Search scoring knowledge version control and vector access methods and systems
By splitting and generating fingerprints for structured scoring cases and distinguishing case types according to deterministic rules, the problems of redundant storage and scoring drift in existing technologies are solved, and efficient version control of scoring knowledge and vector admission are achieved.
Patent Information
- Application Number
- CN202611569157.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-12
- Publication Date
- 2026-11-17
AI Technical Summary
When processing structured scoring cases, existing technologies rely on hashing methods or semantic similarity judgments for the entire record, resulting in redundant storage and repeated vector writing. This makes it difficult to reliably distinguish between complete duplication, identical score revisions, and dissimilar score conflicts, leading to score drift and knowledge pollution.
By splitting the received historical scoring cases, scoring context fingerprints and version fingerprints are generated. Case types are distinguished according to deterministic rules, status identifiers and vector admission identifiers are set, and only currently valid versions that are allowed by vector admission are created, thus blocking conflicting versions.
This avoids redundant storage and duplicate vector writes, reduces the risk of score drift, maintains the accuracy and consistency of scores, and ensures version traceability and the integrity of the audit chain.
Smart Images

Figure CN123114536A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, and in particular to a search scoring knowledge version control and vector admission method and system. Background Technology
[0002] In Internet Search and Retrieval Enhancement Generation (RAG) systems, structured scoring cases are used for manual calibration and knowledge accumulation of search results. A typical process includes: receiving input containing query terms, intent scenarios, and ordered search result entries; retrieving historical scoring cases through the RAG service; inputting matching cases as context into a large language model to generate scoring results. Existing technologies include content hashing for deduplication of scoring cases, conflict detection based on semantic similarity, support for manual feedback write-back, recording request traces, and conducting parameterized strategy evaluation. However, when processing structured scoring cases, these solutions use a whole-record hashing method to calculate unique identifiers, treating changes in manual scores or scoring reasons as new cases. This leads to redundant storage and duplicate vector writing of the same scoring context due to version revisions. Summary of the Invention
[0003] In view of this, the purpose of the present invention is to provide a search scoring knowledge version control and vector admission method and system that can distinguish different cases according to deterministic rules, and avoid redundant storage and repeated vector writing caused by whole record hashing method or semantic similarity judgment.
[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides a method for version control and vector admission of search scoring knowledge, comprising: splitting received historical scoring cases to obtain structured scoring submissions; performing deterministic normalization processing on the structured scoring submissions, and generating scoring context fingerprints and scoring version fingerprints based on the normalized structured scoring submissions; determining the submission type of each historical scoring case based on the scoring context fingerprints and scoring version fingerprints; creating or reusing case subjects based on the submission type, establishing or updating case version records, and setting status identifiers and vector admission identifiers for corresponding versions, and creating online vectors for versions whose status identifiers are currently valid and whose vector admission identifiers are allowed.
[0005] Optionally, the structured score submission includes the project scope, query terms, intent scenario, multiple search result entries in their original relative order, human score, and scoring reason. Based on the normalized structured score submission, a scoring context fingerprint and a scoring version fingerprint are generated, including: concatenating the normalized project scope, query terms, intent scenario, and each search result entry to generate a first summary value, which serves as the scoring context fingerprint; and performing a deterministic summary operation on the scoring context fingerprint, combined with the normalized human score and scoring reason, to generate a second summary value, which serves as the scoring version fingerprint.
[0006] Optionally, based on the scoring context fingerprint and the scoring version fingerprint, the submission type of each historical scoring case is determined, including: based on whether there is the same scoring context fingerprint within the project scope, and whether there is the same scoring version fingerprint, the submission type is determined in turn as a new case, a complete duplicate, a revision with the same score, or a conflict with a different score.
[0007] Optionally, the submission type is determined sequentially as follows: new case, complete duplicate, same-score revision, or dissimilar conflict, including: if there is no identical scoring context fingerprint within the project scope, the submission type of the historical scoring case is determined to be a new case; if there is an identical scoring context fingerprint within the project scope and the scoring version fingerprint of any historical version is the same as the scoring version fingerprint of the historical scoring case, the submission type of the historical scoring case is determined to be a complete duplicate; if there is an identical scoring context fingerprint within the project scope, no identical scoring version fingerprint, and the current human score is equal to the human score of the current valid version, the submission type of the historical scoring case is determined to be a same-score revision; if there is an identical scoring context fingerprint within the project scope, no identical scoring version fingerprint, and the current human score is not equal to the human score of the current valid version, the submission type of the historical scoring case is determined to be a dissimilar conflict.
[0008] Optionally, set the status flag and vector access flag for the corresponding version, including: setting the status flag of the version corresponding to the new case to "currently valid" and the vector access flag to "allowed"; not adding a new version for completely duplicated versions, and not changing the status flag and vector access flag of the corresponding version; setting the status flag of the version corresponding to the same-score revision to "currently valid", the original version to "replaced", and the vector access flag of the new version to "allowed"; setting the status flag of the version corresponding to the dissimilar conflict to "conflict pending adjudication" and the vector access flag to "blocked".
[0009] Optionally, the method further includes: generating a unique tracking identifier during the RAG enhanced scoring process, and associating it with the scoring request, the retrieved case version identifier, the scoring model output, and the human score; wherein, the tracking identifier is used to fix the knowledge snapshot, bind the model and prompt word versions in the strategy evaluation, and ensure that the case version used in the strategy evaluation is consistent with the actual searchable version in the vector storage.
[0010] Optionally, the method further includes: before the strategy evaluation service starts, filtering scoring case versions with a status identifier of currently valid and a vector access identifier of allowed, arranging the filtered scoring case versions in ascending order of case identifier, extracting the scoring context fingerprint and scoring version fingerprint of each scoring case version, and generating a knowledge snapshot summary; responding to the evaluation request, performing stratified sampling of the currently valid and allowed access scoring case versions according to the preset intent scenario and human standard score, fixing the version identifier, search result item order, human standard score, random seed, and model identifier of the selected cases; and executing multiple tests under the same knowledge snapshot summary, the same model identifier, the same prompt word configuration version, and the same RAG service configuration parameters. The system pre-defines search strategies and records the actual case version identifier, ranking, similarity, scoring deviation, RAG service response latency, and the number of knowledge items actually retrieved for each strategy. For each strategy, based on the project's pre-defined hard constraints, it verifies whether each recorded indicator exceeds the corresponding hard constraint threshold. If any indicator exceeds the corresponding hard constraint threshold, the strategy is marked as unqualified. For strategies marked as qualified, the system performs a weighted calculation of each verified indicator based on pre-defined weights to generate a comprehensive utility value. The strategy recommendation result with the highest comprehensive utility value and in a pending confirmation state is output to the authorized personnel's operation interface. Upon receiving confirmation instructions from the authorized personnel, the system updates the default search strategy configuration in the production environment.
[0011] Secondly, this invention provides a search scoring knowledge version control and vector admission system, comprising: a case splitting module for splitting received historical scoring cases to obtain structured scoring submissions; a dual fingerprint generation module for performing deterministic normalization processing on the structured scoring submissions and generating a scoring context fingerprint and a scoring version fingerprint based on the normalized structured scoring submissions; a dual fingerprint determination module for determining the submission type of each historical scoring case based on the scoring context fingerprint and the scoring version fingerprint; and a vector admission module for creating or reusing case subjects based on the submission type, establishing or updating case version records, setting the status identifier and vector admission identifier of the corresponding version, and creating online vectors for versions whose status identifier is currently valid and whose vector admission identifier is allowed.
[0012] Thirdly, the present invention provides an electronic device including a processor and a memory, the memory storing computer-executable instructions executable by the processor, the processor executing the computer-executable instructions to implement the steps of the method provided in any of the first aspects above.
[0013] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, performs the steps of the method provided in any of the first aspects above.
[0014] This invention brings the following beneficial effects: The search scoring knowledge version control and vector admission method and system provided by this invention first splits the received historical scoring cases to obtain structured scoring submissions; secondly, it performs deterministic normalization processing on the structured scoring submissions and generates scoring context fingerprints and scoring version fingerprints based on the normalized structured scoring submissions; then, it determines the submission type of each historical scoring case based on the scoring context fingerprints and scoring version fingerprints; finally, it creates or reuses the case subject based on the submission type, establishes or updates the case version record, and sets the status identifier and vector admission identifier for the corresponding version, creating online vectors for versions whose status identifier is currently valid and whose vector admission identifier is allowed. In the above method, the scoring context identity and scoring version identity of a scoring case are separated by scoring context fingerprint and scoring version fingerprint. The submission type of each historical scoring case is determined based on the scoring context fingerprint and scoring version fingerprint. Case subjects are created or reused according to the submission type, and case version records are established or updated. This allows different cases to be distinguished according to deterministic rules, avoiding redundant storage and duplicate vector writing caused by hashing methods or semantic similarity judgments of the entire record. At the same time, online vectors are created only for versions with a status identifier of "currently valid" and a vector admission identifier of "allowed". This ensures that conflicting versions remain blocked while preserving evidence and audit chains, and versions with an original status identifier of "currently valid" and a vector admission identifier of "allowed" continue to serve, reducing the risk of conflict knowledge contaminating online RAG and causing scoring drift.
[0015] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained in accordance with the structures particularly pointed out in the description, claims and drawings.
[0016] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description
[0017] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0018] Figure 1 A flowchart of a search scoring knowledge version control and vector admission method provided in an embodiment of the present invention; Figure 2 A flowchart of scoring context dual fingerprint generation and joint classification is provided in this embodiment of the invention; Figure 3 This invention provides a case version state and vector admission state transition diagram. Figure 4 This invention provides a timing diagram for online RAG enhancement scoring and Trace experiment acquisition. Figure 5 A flowchart of four-strategy evaluation and constraint-driven selection under the same knowledge snapshot is provided in this embodiment of the invention; Figure 6 A comparison chart of complete consistency rate and P95 latency for four-strategy case recall provided in an embodiment of the present invention; Figure 7 A schematic diagram of the overall system architecture and data flow provided in an embodiment of the present invention; Figure 8 A schematic diagram of the structure of a search scoring knowledge version control and vector admission system provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation
[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, 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.
[0020] Currently, existing technologies include methods for deduplicating scoring cases through content hashing, detecting conflicts based on semantic similarity, supporting manual feedback and write-back, recording request tracking identifiers, and conducting parameterized strategy evaluation. However, these methods have the following technical limitations when processing structured scoring cases: First, search scoring cases have clear structural boundaries: the scope of the project, query terms, intent scenario, and ordered result entries jointly determine the object being scored, while the manual score and scoring reason belong to the scoring version of that object. Whole record hashing can misjudge changes in score or reason as new objects; judging solely by query terms or semantic similarity makes it difficult to definitively express changes in result entries and their order, thus failing to reliably distinguish between complete duplication, identical score revisions, and dissimilar score conflicts.
[0021] Second, when cases generated through step-by-step addition, Excel import, historical migration, or manual revision are directly written into the vector library, heterogeneous knowledge may simultaneously participate in recall and cause score drift; simple overwriting will result in the loss of source and audit evidence. At the same time, if there is a lack of unified control over relation version status and vector visibility, concurrent or asynchronous retries may cause expired vectors to be brought back online; if policy experiments do not fix cases, models, prompts, and knowledge snapshots, it is impossible to reliably attribute experimental differences.
[0022] Based on this, the present invention provides a search scoring knowledge version control and vector admission method and system that can distinguish different cases according to deterministic rules, avoiding redundant storage and repeated vector writing caused by whole record hashing or semantic similarity judgment.
[0023] To facilitate understanding of this embodiment, the technical terms involved in the embodiment of the present invention will first be explained: RAG (Retrieval-Augmented Generation): A technique that retrieves relevant knowledge before invoking the generative model and uses it as the model context.
[0024] Structured scoring cases and scoring context: Structured scoring cases are scoring knowledge units that link human scores, scoring reasons, and versions; their scoring context is determined by the scope of the project, query terms, intent scenario, and search results entries that maintain their original relative order.
[0025] Rating Context Fingerprint Fc: A deterministic summary computed only on the normalized rating context, excluding human scores and rating reasons, used to determine context identity.
[0026] The scoring version fingerprint Fv is a deterministic summary generated by adding normalized human scores and scoring reasons to Fc, used to determine the specific scoring version.
[0027] Dual fingerprint joint determination: Based on the order of Fc, Fv and manual score, the submission certainty is divided into new cases, complete duplicates, revisions with the same score, or conflicts with different scores.
[0028] Current valid version and vector admission status: The former is the only version used for online services in the same scoring context; the latter determines whether the version can be generated, updated or become an online searchable vector, including at least allowed and blocked.
[0029] Conflict vector isolation: When different manual scores appear in the same scoring context, conflicting versions and audit evidence are persistently saved, but they are set to blocked before manual adjudication, and the original currently valid version continues to be used.
[0030] Knowledge Snapshot: A fixed list of currently searchable scoring case versions and their summary identifiers at the start of strategy evaluation, used to ensure consistency of knowledge conditions for multi-strategy comparison.
[0031] Trace, P95 latency, and hard constraints: Trace is a unique traceable identifier that connects scoring requests, RAG retrieval, hit case versions, strategies, and model / human scores; P95 latency is the 95th percentile value of the requested latency; hard constraints are the minimum accuracy, maximum latency, or safety conditions set for the project.
[0032] Next, a detailed description of a search scoring knowledge version control and vector admission method disclosed in this invention will be provided. This method can be executed by electronic devices, such as smartphones, computers, and tablets. See also Figure 1 The flowchart shown illustrates a search scoring knowledge version control and vector admission method, which mainly includes the following steps S101 to S104: Step S101: Split the received historical scoring cases to obtain structured scoring submissions.
[0033] In one implementation, this method can be applied to an AI-RAG knowledge platform and an AI scoring system. The scorer selects "Add to RAG Knowledge Base" in the record details and fills in the required manual score and scoring reason; the client simultaneously submits the project scope, query terms, intent scenario, and result entries in their original order. Importing history from Excel and submitting a single entry enters the same server-side judgment process. The page displays the query terms, intent scenario, current score, version number, conflict status, and vector admission status, and provides an entry point for conflict adjudication for authorized personnel.
[0034] The server breaks down historical scoring cases into structured scoring submissions, as shown in Table 1. Each structured scoring submission includes: the case body, ordered result entries, case version, submission audit, and vector metadata. The case body stores the context identity field and a reference to the current version; result entries retain their content and relative order using serial numbers; the case version stores the human score, reason, Fv, version status, and admission status; the submission audit stores the source, operator, submission type, and time; and the vector metadata includes at least the case ID, version ID, project scope, and visibility status.
[0035] Table 1. Composition of Structured Scoring Case Data
[0036] Step S102: Perform deterministic normalization processing on the structured score submission, and generate a score context fingerprint and a score version fingerprint based on the normalized structured score submission.
[0037] In one implementation, the structured scoring submission includes an item scope P, a query term Q, an intent scenario S, and multiple search result entries R=[r1,r2,…,r] in their original relative order. n The system consists of a score (Y) and a scoring reason (E). The server first performs deterministic normalization (N(·)) to unify full-width / half-width characters, character case, consecutive whitespace, null values, and scene encoding. Resulting entries are calculated according to their original search order and are not converted into an unordered set.
[0038] To avoid ambiguity in field concatenation, we define Enc(x) = Len(N(x)) || "#" || N(x), where Len(·) is the length of the normalized string, || indicates sequential concatenation, and H(·) is the collision-resistant digest function.
[0039] In practice, the standardized project scope, query terms, intent scenarios, and each search result entry are concatenated to generate a first summary value, which serves as the scoring context fingerprint. Specifically, the scoring context fingerprint Fc is: Fc=H(Enc(P)||Enc(Q)||Enc(S)||Enc(r1)||…||Enc(r n )) Fc only expresses the contextual identity of the object being rated, and does not include the human score Y or the reason for the rating E.
[0040] Furthermore, based on the scoring context fingerprint, and combined with the normalized human score and scoring reason, a deterministic digest operation is performed to generate a second digest value, which serves as the scoring version fingerprint. Specifically, the scoring version fingerprint Fv is calculated based on Fc: Fv=H(Enc(Fc)||Enc(Y)||Enc(E)) Fc is used to determine if they belong to the same rating context, and Fv is used to determine if an identical rating version already exists in that context. Manual scores and reasons must not be included in Fc, and the order of result entries must not be ignored in Fc; the boundaries of this field and the dependency relationship between Fc and Fv are the basis for subsequent joint determinations.
[0041] It should be noted that, in this embodiment of the invention, the digest functions of Fc and Fv can employ SHA-256, SHA-3, BLAKE series, or other anti-collision algorithms; field boundaries can be expressed using length prefixes, unambiguous delimiters, or structured serialization. Ordered result entries can be expressed by entry name, business identifier, location information, or a combination thereof; in other business domains, project scope, query terms, and intent scenarios can also be replaced with equivalent contextual identity fields, but manual scoring and scoring reasons remain hierarchical with contextual identity. No limitations are imposed here.
[0042] Step S103: Determine the submission type of each historical scoring case based on the scoring context fingerprint and the scoring version fingerprint.
[0043] In one implementation, based on whether there are identical scoring context fingerprints and identical scoring version fingerprints within the project scope, the submission type is sequentially determined as a new case, a complete duplicate, a revision with the same score, or a conflicting score. Specifically, if there are no identical scoring context fingerprints within the project scope, the submission type of the historical scoring case is determined to be a new case; if there are identical scoring context fingerprints within the project scope and the scoring version fingerprint of any historical version is the same as the scoring version fingerprint of the historical scoring case, the submission type of the historical scoring case is determined to be a complete duplicate; if there are identical scoring context fingerprints within the project scope, no identical scoring version fingerprints, and the current human score is equal to the human score of the current valid version, the submission type of the historical scoring case is determined to be a revision with the same score; if there are identical scoring context fingerprints within the project scope, no identical scoring version fingerprints, and the current human score is not equal to the human score of the current valid version, the submission type of the historical scoring case is determined to be a conflicting score.
[0044] See Figure 2 The flowchart shown illustrates a scoring context dual fingerprint generation and joint classification process. The server performs joint judgment according to a fixed priority. The joint judgment principles include: first, determining the scoring context identity using the Fc; then, eliminating completely duplicate versions using the Fv; and finally, distinguishing between revisions and conflicts based on the current score relationship. Specifically, if no identical Fc exists within the project scope, `new_case` (i.e., a new case) is output; if identical Fc exists within the project scope and any historical version has the same Fv, `exact_duplicate` (i.e., completely duplicate) is output; if identical Fc exists within the project scope and both Fv are different, and the current human score equals the current valid version's human score, `new_revision` (i.e., revision with the same score) is output; for other submissions with different Fcs, `score_conflict` (i.e., conflicting submissions) is output. The judgment result is a direct control input for version creation, current switching, and vector admission. See Table 2 for specific dual fingerprint joint classification and control rules.
[0045] Table 2: Dual Fingerprint Joint Classification and Control Rules
[0046] In this embodiment of the invention, the version, commit audit, and vector task are written within the same transaction or a process with equivalent idempotent constraints, ensuring that there is only one current version for the same case. The vector task carries the target version_id and expected_current_version_id, and re-verifies current and admission_status (status flags) before execution; expired tasks are discarded directly.
[0047] Step S104: Create or reuse the case subject based on the submission type, establish or update the case version record, and set the status identifier and vector admission identifier for the corresponding version. Create online vectors for versions with the status identifier being currently valid and the vector admission identifier being allowed.
[0048] In one implementation, versions with a status flag of "currently valid" and a vector access flag of "allowed" are written to the vector storage. Specifically, the status flag of the version corresponding to the new case is set to "currently valid," and the vector access flag is set to "allowed." For completely duplicated versions, no new version is created, and the status flag and vector access flag of the corresponding version are not changed. The status flag of the version corresponding to the same-score revision is set to "currently valid," the original version is set to "replaced," and the vector access flag of the new version is set to "allowed." The status flag of the version corresponding to the dissimilar conflict is set to "conflict pending," and the vector access flag is set to "blocked."
[0049] See Figure 3 The diagram illustrates a version status and vector admission status transition for one case. The first version of `new_case` is set to `current / allowed` (i.e., the status is currently valid and the vector admission status is allowed), allowing the creation of online vectors. `exact_duplicate` only adds audits, without creating versions or vectors. `new_revision` creates a new version, sets the old version to `superseded` (superseded), sets the new version to `current / allowed`, and updates the online vectors. `score_conflict` creates `conflict_pending / blocked` versions (i.e., the status is set to `conflict pending`, and the vector admission status is set to `blocked`), retaining the source and reason, but does not create or expose online vectors; the original `current / allowed` version continues to serve. Online retrieval simultaneously verifies the project scope, the current validity flag of `version_id`, and `admission_status=allowed`. The relational database is the status control source, and the vector side is synchronized through idempotent tasks or outbox mechanisms.
[0050] Authorized personnel can maintain the current version, convert a conflicting version to a valid version, or create an adjudicated version by combining multiple versions. Switching online vectors is only permitted when the adjudication result simultaneously satisfies current=true and admission_status=allowed. The adjudicator, time, reason, participating versions, and status changes are all written into the audit chain.
[0051] It should be noted that vector storage can employ independent vector databases, relational database vector extensions, or search engine vector indexes; vector admission can be implemented through physical write control, logical visibility fields, versioned namespaces, or search filtering. Cross-storage synchronization can utilize outboxes, message queues, or idempotent tasks, but before task execution, it is still necessary to verify whether the target version is current and allowed.
[0052] Conflict resolution can be achieved through single-person approval, multi-person review, or rule-assisted manual judgment; the scoring model can employ a large language model, a classification / regression model, or a combination thereof. Regardless of the specific implementation, the conflicted version remains unsearchable online before the resolution, and only the current / allowed versions are allowed to switch online vectors after the resolution.
[0053] The search scoring knowledge version control and vector admission method provided in this embodiment of the invention separates the scoring context identity and scoring version identity of a scoring case through scoring context fingerprint and scoring version fingerprint, and determines the submission type of each historical scoring case based on the scoring context fingerprint and scoring version fingerprint. It then creates or reuses the case subject and establishes or updates the case version record according to the submission type. This allows for the differentiation of different cases according to deterministic rules, avoiding redundant storage and repeated vector writing caused by hashing methods or semantic similarity judgments of the entire record. Simultaneously, only versions with a current status mark of validity and a vector admission mark of permission are written to the vector storage. This ensures that conflicting versions remain blocked while preserving evidence and the audit chain, while versions with a current status mark of validity and a vector admission mark of permission continue to serve, reducing the risk of conflicting knowledge contaminating the online RAG and causing scoring drift.
[0054] In one implementation, the method further includes: generating a unique tracking identifier during the RAG enhanced scoring process, and associating it with the scoring request, the retrieved case version identifier, the scoring model output, and the human score; wherein the tracking identifier is used to fix the knowledge snapshot, bind the model and prompt word versions in the strategy evaluation, and ensure that the case version used in the strategy evaluation is consistent with the actual searchable version in the vector storage.
[0055] For specific implementation, see Figure 4The diagram illustrates a timeline for online RAG enhanced scoring and trace experiment data collection. The AI scoring service generates a unique trace_id (unique trace identifier) and sends the project identifier, knowledge base identifier, query terms, intent scenario, and optional strategy_code (unique identifier of the strategy) to the RAG service. The RAG service retrieves data by project scope and current / allowed conditions, returning knowledge text, case_id (project ID), version_id, ranking, and similarity. The enhanced context is then passed to the scoring model to output a score and explanation. The experiment data collection end uses the same trace_id to associate the hit version, strategy and actual parameters, model and prompt identifiers, knowledge snapshot, model score, and human score. If a human score is generated later, the trace and hit version are saved first, and the deviation is recalculated after backfilling. If data collection fails, the process is downgraded without blocking online scoring.
[0056] When a specific human score exists, the system calculates the following indicators: Absolute rating deviation d: d = |Y model Y human |; Complete consistency index: I0=1[d=0]; Index with an error within one minute: I1=1 [d≤1].
[0057] Among them, Y model To score the model, Y human For records with only a manual "pass / fail" rating without a specific score, the system does not calculate scoring bias. Failure to collect data during an experiment will result in a downgraded process, but will not affect the main AI scoring workflow.
[0058] In one implementation, the method further includes: First, before the policy evaluation service is started, selecting scoring case versions with a current valid status and an allowed vector access status, and arranging the selected scoring case versions in ascending order of case identifiers, extracting the scoring context fingerprint and scoring version fingerprint of each scoring case version, and generating a knowledge snapshot summary; Second, in response to the evaluation request, performing stratified sampling of the currently valid and allowed access scoring case versions according to a preset intent scenario and human standard score, fixing the version identifier, search result item order, human standard score, random seed, and model identifier of the selected cases; Then, under the same knowledge snapshot summary, the same model identifier, the same prompt word configuration version, and the same RAG service configuration parameters, respectively executing... Multiple preset search strategies are implemented, and for each strategy, the actual case version identifier, ranking, similarity, scoring deviation, RAG service response latency, and the number of knowledge items actually retrieved are recorded. Next, for each strategy, based on the project's pre-set hard constraints, each recorded indicator is checked to see if it exceeds the corresponding hard constraint threshold. If any indicator exceeds the corresponding hard constraint threshold, the strategy is marked as unqualified. Then, for the marked qualified strategies, the indicators that have passed the checks are weighted according to preset weights to generate a comprehensive utility value. Finally, the strategy recommendation result with the highest comprehensive utility value and in the pending confirmation state is output to the authorized personnel's operation interface, and after receiving the authorized personnel's confirmation instruction, the default search strategy configuration in the production environment is updated.
[0059] For details, see Figure 5 The flowchart shown is a four-strategy evaluation and constraint-driven selection process under the same knowledge snapshot. Before the first call, the scoring case versions with the status mark of being currently valid and the vector admission mark of being allowed are filtered to generate a knowledge snapshot summary containing all searchable case versions and the fingerprint Fv of the scoring version. If any subsequent call finds that the snapshot has changed, the current experiment is terminated and renumbering is required.
[0060] Upon receiving an evaluation request, the strategy evaluation service performs stratified sampling from the current / allowed structured scoring cases according to the preset intent scenario and the human standard score, fixing the case ID, version ID (version identifier), search result item order, human standard score, random seed, model identifier, and evaluation set hash.
[0061] Under the conditions that the case input, scoring model (i.e., model identifier), prompt (keyword), RAG configuration version (RAG service configuration parameters), and knowledge snapshot summary are the same, the default baseline, time-priority, balanced, and accuracy-priority strategies are executed respectively, and the time-period bias is reduced by rotating or randomizing the order. The retrieval strategy template and technical objectives are shown in Table 3.
[0062] Table 3 Search Strategy Templates and Technical Objectives
[0063] After each execution, the system saves indicators such as whether the correct case is recalled, the hit version ID (i.e., the case version identifier), ranking, similarity, scoring bias, RAG model latency (i.e., RAG service response latency), and actual candidate knowledge volume (i.e., the number of knowledge items actually involved in the retrieval). Based on pre-set hard constraints, unqualified strategies are eliminated, and the comprehensive utility value of qualified strategies is calculated.
[0064] For example: for project j and strategy k, let the case recall rate be R. jk The accuracy index for scoring is A. jk P95 latency is L jk Safety violation indicator is S jk The server first determines based on A jk ≥A min (Minimum value of the preset scoring accuracy index), L jk ≤L max (The maximum value of P95 delay is preset) and Sjk=0 and other pre-set hard constraints are used to eliminate unqualified strategies. Then, the comprehensive utility value U is calculated for the remaining qualified strategies. jk : U jk =w A ·A jk +w R ·R jk w L ·Norm(L jk ) w S ·S jk Among them, w A w R w L and w S The system stores the project weights, and Norm(·) is the normalization function. The system saves the metrics, eliminated hard constraints, utility values, and reasons for recommendation; the recommendation results only enter a pending confirmation state, and the production strategy can only be updated after authorized personnel confirm them.
[0065] To facilitate understanding, this embodiment of the invention also provides a specific example where a user submits a score of 5 points and the reason for the score for the query term "air tickets to Kunming", the intent scenario "traffic_air", and eight ordered search results. Since no duplicate Fc exists in the database, the system determines it as a new_case, creates case V1, sets V1 as current / allowed, and admits it into the online vector.
[0066] When the same project scope, query terms, intent scenario, item content and order are submitted again, and the human score and scoring reason are also the same, the system first hits the same Fc, and then finds that Fv is the same as the historical V1. Therefore, it is judged as exact_duplicate. Only a new submission is added for auditing, and no new version is created or vectors are generated repeatedly.
[0067] If a score of 5 is submitted under the same scoring context, but a new scoring reason is added, and the Fc is the same, the Fv is different from the historical version, and the human score is the same as the currently valid V1, it is judged as new_revision. The system creates V2, sets V1 to superseded, sets V2 to current / allowed, and updates the online vectors; V1 retains the audit.
[0068] If another person submits a score of 0 for the same scoring context, and the Fc is the same and there is no identical Fv, but the human score differs from the currently valid V2, it is judged as a score_conflict. The system creates a conflicting version Vc and saves its source, sets Vc to conflict_pending / blocked, and V2 continues to be online. Only after a human decision confirms a score of 0 is Vc or the newly created decision version allowed to become current / allowed and switch online vectors.
[0069] A project requires a P95 latency of no more than 2 seconds and a complete consistency rate of no less than 97%. The system runs four sets of policies under a fixed evaluation set, model, prompt, and knowledge snapshot. Policies that violate any hard constraints are first eliminated. Then, the utility of the remaining policies is calculated, and recommendations are generated for human confirmation. If any version of a scoring case changes during the experiment, or if the knowledge snapshot verification fails, the current experiment is immediately stopped and re-numbered for execution.
[0070] Experiment number patent-eval-001 used 200 fixed structured scoring cases, covering 71 intent scenarios. Each case executed four strategies, for a total of 800 trials. The initial success rate was 789. After excluding missing prompts and single-strategy rate limiting failures, 195 complete four-strategy pairing cases were obtained. The model was qwen3.6-flash, and the evaluation set hash and knowledge snapshot were fixed. Metrics included case recall, absolute scoring bias, perfect consistency rate, average / P95 RAG latency, and actual knowledge usage. The experimental results for the 195 complete pairing cases are shown in Table 4.
[0071] Table 4. Results of the four-strategy experiment with 195 complete paired cases.
[0072] Further, see Figure 6The chart showing a comparison of case recall, complete agreement rate, and P95 latency for a four-strategy approach demonstrates that, compared to the default baseline, the time-priority approach reduces the average actual knowledge usage by approximately 40.5%, lowers P95 RAG latency by approximately 25.8%, and decreases case recall by approximately 1.54 percentage points. This indicates that the system can quantify the "latency-recall" tradeoff under fixed knowledge conditions. The accuracy-priority approach exhibits an engineering trend of lower MAE scores and higher complete agreement rates; however, due to the limited sample size, this trend is not presented as having achieved statistical significance.
[0073] See Figure 7 The diagram illustrates the overall system architecture and data flow. The client or batch import end generates structured score submissions; the server performs normalization, Fc / Fv calculation, four types of submission judgment, version state transition, and vector admission control; the relational database stores cases, versions, audits, rulings, and traces; the vector storage only exposes current / allowed versions to online retrieval; the scoring model outputs model scores and explanations. Trace attribution, knowledge snapshots, and policy evaluation are preferred extension modules.
[0074] The method provided in this embodiment of the invention establishes a hierarchical representation of scoring context identity and scoring version identity, so that the scope of the project, query terms, intent scenario and ordered result items stably determine the scored object, and changes in manual scores or reasons do not change the context identity of the object.
[0075] Within the same scoring context, it definitively distinguishes between complete duplicates, identical score revisions, and dissimilar score conflicts, and enables the judgment results to directly drive version creation, unique current version switching, and vector admission, rather than merely generating prompts or hash identifiers.
[0076] In the event of a conflict, the conflicting version, source, and audit evidence are fully preserved, while preventing it from entering the online searchable vector, ensuring that the original valid version does not interrupt the service; and the relationship version state is kept consistent with the vector visibility through idempotent tasks and state verification.
[0077] By linking the actual hit case versions, model scores, human scores, and biases through traces, and comparing retrieval strategies under fixed evaluation sets, models, prompts, and knowledge snapshots, a reproducible, auditable, and manually verifiable basis for strategy selection is formed.
[0078] In summary, the present invention has the following technical effects: (1) By separating the scoring context identity and the "scoring version identity" through Fc and Fv, and taking the ordered search results as part of the context identity, complete duplication, same score revision and different score conflict can be distinguished according to deterministic rules, avoiding identity confusion caused by whole record hash or semantic similarity judgment.
[0079] (2) By using a continuous control chain of joint judgment result → version status → vector admission status, the conflicting versions remain blocked while preserving evidence and audit chain, and the original current / allowed versions continue to serve, thereby reducing the risk of conflict knowledge polluting online RAG and causing score drift.
[0080] (3) By using a unique current version constraint, an idempotent vector task carrying version_id, and a state review before execution, we prevent expired vectors from being re-uploaded by concurrent submissions or asynchronous retries; at the same time, we retain actual hit evidence with a trace of case ID + version ID to improve the consistency and traceability of version evolution.
[0081] (4) By using fixed evaluation sets, models, prompts, and knowledge snapshots for conditional evaluation, the differences in accuracy and latency of different retrieval strategies can be reproduced and audited. The completed paired experiments show that the timeliness-first strategy can reduce the amount of knowledge actually used and reduce the P95 RAG latency compared to the default baseline, providing a quantitative basis for project strategy selection.
[0082] In addition to the search scoring knowledge version control and vector admission method provided in the foregoing embodiments, this invention also provides a search scoring knowledge version control and vector admission system, see [link to relevant documentation]. Figure 8 The diagram shown illustrates the structure of a search scoring knowledge version control and vector admission system, indicating that the system mainly includes: The case splitting module 801 is used to split the received historical scoring cases to obtain structured scoring submissions.
[0083] The dual fingerprint generation module 802 is used to perform deterministic normalization processing on the structured score submission and generate a score context fingerprint and a score version fingerprint based on the normalized structured score submission.
[0084] The dual fingerprint determination module 803 is used to determine the submission type of each historical scoring case based on the scoring context fingerprint and the scoring version fingerprint.
[0085] The vector admission module 804 is used to create or reuse case subjects based on submission type, establish or update case version records, and set the status identifier and vector admission identifier of the corresponding version. It creates online vectors for versions with a status identifier of "currently valid" and a vector admission identifier of "allowed".
[0086] The search scoring knowledge version control and vector admission system provided in this embodiment of the invention separates the scoring context identity and scoring version identity of a scoring case through scoring context fingerprints and scoring version fingerprints. It then determines the submission type of each historical scoring case based on these fingerprints and creates or reuses the case subject and establishes or updates the case version record according to the submission type. This allows for the differentiation of different cases based on deterministic rules, avoiding redundant storage and duplicate vector writing caused by hashing methods or semantic similarity judgments of the entire record. Simultaneously, only versions with a current status mark indicating validity and a vector admission mark indicating permission are written to the vector storage. This ensures that conflicting versions remain blocked while preserving evidence and the audit chain, while versions with a current status mark indicating validity and a vector admission mark indicating permission continue to serve, reducing the risk of conflicting knowledge contaminating the online RAG and causing scoring drift.
[0087] In one implementation, the structured score submission includes the project scope, query terms, intent scenario, multiple search result entries in their original relative order, human score, and scoring reason; the aforementioned dual fingerprint generation module 802 is specifically used to: concatenate the normalized project scope, query terms, intent scenario, and each search result entry to generate a first summary value, which serves as the scoring context fingerprint; based on the scoring context fingerprint, and combining the normalized human score and scoring reason, perform a deterministic summary operation to generate a second summary value, which serves as the scoring version fingerprint.
[0088] In one implementation, the dual fingerprint determination module 803 is specifically used to: determine the submission type as a new case, a complete duplicate, a revision with the same score, or a conflicting score, based on whether there is the same scoring context fingerprint within the project scope and whether there is the same scoring version fingerprint.
[0089] In one implementation, the dual-fingerprint determination module 803 is specifically used to: if there is no identical scoring context fingerprint within the project scope, determine that the submission type of the historical scoring case is a new case; if there is an identical scoring context fingerprint within the project scope and the scoring version fingerprint of any historical version is the same as the scoring version fingerprint of the historical scoring case, determine that the submission type of the historical scoring case is a complete duplicate; if there is an identical scoring context fingerprint within the project scope, no identical scoring version fingerprint, and the current human score is equal to the current valid version's human score, determine that the submission type of the historical scoring case is a same-score revision; if there is an identical scoring context fingerprint within the project scope, no identical scoring version fingerprint, and the current human score is not equal to the current valid version's human score, determine that the submission type of the historical scoring case is a dissimilar conflict.
[0090] In one implementation, the vector admission module 804 is specifically used to: set the status identifier of the version corresponding to the new case to currently valid and the vector admission identifier to allowed; for completely duplicated versions, no new version is added, and the status identifier and vector admission identifier of the corresponding version are not changed; set the status identifier of the version corresponding to the same revision to currently valid, the original version to replaced, and the vector admission identifier of the new version to allowed; set the status identifier of the version corresponding to the dissimilar conflict to conflict pending adjudication and the vector admission identifier to blocked.
[0091] In one implementation, the system further includes a RAG enhanced scoring module, used to: generate a unique tracking identifier during the RAG enhanced scoring process, and associate it with the scoring request, the retrieved case version identifier, the scoring model output, and the human score; wherein, the tracking identifier is used to fix the knowledge snapshot, bind the model and prompt word versions in the strategy evaluation, and ensure that the case version used in the strategy evaluation is consistent with the actual searchable version in the vector storage.
[0092] In one implementation, the system further includes a strategy evaluation module, configured to: before the strategy evaluation service is started, filter scoring case versions whose status is currently valid and whose vector access is allowed, and arrange the filtered scoring case versions in ascending order of case identifiers; extract the scoring context fingerprint and scoring version fingerprint of each scoring case version; and generate a knowledge snapshot summary; respond to the evaluation request, perform stratified sampling of the currently valid and allowed access scoring case versions according to a preset intent scenario and human standard score, and fix the version identifier, search result item order, human standard score, random seed, and model identifier of the selected cases; and under the same knowledge snapshot summary, the same model identifier, the same prompt word configuration version, and the same RAG service configuration parameters. The system executes multiple preset search strategies, recording the actual case version identifier, ranking, similarity, scoring deviation, RAG service response latency, and the number of knowledge items actually retrieved for each strategy. For each strategy, based on the project's pre-defined hard constraints, it verifies whether the recorded indicators exceed the corresponding hard constraint thresholds. If any indicator exceeds the corresponding hard constraint threshold, the strategy is marked as unqualified. For strategies marked as qualified, a weighted calculation is performed on each verified indicator based on preset weights to generate a comprehensive utility value. The strategy recommendation result with the highest comprehensive utility value and in a pending confirmation state is output to the authorized personnel's operation interface, and the default search strategy configuration in the production environment is updated after receiving confirmation instructions from the authorized personnel.
[0093] It should be noted that the system provided in this embodiment of the invention has the same implementation principle and technical effects as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the system embodiment can be referred to the corresponding content in the aforementioned method embodiment. The specific numerical values provided in the implementation of this invention are merely exemplary and are not intended to limit the scope of the invention.
[0094] This invention also provides an electronic device, specifically, the electronic device includes a processor and a storage device; the storage device stores a computer program, and the computer program, when run by the processor, executes the method described in any of the above embodiments.
[0095] Figure 9 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. The electronic device 100 includes: a processor 90, a memory 91, a bus 92, and a communication interface 93. The processor 90, the communication interface 93, and the memory 91 are connected through the bus 92. The processor 90 is used to execute executable modules, such as computer programs, stored in the memory 91.
[0096] The memory 91 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 93 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc.
[0097] Bus 92 can be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 9 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.
[0098] The memory 91 is used to store programs. After receiving an execution instruction, the processor 90 executes the programs. The method executed by the device for defining the flow process disclosed in any of the foregoing embodiments of the present invention can be applied to the processor 90 or implemented by the processor 90.
[0099] The processor 90 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of the processor 90 or by instructions in software form. The processor 90 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in memory 91. Processor 90 reads the information in memory 91 and, in conjunction with its hardware, completes the steps of the above method.
[0100] The computer program product of the readable storage medium provided in the embodiments of the present invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the foregoing method embodiments. For specific implementation, please refer to the foregoing method embodiments, which will not be repeated here.
[0101] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned 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.
[0102] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A search scoring knowledge version control and vector admission method, characterized in that, include: The received historical scoring cases are broken down to obtain structured scoring submissions; The structured score submission is subjected to deterministic normalization processing, and a score context fingerprint and a score version fingerprint are generated based on the normalized structured score submission. Based on the scoring context fingerprint and the scoring version fingerprint, the submission type of each historical scoring case is determined; Based on the submission type, create or reuse the case subject, establish or update the case version record, and set the status identifier and vector admission identifier for the corresponding version. Create online vectors for versions whose status identifier is currently valid and whose vector admission identifier is allowed.
2. The method according to claim 1, characterized in that, The structured scoring submission includes the project scope, query terms, intent scenario, multiple search results entries in their original relative order, human score, and scoring reason; Based on the normalized structured score submission, a score context fingerprint and a score version fingerprint are generated, including: The standardized project scope, query terms, intent scenarios, and each search result item are concatenated to generate a first summary value, which serves as the scoring context fingerprint. Based on the scoring context fingerprint, combined with the normalized human score and the scoring reason, a deterministic summary operation is performed to generate a second summary value, which serves as the scoring version fingerprint.
3. The method according to claim 1, characterized in that, Based on the scoring context fingerprint and the scoring version fingerprint, the submission type of each historical scoring case is determined, including: Based on whether the same scoring context fingerprint exists within the project scope, and whether the same scoring version fingerprint exists, the submission type is sequentially determined as a new case, a complete duplicate, a revision with the same score, or a conflict with a different score.
4. The method according to claim 3, characterized in that, The submission types are determined sequentially as follows: new case, complete duplicate, same score revision, or different score conflict, including: If no identical scoring context fingerprint exists within the project scope, the submission type of the historical scoring case is determined to be a new case; If the same scoring context fingerprint exists within the project scope and any historical version of the scoring version fingerprint is the same as the scoring version fingerprint of the historical scoring case, then the submission type of the historical scoring case is determined to be completely duplicated. If there are identical scoring context fingerprints within the project scope, no identical scoring version fingerprints, and the current human score is equal to the current valid version's human score, then the submission type of the historical scoring case is determined to be a revision with the same score. If there are identical scoring context fingerprints within the project scope, but no identical scoring version fingerprints, and the current human score is not equal to the current valid version's human score, then the submission type of the historical scoring case is determined to be a heterogeneous conflict.
5. The method according to claim 3, characterized in that, Configure the corresponding version's status flags and vector admission flags, including: Set the status flag of the version corresponding to the new case to currently valid, and set the vector access flag to allowed; For the completely duplicated version, no new version is added, and the status identifier and vector admission identifier of the corresponding version are not changed; Set the status identifier of the version corresponding to the same revision to currently valid, set the original version to replaced, and set the vector access identifier of the new version to allowed. Set the status flag of the version corresponding to the heterogeneous conflict to conflict pending adjudication, and set the vector admission flag to blocked.
6. The method according to claim 1, characterized in that, The method further includes: During the RAG enhanced scoring process, a unique tracking identifier is generated and associated with the scoring request, the retrieved case version identifier, the scoring model output, and the human score. The tracking identifier is used to fix the knowledge snapshot, bind the model and prompt word versions in the strategy evaluation, and ensure that the case version used in the strategy evaluation is consistent with the actual searchable version in the vector storage.
7. The method according to claim 1, characterized in that, The method further includes: Before the strategy evaluation service is launched, the scoring case versions with the status identifier of "currently valid" and the vector admission identifier of "allowed" are selected, and the selected scoring case versions are arranged in ascending order of case identifier. The scoring context fingerprint and scoring version fingerprint of each scoring case version are extracted to generate a knowledge snapshot summary. In response to the evaluation request, the system performs stratified sampling of the currently valid and permitted scoring case versions based on the preset intent scenario and human standard score, and fixes the version identifier, search result item order, human standard score, random seed and model identifier of the selected cases. Under the same knowledge snapshot summary, the same model identifier, the same prompt word configuration version, and the same RAG service configuration parameters, multiple preset search strategies are executed respectively. For each strategy, the actual hit case version identifier, ranking, similarity, scoring deviation, RAG service response latency, and the number of knowledge items actually involved in the search are recorded. For each strategy, based on the pre-set hard constraints of the project, each recorded indicator is checked to see if it exceeds the corresponding hard constraint threshold. If any indicator exceeds the corresponding hard constraint threshold, the strategy is marked as unqualified. For the strategy marked as qualified, a weighted calculation is performed on each verified indicator based on preset weights to generate a comprehensive utility value; The strategy recommendation result with the highest comprehensive utility value and in the pending confirmation state is output to the authorized personnel's operation interface, and the default search strategy configuration in the production environment is updated after receiving the confirmation instruction from the authorized personnel.
8. A search scoring knowledge version control and vector admission system, characterized in that, include: The case splitting module is used to split the received historical scoring cases to obtain structured scoring submissions; The dual fingerprint generation module is used to perform deterministic normalization processing on the structured score submission, and generate a score context fingerprint and a score version fingerprint based on the normalized structured score submission. The dual fingerprint determination module is used to determine the submission type of each historical scoring case based on the scoring context fingerprint and the scoring version fingerprint. The vector admission module is used to create or reuse case subjects based on the submission type, establish or update case version records, and set the status identifier and vector admission identifier of the corresponding version. It creates online vectors for versions whose status identifier is currently valid and whose vector admission identifier is allowed.
9. An electronic device, characterized in that, The method includes a processor and a memory, the memory storing computer-executable instructions executable by the processor, the processor executing the computer-executable instructions to implement the steps of the method according to any one of claims 1 to 7.
10. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program, when run by a processor, performs the steps of the method described in any one of claims 1 to 7.
