop path that satisfies the hop-by-hop determination and semantic situation label sequence verification is identified, outputting the closed-loop path and recording the evidence fragment and trigger rule chain of the closed loop formation.

5. The national security policy credible question-answering method integrating knowledge graph and GraphRAG as described in claim 1, characterized in that, In step S4, the controlled secondary corroboration introduction includes: evaluating the original document corresponding to each document node in the blocked node set based on consistency with the source document according to the immutable metadata fields, and generating true / false flags for each immutable metadata field to form an independence fingerprint; based on the independence fingerprint, according to the complementarity judgment rule that requires the true / false flag combination to cover all necessary fields, selecting the minimum complementary combination that satisfies the complementary truth value, and evaluating the reference cross-references between the minimum complementary combinations to determine circular dependencies; when the minimum complementary combination passes the complementarity judgment and has no circular dependencies, the minimum complementary combination is introduced in a controlled secondary corroboration manner, and an auditable trigger record and discount description are generated at the time of introduction.

6. The national security policy credible question-answering method integrating knowledge graph and GraphRAG as described in claim 1, characterized in that, In step S5, the causal consistency verification is performed, specifically including: reading the intention category field of each node on the admitted closed-loop path and verifying the binding status of the starting node of the closed-loop path; reading the binding status of each document node in the introduced controlled secondary evidence combination; obtaining a timestamp sequence based on the publication time field and verifying the timestamp sequence of the closed-loop path to determine whether the time order is consistent with the causal arrangement of the intention category; verifying whether the timestamp sequence of the document nodes in the controlled secondary evidence combination is consistent with the causal arrangement; when any verification fails, the corresponding closed-loop path or the controlled secondary evidence combination is marked as the unbound state or the restricted state, and the rule matching summary of the verification failure is written to the audit pointer to trigger the manual review process or automatically initiate a self-calibration suggestion.

7. The national security policy credible question-answering method integrating knowledge graph and GraphRAG as described in claim 1, characterized in that, In step S6, generating trusted output specifically includes: compiling a structured output report based on the closed-loop path of admission or the introduced controlled secondary evidence combination, and listing the document node identifier and the trigger statement fragment for each basis in the report; writing each rule trigger chain and decision node into the audit storage as an immutable record, and generating the audit pointer for the management terminal to query; in the structured output, clearly indicating the restricted status of any closed-loop path or controlled secondary evidence combination marked as the unbound state or the restricted state, and attaching corresponding discount explanations and review suggestions.

8. The national security policy credible question-answering method integrating knowledge graph and GraphRAG as described in claim 1, characterized in that, When the semantic loop closure verifier fails to identify the loop closure path and the independence fingerprint determination cannot generate the minimum complementary combination that can be introduced, an advisory feedback connection statement is automatically generated, and the advisory feedback connection statement and the corresponding document fragment are submitted to the designated compliance subject for rule verification. After the compliant entity passes the rule verification and writes the suggested feedback as a formal entry into the structured knowledge network, the subgraph retrieval and closed-loop verification process from step S2 onwards is retried. All self-calibration triggers, compliance entity verification processes, and write operations are recorded as audit entries, and the audit pointers are disclosed in the structured output for retrospective review.

Images