An intelligent cockpit dynamic knowledge service system based on angel star

Abstract

The application discloses an intelligent cockpit dynamic knowledge service system based on Angel star. In the scheme, the interaction data is structured and buried at the vehicle end, and is reported after being filtered by a confidence double threshold, so that bandwidth consumption is reduced, original corpus is not included, and privacy compliance is ensured. The cloud end merges the question and answer into an intention cluster through semantic clustering, generates a cluster fingerprint, and deploys the light-weight coded cluster fingerprint to the vehicle end, so that the local intention attribution of the original corpus is realized without leaving the vehicle. The knowledge operation and verification module automatically generates and updates work orders according to priorities, and generates test cases based on the intention cluster question, so as to ensure revision quality. The distribution module issues knowledge package snapshots and cluster fingerprints, and the vehicle end intelligent agent updates the knowledge base and responds locally according to the knowledge package snapshots and cluster fingerprints, and feedback is changed into feedback indicators, so that the operation cost is greatly reduced. The effect backflow verification module compares the feedback indicators with historical benchmarks based on the cluster fingerprint, forms a closed loop, makes the knowledge updating process traceable, realizes data-driven autonomous iteration, and improves reliability and efficiency.

Description

Technical Field

[0001] This application relates to the field of intelligent connected vehicle technology, and in particular to an intelligent cockpit dynamic knowledge service system based on OnStar. Background Technology

[0002] In in-vehicle voice interaction scenarios, the vehicle's intelligent agent relies on a customer service knowledge base to answer user questions. Currently, major automakers have deployed two systems: an in-vehicle intelligent agent and a customer service knowledge base. However, these two systems have long operated in isolation: the vehicle generates a large amount of question-and-answer interaction data daily, but this data is not structured and collected; updates to the customer service knowledge base rely on manual review of customer complaint tickets, and it typically takes several weeks for hot issues to go from generation to knowledge deployment; knowledge modifications take effect immediately, lacking a vehicle-side effectiveness verification process. These data flow gaps prevent the formation of an effective closed loop between vehicle-side problem feedback and cloud-based knowledge updates.

[0003] Furthermore, existing verification schemes typically employ a canary release + statistical testing paradigm: first, the updated knowledge is distributed to a subset of vehicles, and the response performance metrics for the same intent cluster on these vehicles are collected and statistically compared with the baseline values ​​before the update. If the metrics show a significant improvement, the entire dataset is released. However, this paradigm has certain drawbacks in practical applications: knowledge package updates are often accompanied by semantic model version upgrades, causing the same semantic intent to be assigned to different clusters before and after the update, resulting in intra-cluster member drift. At this point, the samples from the canary release period and the baseline period no longer belong to the same semantic population, the identically distributed hypothesis of the statistical test is not valid, and the test results cannot reflect the true effect of the knowledge update, seriously affecting the user experience.

[0004] How to reliably measure the effectiveness of knowledge updates on the same semantic entity in order to achieve complete closed-loop management of knowledge services is a technical problem that urgently needs to be solved. Summary of the Invention

[0005] To address the aforementioned issues, this application provides an intelligent cockpit dynamic knowledge service system based on OnStar, aiming to reliably measure the effectiveness of knowledge updates on the same semantic entity in order to achieve complete closed-loop management of knowledge services.

[0006] The embodiments of this application disclose the following technical solutions: The first aspect of this application provides an intelligent cockpit dynamic knowledge service system based on OnStar, the system comprising: a vehicle-side interactive perception module, a cloud-based demand mining module, a knowledge operation and verification module, a distribution module, a vehicle-side intelligent agent, and an effect feedback verification module; The vehicle-side interaction perception module is used to perform structured data collection based on the question-and-answer data during the interaction process, obtain structured data collection data, and filter the structured data collection data according to the confidence level dual threshold diversion strategy to obtain the events to be optimized and report them to the cloud-based demand mining module; the structured data collection data does not contain the original corpus. The cloud-based demand mining module is used to perform semantic clustering on the reported events to be optimized, obtain multiple intent clusters, generate a corresponding cluster fingerprint for each intent cluster, and output the cluster fingerprint after lightweight encoding according to the vehicle-side resource constraints; the cluster fingerprint is used to uniquely identify the semantic identity of the corresponding intent cluster. The knowledge operation and verification module is used to generate corresponding knowledge update work orders based on each intent cluster, revise the knowledge items corresponding to each knowledge update work order according to the priority of the knowledge update work orders, generate corresponding test case sets based on the question of each intent cluster, verify the corresponding revised knowledge items using the test case sets, and package the verified revised knowledge items into the corresponding knowledge package version snapshot; the knowledge update work order includes the question of the corresponding intent cluster, sample size, and knowledge item location information. The distribution module is used to send the snapshot of each knowledge package version and the cluster fingerprint corresponding to each intent cluster after lightweight encoding to the vehicle-side intelligent agent; The vehicle-side intelligent agent is used to update the vehicle-side knowledge base based on the snapshots of each knowledge package version, determine the answer to the user's current question locally on the vehicle-side based on the updated vehicle-side knowledge base, and, without uploading the original corpus, determine the intent attribution of the current question and answer content based on the cluster fingerprints corresponding to each intent cluster after lightweight encoding, and generate corresponding effect feedback indicators based on the judgment results and report them to the effect feedback verification module. The effect feedback verification module is used to statistically compare the effect feedback indicators of the corresponding intent cluster with the historical benchmark value based on the fingerprint of each cluster, and determine the actual effect of the corresponding knowledge package version snapshot based on the comparison results.

[0007] Optionally, the system further includes: a prediction module; the prediction module includes: The change function point extraction unit is used to obtain the vehicle software version release plan and extract the change function points based on the baseline change list of the vehicle software version release plan. The prediction unit is used to determine the corresponding prediction intent cluster, as well as the prediction knowledge entry and cluster fingerprint corresponding to each prediction intent cluster, based on the changed function point. The prediction result distribution unit is used to distribute the prediction knowledge entries and cluster fingerprints corresponding to each prediction intention cluster to the vehicle along with the vehicle software version in the vehicle software version release plan.

[0008] Optionally, the prediction unit is used for: The modified function points are assigned to the corresponding function domains, and each function domain is queried to see if there is a corresponding intent cluster. If no corresponding intent cluster exists for any functional domain, multiple prediction question statements are generated based on the functional description of the corresponding changed functional point, and then processed by the intent cluster. Figure 1 After consistency refeedback verification, the corresponding predicted intent cluster is obtained, and the corresponding predicted knowledge entries and cluster fingerprints are generated based on the predicted intent cluster; If any functional domain has a corresponding intent cluster, then the intent cluster is incrementally updated based on the changed functional point; the incremental update includes adding new question variants and corresponding knowledge entries to the intent cluster, and updating the cluster fingerprint of the intent cluster.

[0009] Optionally, the prediction module further includes: a prediction result verification unit; The prediction result verification unit is used to collect the vehicle-side performance feedback indicators for each prediction intent cluster based on the cluster fingerprint corresponding to each prediction intent cluster. If the semantic similarity between the actual problem statement of any predicted intent cluster and the predicted problem statement of the corresponding predicted intent cluster is greater than or equal to a preset similarity threshold, then the corresponding predicted intent cluster will be converted into an intent cluster. If the semantic similarity between the actual problem statement of any predicted intent cluster and the predicted problem statement of the corresponding predicted intent cluster is less than a preset similarity threshold, or if no actual sample appears for a consecutive preset number of rounds, the corresponding predicted intent cluster will be taken offline and archived.

[0010] Optionally, the effect reflow verification module is used for: The effect feedback index of each intent cluster was statistically significant compared with the historical baseline, and the corresponding test results were obtained. If the corresponding performance feedback metric is greater than the historical baseline value, and the corresponding test result indicates that the difference between the corresponding performance feedback metric and the historical baseline value is statistically significant, then the corresponding knowledge package version snapshot is valid; the knowledge package version snapshot includes the content of the knowledge item, version number, correspondence between the applicable vehicle model and the software version baseline, digital signature, and rollback pointer information; If the test result indicates that the corresponding effect feedback index is less than or equal to the historical benchmark value, or if the difference between the corresponding effect feedback index and the historical benchmark value is not statistically significant, then the system will automatically roll back to the previous knowledge package version.

[0011] Optionally, the vehicle-side interactive perception module is used for: If the confidence value of the answer is less than or equal to the first preset threshold, the corresponding question and answer data is reported as the event to be optimized; if the confidence value of the answer is greater than the first preset threshold and less than the second preset threshold, a clarifying follow-up question is initiated to the user. If the user's reply to the clarifying follow-up question indicates that the original question has been answered, it is not reported; otherwise, the corresponding question and answer data is reported as the event to be optimized; if the confidence value of the answer is greater than or equal to the second preset threshold, local anonymous counting is performed, and it is not reported.

[0012] The second aspect of this application provides a method for providing dynamic knowledge services for intelligent cockpits based on OnStar. This method is applied to the OnStar-based dynamic knowledge service system described in the first aspect, and includes: The vehicle-side interaction perception module performs structured data collection based on the question-and-answer data during the interaction process, obtains structured data collection data, and filters the structured data collection data according to the confidence level dual threshold diversion strategy to obtain the events to be optimized and report them to the cloud-based demand mining module; the structured data collection data does not contain the original corpus. The cloud-based demand mining module performs semantic clustering on the reported events to be optimized, obtains multiple intent clusters, and generates a corresponding cluster fingerprint for each intent cluster. The cluster fingerprint is then output after lightweight encoding based on vehicle-side resource constraints. The cluster fingerprint is used to uniquely identify the semantic identity of the corresponding intent cluster. The knowledge operation and verification module generates corresponding knowledge update work orders based on each intent cluster, revises the knowledge items corresponding to each knowledge update work order according to the priority of the knowledge update work orders, generates corresponding test case sets based on the question of each intent cluster, uses the test case sets to verify the corresponding revised knowledge items, and packages the verified revised knowledge items into the corresponding knowledge package version snapshot; the knowledge update work order includes the question of the corresponding intent cluster, sample size, and knowledge item location information. The distribution module sends the version snapshots of each knowledge package and the cluster fingerprints corresponding to each intent cluster after lightweight encoding to the vehicle-side intelligent agent. The vehicle-side intelligent agent updates the vehicle-side knowledge base based on the snapshots of each knowledge package version, determines the answer to the user's current question locally on the vehicle-side based on the updated vehicle-side knowledge base, and performs intent attribution judgment on the current question and answer content based on the cluster fingerprints corresponding to each intent cluster after lightweight encoding without uploading the original corpus. Based on the judgment result, it generates corresponding effect feedback indicators and reports them to the effect feedback verification module. The effect feedback verification module statistically compares the effect feedback indicators of the corresponding intent cluster with the historical benchmark value based on the fingerprint of each cluster, and determines the actual effect of the corresponding knowledge package version snapshot based on the comparison results.

[0013] Optionally, the method further includes: Obtain the vehicle software version release plan, and extract the change function points based on the baseline change list of the vehicle software version release plan; Based on the changed function points, determine the corresponding prediction intent clusters, as well as the prediction knowledge entries and cluster fingerprints corresponding to each prediction intent cluster; The prediction knowledge entries and cluster fingerprints corresponding to each prediction intent cluster are distributed to the vehicle along with the vehicle software version in the vehicle software version release plan.

[0014] Optionally, determining the corresponding prediction intent cluster, and the prediction knowledge entry and cluster fingerprint corresponding to each prediction intent cluster, based on the changed function point, includes: The modified function points are assigned to the corresponding function domains, and each function domain is queried to see if there is a corresponding intent cluster. If no corresponding intent cluster exists for any functional domain, multiple prediction question statements are generated based on the functional description of the corresponding changed functional point, and then processed by the intent cluster. Figure 1 After consistency refeedback verification, the corresponding predicted intent cluster is obtained, and the corresponding predicted knowledge entries and cluster fingerprints are generated based on the predicted intent cluster; If any functional domain has a corresponding intent cluster, then the intent cluster is incrementally updated based on the changed functional point; the incremental update includes adding new question variants and corresponding knowledge entries to the intent cluster, and updating the cluster fingerprint of the intent cluster.

[0015] Optionally, after distributing the prediction knowledge entries and cluster fingerprints corresponding to each prediction intent cluster to the vehicle along with the vehicle software version in the vehicle software version release plan, the method further includes: Based on the cluster fingerprint corresponding to each predicted intent cluster, collect the vehicle-side performance feedback indicators for each predicted intent cluster. If the semantic similarity between the actual problem statement of any predicted intent cluster and the predicted problem statement of the corresponding predicted intent cluster is greater than or equal to a preset similarity threshold, then the corresponding predicted intent cluster will be converted into an intent cluster. If the semantic similarity between the actual problem statement of any predicted intent cluster and the predicted problem statement of the corresponding predicted intent cluster is less than a preset similarity threshold, or if no actual sample appears for a consecutive preset number of rounds, the corresponding predicted intent cluster will be taken offline and archived.

[0016] Compared with the prior art, this application has the following beneficial effects: The intelligent cockpit dynamic knowledge service system based on OnStar provided in this application embodiment uses a vehicle-side interaction perception module to perform structured data embedding for question-and-answer interaction data and combines a confidence-based dual-threshold diversion strategy for reporting and filtering, effectively reducing the bandwidth consumption of the vehicle-cloud link. Simultaneously, the structured data embedding does not contain the original corpus, ensuring user privacy data compliance from the source. The cloud-based demand mining module uses semantic clustering to group massive, scattered question-and-answer events into intent clusters and generates cross-version fixed cluster fingerprints, providing stable semantic anchors for subsequent effect feedback. This solves the technical problem of unreliable attribution due to measurement object drift caused by model version iteration. The cluster fingerprints are lightweightly encoded according to vehicle-side resource constraints, enabling deployment in vehicle-side environments with limited storage capacity and computing power. This achieves local intent attribution judgment on the vehicle side, completing attribution without the original corpus leaving the vehicle, meeting regulatory requirements for in-vehicle data processing, and contributing to improved user interaction experience. The knowledge operation and verification module automatically generates knowledge update work orders based on priority, ensuring that high-frequency issues are resolved first. It automatically generates and verifies test case sets based on the question phrasing of each intent cluster, ensuring that test cases are semantically consistent with real-world scenarios. This provides verifiable quality control for knowledge revisions, significantly improving the success rate of deployment. The distribution module sends knowledge package version snapshots and cluster fingerprints to the vehicle, enabling knowledge updates to be deployed gradually in a controlled, gray-scale manner, reducing the risk of large-scale releases. The vehicle-side intelligent agent updates the vehicle's knowledge base based on the knowledge package version snapshot and determines the answer to the user's current question locally on the vehicle, ensuring that knowledge updates take effect immediately on the vehicle without waiting for cloud interaction, resulting in faster response times. Simultaneously, it completes intent attribution determination based on cluster fingerprints without uploading original corpora, changing reported data from raw questions to structured feedback indicators, significantly reducing the amount of reported data and substantially lowering operating costs. The effect feedback verification module uses cluster fingerprinting to statistically compare the effect feedback indicators of the corresponding intent cluster with historical benchmark values, forming a complete closed-loop link. This ensures that every knowledge update is automatically verified by real key performance indicators on the vehicle side, transforming knowledge services from passive manual maintenance to data-driven closed-loop autonomy. This enables the entire knowledge update process to be traceable, measurable, and controllable, improving the iteration efficiency and reliability of knowledge services. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A schematic diagram of the structure of an intelligent cockpit dynamic knowledge service system based on OnStar provided for an embodiment of this application; Figure 2 A flowchart illustrating a dynamic knowledge service method for an intelligent cockpit based on OnStar, provided as an embodiment of this application. Detailed Implementation

[0019] As described earlier, most major automakers have deployed two systems: an in-vehicle intelligent agent and a customer service knowledge base. However, these two systems have long operated in isolation: the vehicle generates a large amount of Q&A interaction data daily, but this data is not structured and collected; updates to the customer service knowledge base rely on manual review of customer complaint tickets, and it typically takes several weeks for hot issues to go from generation to knowledge deployment; knowledge changes take effect immediately, lacking a vehicle-side effectiveness verification process. These data flow gaps prevent the formation of an effective closed loop between vehicle-side issue feedback and cloud-based knowledge updates.

[0020] Furthermore, existing verification schemes typically employ a canary release + statistical testing paradigm: first, the updated knowledge is distributed to a subset of vehicles, and the response performance metrics for the same intent cluster on these vehicles are collected and statistically compared with the baseline values ​​before the update. If the metrics show a significant improvement, the entire dataset is released. However, this paradigm has certain drawbacks in practical applications: knowledge package updates are often accompanied by semantic model version upgrades, causing the same semantic intent to be assigned to different clusters before and after the update, resulting in intra-cluster member drift. At this point, the samples from the canary release period and the baseline period no longer belong to the same semantic population, the identically distributed hypothesis of the statistical test is not valid, and the test results cannot reflect the true effect of the knowledge update, seriously affecting the user experience.

[0021] In view of the above problems, this application proposes an intelligent cockpit dynamic knowledge service system based on OnStar. The system includes: a vehicle-side interaction perception module, a cloud-based demand mining module, a knowledge operation and verification module, a distribution module, a vehicle-side intelligent agent, and an effect feedback verification module. The vehicle-side interaction perception module is used to perform structured embedding based on question-and-answer data during the interaction process to obtain structured embedding data. It then filters the structured embedding data according to a dual-threshold confidence-based diversion strategy to obtain events to be optimized and reports them to the cloud-based demand mining module. The structured embedding data does not contain raw corpus. The cloud-based demand mining module is used to perform semantic clustering on the reported events to be optimized to obtain multiple intent clusters. It generates a corresponding cluster fingerprint for each intent cluster and outputs the cluster fingerprint after lightweight encoding according to vehicle-side resource constraints. The cluster fingerprint is used to uniquely identify the semantic identity of the corresponding intent cluster. The knowledge operation and verification module is used to generate corresponding knowledge update work orders based on each intent cluster, revise the knowledge entries corresponding to each knowledge update work order according to the priority of the knowledge update work orders, and... Based on the question format of each intent cluster, a corresponding test case set is generated. The test case set is used to verify the corresponding revised knowledge items. The verified revised knowledge items are packaged into a corresponding knowledge package version snapshot. The knowledge update work order includes the question format, sample size, and knowledge item location information for the corresponding intent cluster. The distribution module is used to send the knowledge package version snapshots and the cluster fingerprints corresponding to each intent cluster after lightweight encoding to the vehicle-side intelligent agent. The vehicle-side intelligent agent is used to update the vehicle-side knowledge base based on the knowledge package version snapshots, determine the answer to the user's current question locally on the vehicle based on the updated vehicle-side knowledge base, and, without uploading the original corpus, determine the intent attribution of the current question and answer content based on the cluster fingerprints corresponding to each intent cluster after lightweight encoding. Based on the determination result, the agent generates corresponding effect feedback indicators and reports them to the effect feedback verification module. The effect feedback verification module is used to statistically compare the effect feedback indicators of the corresponding intent clusters with historical benchmark values ​​based on each cluster fingerprint, and determine the actual effect of the corresponding knowledge package version snapshot based on the comparison results.

[0022] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0023] See Figure 1 This figure is a schematic diagram of the structure of an intelligent cockpit dynamic knowledge service system based on OnStar, provided in an embodiment of this application. Figure 1As shown, the system includes: a vehicle-side interaction perception module, a cloud-based demand mining module, a knowledge operation and verification module, a distribution module, a vehicle-side intelligent agent, and an effect feedback verification module.

[0024] The vehicle-side interaction perception module is used to perform structured tracking based on the question-and-answer data during the interaction process, obtain structured tracking data, and filter the structured tracking data according to the confidence level dual threshold diversion strategy to obtain the events to be optimized and report them to the cloud-based demand mining module; the structured tracking data does not contain the original corpus.

[0025] In one feasible implementation, the vehicle-side interactive sensing module is used for: If the confidence value of the answer is less than or equal to the first preset threshold, the corresponding question and answer data is reported as the event to be optimized; if the confidence value of the answer is greater than the first preset threshold and less than the second preset threshold, a clarifying follow-up question is initiated to the user. If the user's reply to the clarifying follow-up question indicates that the original question has been answered, it is not reported; otherwise, the corresponding question and answer data is reported as the event to be optimized; if the confidence value of the answer is greater than or equal to the second preset threshold, local anonymous counting is performed, and it is not reported.

[0026] Here, the first preset threshold is lower than the second preset threshold. The selection criteria are as follows: Samples from the past six months of in-vehicle voice recognition question-and-answer logs, manually labeled with "whether the answer was actually accepted by the user," are used to plot the precision-recall curve of the confidence score. The first preset threshold is the confidence value corresponding to a recall rate of 95%, and the second preset threshold is the confidence value corresponding to a precision rate of 98%. For example, the first preset threshold is set to 0.60, and the second preset threshold is set to 0.85. These two thresholds are not static; they are recalculated and calibrated according to the same standards each time the natural language processing model is updated.

[0027] By setting a first preset threshold and a second preset threshold, the confidence level of the responses is processed in three levels: low confidence is reported directly, medium confidence requires a clarifying follow-up question, and the decision to report is based on the user's response; high confidence only requires local anonymous counting and is not reported. Compared to the full reporting scheme, this embodiment significantly reduces the bandwidth consumption of vehicle-to-cloud communication while ensuring the integrity of demand recovery. Compared to a simple threshold filtering scheme, this embodiment avoids the loss of effective interactions in the intermediate confidence level range through a clarifying follow-up question mechanism, achieving a better balance between bandwidth saving and the integrity of demand recovery, and effectively solving the technical problem of limited bandwidth resources on the vehicle side.

[0028] The cloud-based demand mining module is used to perform semantic clustering on the reported events to be optimized, obtain multiple intent clusters, and generate a corresponding cluster fingerprint for each intent cluster. The cluster fingerprint is output after lightweight encoding based on vehicle-side resource constraints. The cluster fingerprint is used to uniquely identify the semantic identity of the corresponding intent cluster.

[0029] The vehicle-side resource constraints include storage capacity limitations and computation latency limitations.

[0030] The goal of the encoding process is to compress the cluster fingerprint to a size that can be stored on the vehicle side, while ensuring that intent attribution determination can be completed within 10ms latency on the vehicle side. Specific encoding methods include, but are not limited to, the following: quantization encoding: quantizing a high-dimensional floating-point vector into a low-bit integer representation, such as quantizing a 32-bit floating-point vector into an 8-bit integer; dimensionality reduction encoding: using dimensionality reduction methods such as principal component analysis or autoencoders to retain the main semantic information while reducing dimensionality; hash encoding: mapping the semantic encoding to a fixed-length hash value.

[0031] The size of the encoded cluster fingerprint must meet the vehicle-side storage constraints. For example, in a fleet of 5 million vehicles, assuming that each vehicle needs to store thousands of cluster fingerprints on average, each cluster fingerprint occupies about several hundred bytes after encoding, and the total storage space is within 2MB.

[0032] By using lightweight encoding of cluster fingerprints, high-dimensional semantic vectors are compressed into low-dimensional compact encodings, enabling cluster fingerprints to be stored locally on the vehicle within a few MB of storage space and to complete intent attribution determination within millisecond latency. At the same time, it achieves the data compliance requirement that the original corpus does not leave the vehicle—the vehicle only reports structured backflow data such as cluster identifiers, knowledge package version numbers, and interaction counts. This fundamentally solves the quadruple constraints of vehicle storage, computing power, bandwidth, and data compliance, making local intent attribution determination and effect attribution verification based on cluster fingerprints possible on mass-produced vehicles.

[0033] The cloud-based demand mining service performs vectorized semantic clustering on the interaction samples reported by the vehicle, grouping user questions with different expressions but consistent semantics into the same intent cluster. Semantic clustering is implemented based on a semantic embedding model, which converts the text of user questions into high-dimensional vector representations, and then uses clustering algorithms, such as density-based clustering or hierarchical clustering, to group semantically similar vectors together.

[0034] A cross-version fixed semantic identifier is generated for each intent cluster. In this embodiment, the semantic identifier is specifically a cluster fingerprint, generated by combining the semantic codes of all anchor samples under that intent cluster. Anchor samples are representative samples selected from each intent cluster, covering the main semantic variants of that cluster. Semantic encoding is obtained through a preset semantic embedding model, mapping the anchor samples to fixed-dimensional vector representations. The semantic codes of multiple anchor samples are combined, such as through concatenation, hash aggregation, or other encoding methods, to form the cluster fingerprint. This cluster fingerprint serves as the identity identifier of the intent cluster and remains unchanged when the knowledge package version is updated.

[0035] The knowledge operation and verification module is used to generate corresponding knowledge update work orders based on each intent cluster, revise the knowledge items corresponding to each knowledge update work order according to the priority of the knowledge update work orders, generate corresponding test case sets based on the question of each intent cluster, verify the corresponding revised knowledge items using the test case sets, and package the verified revised knowledge items into the corresponding knowledge package version snapshot; the knowledge update work order includes the question of the corresponding intent cluster, sample size, and knowledge item location information.

[0036] The knowledge package version snapshot here includes the content of the knowledge entries, version number, applicable vehicle models and the correspondence between the software version baseline, digital signature and rollback pointer information.

[0037] The priority of knowledge update work orders is determined by a weighted calculation based on three dimensions: the popularity of the intent cluster, the scope of impact, and the severity. Issues with high popularity, wide impact, and high severity are automatically assigned as expedited work orders, reducing the time from discovery to deployment of common issues from weeks to days. Popularity represents the number of samples per unit time and the month-on-month change trend; the scope of impact represents the number of affected vehicles on the network and the number of different brands and models affected; severity represents whether it involves safety-related areas and whether there is negative public opinion.

[0038] Test case sets can be generated through the following paths: Path 1: Prepare a thesaurus and a pool of colloquial expression templates for high-frequency functional domains in the vehicle beforehand. Break down the questions in each intent cluster into intent slots, function slots, and condition slots. Generate different question variations in batches using methods such as slot replacement, word order adjustment, and component deletion. For multi-turn dialogues, generate follow-up questions, counter-questions, and conditional supplementation based on the designed dialogue state machine template. Path 2: Use the company's proprietary large language model, using representative questioning methods from each intent cluster as seeds, to generate different questioning methods and corresponding multi-turn dialogue paths based on given prompts.

[0039] Test cases generated using any of the above methods must be reviewed. Figure 1 Consistency re-feedback verification and deduplication coverage verification. Figure 1Consistency refeeding verification involves refeeding candidate variants into the same natural language understanding engine on the vehicle side for recognition. Variants with the same result as the original intent are retained, while those with different results are deleted. Deduplication coverage verification ensures that variants with semantic similarity greater than or equal to 0.9 are deduplicated, and that the variant set covers 100% of the original intent cluster. Optimized knowledge entries must pass the overall test on the regression test set (e.g., accuracy exceeding 95%) before entering the release candidate pool.

[0040] The distribution module is used to send the version snapshots of each knowledge package and the cluster fingerprints corresponding to each intent cluster after lightweight encoding to the vehicle-side intelligent agent.

[0041] The encoded cluster fingerprint is sent to the vehicle via the SOTA channel along with the knowledge package version snapshot and stored locally on the vehicle. When the knowledge package version is updated, only the changed cluster fingerprint is synchronized to the vehicle to reduce the amount of data transmitted.

[0042] The vehicle-side intelligent agent is used to update the vehicle-side knowledge base based on the version snapshots of each knowledge package, determine the answer to the user's current question locally on the vehicle based on the updated vehicle-side knowledge base, and, without uploading the original corpus, determine the intent attribution of the current question and answer content based on the cluster fingerprints corresponding to each intent cluster after lightweight encoding, generate corresponding effect feedback indicators based on the judgment results, and report them to the effect feedback verification module.

[0043] The performance feedback metrics are structured feedback data, which include at least cluster identifiers, knowledge package version numbers, and number of interactions.

[0044] During each user question-and-answer interaction, the vehicle-side intelligent agent determines intent attribution based on locally cached cluster fingerprints. Specifically, the user's question is converted into a vector representation using a natural language understanding model deployed on the vehicle, and then matched against the locally stored cluster fingerprints to determine the intent cluster to which the question belongs. This process is completed locally on the vehicle, without uploading the original corpus to the cloud, with a latency of no more than 10ms. When the vehicle is in a network-free environment (such as underground parking garages or tunnels with weak networks), the vehicle still uses the locally cached cluster fingerprints for attribution determination and data logging. Incremental structured data is added after network recovery to ensure uninterrupted attribution. Because the structured data does not contain the original interaction corpus, the original voice or text information is not transmitted from the vehicle, thus meeting data compliance requirements.

[0045] The effect feedback verification module is used to statistically compare the effect feedback indicators of the corresponding intent cluster with the historical benchmark values ​​based on the fingerprint of each cluster, and determine the actual effect of the corresponding knowledge package version snapshot based on the comparison results.

[0046] In one feasible implementation, the effect feedback verification module is used to: The effect feedback index of each intent cluster was statistically significant compared with the historical baseline, and the corresponding test results were obtained. If the corresponding performance feedback metric is greater than the historical baseline value, and the corresponding test result indicates that the difference between the corresponding performance feedback metric and the historical baseline value is statistically significant, then the corresponding knowledge package version snapshot is valid; the knowledge package version snapshot includes the content of the knowledge item, version number, correspondence between the applicable vehicle model and the software version baseline, digital signature, and rollback pointer information; If the test result indicates that the corresponding effect feedback index is less than or equal to the historical benchmark value, or if the difference between the corresponding effect feedback index and the historical benchmark value is not statistically significant, then the system will automatically roll back to the previous knowledge package version.

[0047] Verified knowledge updates are packaged into knowledge package version snapshots and distributed in batches by vehicle type, region, and configuration via the over-the-air upgrade channel. During the gray-scale release, performance metrics (intent hit rate, first-time resolution rate, referral rate to human agent, and user negative feedback rate) for the same intent cluster on the vehicle side are continuously collected and statistically significant compared with the baseline values ​​for the same cluster in the first seven days of the gray-scale release.

[0048] The observation period is 7 calendar days by default, and the minimum effective sample size is 1000 effective interactions for this intent cluster. Two ratio Z-tests are performed on the first-time resolution rate during the gray-scale and baseline periods, with a significance level of 0.05. The target conditions must be met simultaneously: a statistically significant increase in the first-time resolution rate (p-value less than 0.05), an improvement in point estimation exceeding 5 percentage points, and no deterioration in safety-related indicators. Rollback is triggered if any one of the following conditions is met: a decrease in the hit rate of safety-related functional domains, a statistically significant increase in the user negative feedback rate exceeding 2 percentage points, or failure to reach the target value by the end of the observation period. When a rollback is triggered, the system automatically rolls back the vehicle-side knowledge to the previous version snapshot.

[0049] Using real key performance indicators from the vehicle side as the basis for release, combined with statistical significance testing, ensures the reliability of the judgment conclusions and effectively avoids misjudgment and over-release due to sampling errors. At the same time, by using information such as version number, applicable vehicle model and software version baseline correspondence, digital signature and rollback pointer contained in the knowledge package version snapshot, the accuracy and traceability of rollback operations are ensured. Every knowledge update is tested and verified, and unqualified updates are automatically rolled back, which greatly improves the reliability and controllability of knowledge service iteration.

[0050] Since the above content is a passive response update based on real interactive feedback, a knowledge pre-update will be performed before the vehicle software version is released, as follows: In one feasible implementation, the system further includes: a prediction module; the prediction module includes: The change function point extraction unit is used to obtain the vehicle software version release plan and extract the change function points based on the baseline change list of the vehicle software version release plan. The prediction unit is used to determine the corresponding prediction intent cluster, as well as the prediction knowledge entry and cluster fingerprint corresponding to each prediction intent cluster, based on the changed function point. The prediction result distribution unit is used to distribute the prediction knowledge entries and cluster fingerprints corresponding to each prediction intention cluster to the vehicle along with the vehicle software version in the vehicle software version release plan.

[0051] The changes to functional points include information on adding, deleting, and modifying functional points. The change list can come from API calls or data pushes from the engineering change management system, and includes information such as the function description, original operation path, new operation path, old function name, and new function name for each changed functional point.

[0052] The system automatically transforms the change list of the vehicle software release plan into predicted intent clusters and pre-prepared knowledge, ensuring that the knowledge is ready synchronously when the function goes live. This effectively avoids service gaps caused by knowledge gaps after the release, eliminates the problem of users being unable to get answers due to knowledge lag after the version update, and significantly improves the user experience.

[0053] In one feasible implementation, the prediction unit is used to: The modified function points are assigned to the corresponding function domains, and each function domain is queried to see if there is a corresponding intent cluster. If no corresponding intent cluster exists for any functional domain, multiple prediction question statements are generated based on the functional description of the corresponding changed functional point, and then processed by the intent cluster. Figure 1 After consistency refeedback verification, the corresponding predicted intent cluster is obtained, and the corresponding predicted knowledge entries and cluster fingerprints are generated based on the predicted intent cluster; If any functional domain has a corresponding intent cluster, then the intent cluster is incrementally updated based on the changed functional point; the incremental update includes adding new question variants and corresponding knowledge entries to the intent cluster, and updating the cluster fingerprint of the intent cluster.

[0054] For example, the generation of the predicted question statement can use a generative model, with the following prompt: "You are a user of an in-vehicle voice assistant. Regarding the following change: {Change of function description}, the original access method was {Original operation path}, now the access method is {New operation path}, the original function name was {Old Name}, and the current function name is {New Name}. Please provide N questions that car owners might ask the in-vehicle voice assistant, roughly divided into four parts: questions about function location, how to use this function, why there is a change, and questions using the old name. Each part accounts for approximately 25%, each line contains only one sentence, and technical jargon and formal language should be avoided." Here, N is determined based on the frequency with which the changed function item is encountered by the user, ranging from 40 to 50.

[0055] The generated prediction question needs to be processed. Figure 1 Consistency check: The predicted question is fed back into the same natural language understanding engine on the vehicle for recognition. Questions with the same recognition result as the original intent are retained, while those with different results are deleted. Deduplication check: Variants with a semantic similarity greater than or equal to 0.9 are deduplicated. Predicted questions that pass the verification constitute a predicted intent group.

[0056] Before the vehicle software version is released, a pre-update knowledge work order and test case set are automatically generated a predetermined number of days (e.g., 7 days). Knowledge engineers write or modify the corresponding knowledge items according to the predicted intent groups. After verification through regression testing, the knowledge update is packaged into a knowledge package version snapshot and synchronously distributed to the vehicle in the same gray-scale batch as the vehicle software version.

[0057] The above approach can accurately pinpoint the knowledge scope affected by the changes in functional points, ensuring a precise match between knowledge updates and functional changes. At the same time, by incrementally updating existing intent clusters rather than rebuilding them entirely, it effectively avoids duplication of work and waste of knowledge assets, and improves the efficiency and accuracy of knowledge pre-updates.

[0058] In one feasible implementation, the prediction module further includes: a prediction result verification unit; The prediction result verification unit is used to collect the vehicle-side performance feedback indicators for each prediction intent cluster based on the cluster fingerprint corresponding to each prediction intent cluster. If the semantic similarity between the actual problem statement of any predicted intent cluster and the predicted problem statement of the corresponding predicted intent cluster is greater than or equal to a preset similarity threshold, then the corresponding predicted intent cluster will be converted into an intent cluster. If the semantic similarity between the actual problem statement of any predicted intent cluster and the predicted problem statement of the corresponding predicted intent cluster is less than a preset similarity threshold, or if no actual sample appears for a consecutive preset number of rounds, the corresponding predicted intent cluster will be taken offline and archived.

[0059] For example, the preset similarity threshold is 60%, and the preset number of consecutive rounds is 2 rounds.

[0060] By using semantic similarity thresholds, only predicted intent clusters that have been verified through real interactions on the vehicle side can be officially recognized as such, thus avoiding the blind deployment of unverified knowledge. At the same time, predicted intent clusters with insufficient semantic similarity or no actual samples for a long time are promptly taken offline and archived, effectively preventing the redundant expansion of the knowledge base and ensuring its accuracy and maintainability.

[0061] To compare the effectiveness of the proposed solutions with existing technologies and intermediate solutions, three control groups were set up using the same vehicle platform, the same time window (March-May 2026), and the same functional domain (vehicle comfort and parking assistance). All three groups were deployed on real mass-produced vehicles and allocated the same vehicle-side resources (2MB storage, 5MB monthly bandwidth per vehicle, and 10ms interaction latency). The results of the three control experiments are shown in Table 1: Group A used the complete solution of the proposed solutions; Group B used the baseline solution, with full-text reporting, manual scheduling, and open-loop distribution; Group C used the intermediate solution, with dual-threshold traffic splitting, automatic weighted scheduling, ordinary gray-scale distribution, and manual rollback, but without cluster fingerprint attribution and predictive pre-update functions.

[0062] In Group B, under the same conditions, the daily reported traffic exceeded the monthly bandwidth limit and could not be launched, which was deemed unacceptable by the platform—this is the result of the control group, indicating that the conventional full-volume reporting scheme is not feasible under vehicle-side constraints. The launch cycle time (42 days) for Group B in the table was obtained from work orders completed by the same group of people using the same method. The calculation method for indicators was the same for all groups. Traceability refers to the proportion of the return sample clusters divided by the system that are consistent with the manual review annotations, which remained consistent across the three groups. "Misjudged release" refers to a situation where the data was considered qualified through statistical testing during the gray-scale period (p<0.05), but key indicators were found to be unqualified after manual review following the official release. Group C had two misjudgments, both of which can be confirmed by relevant work orders. The experimental results for the three groups are shown in Table 1.

[0063] Table 1 Results of the three control experiments

[0064] Table 1 shows that while Group C also has automated factors such as automatic scheduling and canary releases, the lack of a cluster fingerprint attribution mechanism led to two instances of misjudged volume increases after model version updates due to measurement object drift causing statistical test errors. Furthermore, the lack of a predictive pre-update mechanism resulted in a 14-day service gap after version launch. Compared to Group C, Group A showed improved attributability from 71% to 98.6%, and a reduced gap from 14 days to 0 days. The difference between Group A and Group B is that Group A only experienced 3% of the traffic instead of 100%, highlighting the importance of lightweight design under the four constraints on the vehicle side.

[0065] The intelligent cockpit dynamic knowledge service system based on OnStar provided in this application embodiment uses a vehicle-side interaction perception module to perform structured data embedding for question-and-answer interaction data and combines a confidence-based dual-threshold diversion strategy for reporting and filtering, effectively reducing the bandwidth consumption of the vehicle-cloud link. Simultaneously, the structured data embedding does not contain the original corpus, ensuring user privacy data compliance from the source. The cloud-based demand mining module uses semantic clustering to group massive, scattered question-and-answer events into intent clusters and generates cross-version fixed cluster fingerprints, providing stable semantic anchors for subsequent effect feedback. This solves the technical problem of unreliable attribution due to measurement object drift caused by model version iteration. The cluster fingerprints are lightweightly encoded according to vehicle-side resource constraints, enabling deployment in vehicle-side environments with limited storage capacity and computing power. This achieves local intent attribution judgment on the vehicle side, completing attribution without the original corpus leaving the vehicle, meeting regulatory requirements for in-vehicle data processing, and contributing to improved user interaction experience. The knowledge operation and verification module automatically generates knowledge update work orders based on priority, ensuring that high-frequency issues are resolved first. It automatically generates and verifies test case sets based on the question phrasing of each intent cluster, ensuring that test cases are semantically consistent with real-world scenarios. This provides verifiable quality control for knowledge revisions, significantly improving the success rate of deployment. The distribution module sends knowledge package version snapshots and cluster fingerprints to the vehicle, enabling knowledge updates to be deployed gradually in a controlled, gray-scale manner, reducing the risk of large-scale releases. The vehicle-side intelligent agent updates the vehicle's knowledge base based on the knowledge package version snapshot and determines the answer to the user's current question locally on the vehicle, ensuring that knowledge updates take effect immediately on the vehicle without waiting for cloud interaction, resulting in faster response times. Simultaneously, it completes intent attribution determination based on cluster fingerprints without uploading original corpora, changing reported data from raw questions to structured feedback indicators, significantly reducing the amount of reported data and substantially lowering operating costs. The effect feedback verification module uses cluster fingerprinting to statistically compare the effect feedback indicators of the corresponding intent cluster with historical benchmark values, forming a complete closed-loop link. This ensures that every knowledge update is automatically verified by real key performance indicators on the vehicle side, transforming knowledge services from passive manual maintenance to data-driven closed-loop autonomy. This enables the entire knowledge update process to be traceable, measurable, and controllable, improving the iteration efficiency and reliability of knowledge services.

[0066] Based on the OnStar-based intelligent cockpit dynamic knowledge service system described in the preceding embodiments, this application also provides an OnStar-based intelligent cockpit dynamic knowledge service method. Figure 2 This is a flowchart of the method. For example... Figure 2 As shown, the method is applied to the OnStar-based intelligent cockpit dynamic knowledge service system described in the previous embodiment, and the method includes the following steps: S201, the vehicle-side interaction perception module performs structured data collection based on the question-and-answer data during the interaction process, obtains structured data collection data, and filters the structured data collection data according to the confidence level dual threshold diversion strategy to obtain the events to be optimized and report them to the cloud demand mining module.

[0067] The structured data embedding does not include the original corpus.

[0068] In one feasible implementation, the structured embedded data is filtered according to a confidence-based dual-threshold triage strategy to obtain events to be optimized, including: If the confidence value of the answer is less than or equal to the first preset threshold, the corresponding question and answer data is reported as the event to be optimized; if the confidence value of the answer is greater than the first preset threshold and less than the second preset threshold, a clarifying follow-up question is initiated to the user. If the user's reply to the clarifying follow-up question indicates that the original question has been answered, it is not reported; otherwise, the corresponding question and answer data is reported as the event to be optimized; if the confidence value of the answer is greater than or equal to the second preset threshold, local anonymous counting is performed, and it is not reported.

[0069] Here, the first preset threshold is lower than the second preset threshold. The selection criteria are as follows: Samples from the past six months of in-vehicle voice recognition question-and-answer logs, manually labeled with "whether the answer was actually accepted by the user," are used to plot the precision-recall curve of the confidence score. The first preset threshold is the confidence value corresponding to a recall rate of 95%, and the second preset threshold is the confidence value corresponding to a precision rate of 98%. For example, the first preset threshold is set to 0.60, and the second preset threshold is set to 0.85. These two thresholds are not static; they are recalculated and calibrated according to the same standards each time the natural language processing model is updated.

[0070] Low-confidence events are directly reported to ensure timely problem detection; medium-confidence events are confirmed through clarification questions to ensure user satisfaction, avoiding false reporting and invalid data while also improving user experience; high-confidence events are counted locally without reporting, minimizing cloud communication overhead. This strategy ensures effective data collection while further optimizing bandwidth consumption and privacy protection, making the reported data more refined and efficient.

[0071] S202 The cloud-based demand mining module performs semantic clustering on the reported events to be optimized, obtains multiple intent clusters, generates corresponding cluster fingerprints for each intent cluster, and outputs the cluster fingerprints after lightweight encoding based on vehicle-side resource constraints.

[0072] The cluster fingerprint is used to uniquely identify the semantic identity of the corresponding intent cluster.

[0073] The vehicle-side resource constraints include storage capacity limitations and computation latency limitations.

[0074] The goal of the encoding process is to compress the cluster fingerprint to a size that can be stored on the vehicle side, while ensuring that intent attribution determination can be completed within 10ms latency on the vehicle side. Specific encoding methods include, but are not limited to, the following: quantization encoding: quantizing a high-dimensional floating-point vector into a low-bit integer representation, such as quantizing a 32-bit floating-point vector into an 8-bit integer; dimensionality reduction encoding: using dimensionality reduction methods such as principal component analysis or autoencoders to retain the main semantic information while reducing dimensionality; hash encoding: mapping the semantic encoding to a fixed-length hash value.

[0075] The size of the encoded cluster fingerprint must meet the vehicle-side storage constraints. For example, in a fleet of 5 million vehicles, assuming that each vehicle needs to store thousands of cluster fingerprints on average, each cluster fingerprint occupies about several hundred bytes after encoding, and the total storage space is within 2MB.

[0076] By using lightweight encoding of cluster fingerprints, high-dimensional semantic vectors are compressed into low-dimensional compact encodings, enabling cluster fingerprints to be stored locally on the vehicle within a few MB of storage space and to complete intent attribution determination within millisecond latency. At the same time, it achieves the data compliance requirement that the original corpus does not leave the vehicle—the vehicle only reports structured backflow data such as cluster identifiers, knowledge package version numbers, and interaction counts. This fundamentally solves the quadruple constraints of vehicle storage, computing power, bandwidth, and data compliance, making local intent attribution determination and effect attribution verification based on cluster fingerprints possible on mass-produced vehicles.

[0077] S203. The knowledge operation and verification module generates corresponding knowledge update work orders based on each intent cluster, revises the knowledge items corresponding to each knowledge update work order according to the priority of the knowledge update work orders, generates corresponding test case sets based on the question of each intent cluster, uses the test case sets to verify the corresponding revised knowledge items, and packages the verified revised knowledge items into the corresponding knowledge package version snapshot.

[0078] The knowledge update work order includes the question type, sample size, and knowledge item location information for the corresponding intent cluster.

[0079] The knowledge package version snapshot here includes the content of the knowledge entries, version number, applicable vehicle models and the correspondence between the software version baseline, digital signature and rollback pointer information.

[0080] The priority of knowledge update work orders is determined by a weighted calculation based on three dimensions: the popularity of the intent cluster, the scope of impact, and the severity. Issues with high popularity, wide impact, and high severity are automatically assigned as expedited work orders, reducing the time from discovery to deployment of common issues from weeks to days. Popularity represents the number of samples per unit time and the month-on-month change trend; the scope of impact represents the number of affected vehicles on the network and the number of different brands and models affected; severity represents whether it involves safety-related areas and whether there is negative public opinion.

[0081] Test case sets can be generated through the following paths: Path 1: Prepare a thesaurus and a pool of colloquial expression templates for high-frequency functional domains in the vehicle beforehand. Break down the questions in each intent cluster into intent slots, function slots, and condition slots. Generate different question variations in batches using methods such as slot replacement, word order adjustment, and component deletion. For multi-turn dialogues, generate follow-up questions, counter-questions, and conditional supplementation based on the designed dialogue state machine template. Path 2: Use the company's proprietary large language model, using representative questioning methods from each intent cluster as seeds, to generate different questioning methods and corresponding multi-turn dialogue paths based on given prompts.

[0082] Test cases generated using any of the above methods must be reviewed. Figure 1 Consistency re-feedback verification and deduplication coverage verification. Figure 1 Consistency refeeding verification involves refeeding candidate variants into the same natural language understanding engine on the vehicle side for recognition. Variants with the same result as the original intent are retained, while those with different results are deleted. Deduplication coverage verification ensures that variants with semantic similarity greater than or equal to 0.9 are deduplicated, and that the variant set covers 100% of the original intent cluster. Optimized knowledge entries must pass the overall test on the regression test set (e.g., accuracy exceeding 95%) before entering the release candidate pool.

[0083] S204. The distribution module sends the version snapshots of each knowledge package and the cluster fingerprints corresponding to each intent cluster after lightweight encoding to the vehicle-side intelligent agent.

[0084] The encoded cluster fingerprint is sent to the vehicle via the SOTA channel along with the knowledge package version snapshot and stored locally on the vehicle. When the knowledge package version is updated, only the changed cluster fingerprint is synchronized to the vehicle to reduce the amount of data transmitted.

[0085] S205. The vehicle-side intelligent agent updates the vehicle-side knowledge base based on the snapshot of each knowledge package version. Based on the updated vehicle-side knowledge base, it determines the answer to the user's current question locally on the vehicle. Without uploading the original corpus, it performs intent attribution judgment on the current question and answer content based on the cluster fingerprint corresponding to each intent cluster after lightweight encoding. Based on the judgment result, it generates the corresponding effect feedback indicators and reports them to the effect feedback verification module.

[0086] The performance feedback metrics are structured feedback data, which include at least cluster identifiers, knowledge package version numbers, and number of interactions.

[0087] During each user question-and-answer interaction, the vehicle-side intelligent agent determines intent attribution based on locally cached cluster fingerprints. Specifically, the user's question is converted into a vector representation using a natural language understanding model deployed on the vehicle, and then matched against the locally stored cluster fingerprints to determine the intent cluster to which the question belongs. This process is completed locally on the vehicle, without uploading the original corpus to the cloud, with a latency of no more than 10ms. When the vehicle is in a network-free environment (such as underground parking garages or tunnels with weak networks), the vehicle still uses the locally cached cluster fingerprints for attribution determination and data logging. Incremental structured data is added after network recovery to ensure uninterrupted attribution. Because the structured data does not contain the original interaction corpus, the original voice or text information is not transmitted from the vehicle, thus meeting data compliance requirements.

[0088] S206. The effect feedback verification module statistically compares the effect feedback indicators of the corresponding intent cluster with the historical benchmark value based on the fingerprint of each cluster, and determines the actual effect of the corresponding knowledge package version snapshot based on the comparison results.

[0089] In one feasible implementation, the effect feedback index of the corresponding intent cluster is statistically compared with the historical benchmark value based on the fingerprint of each cluster, and the actual effect of the corresponding knowledge package version snapshot is determined according to the comparison result, including: The effect feedback index of each intent cluster was statistically significant compared with the historical baseline, and the corresponding test results were obtained. If the corresponding performance feedback metric is greater than the historical baseline value, and the corresponding test result indicates that the difference between the corresponding performance feedback metric and the historical baseline value is statistically significant, then the corresponding knowledge package version snapshot is valid; the knowledge package version snapshot includes the content of the knowledge item, version number, correspondence between the applicable vehicle model and the software version baseline, digital signature, and rollback pointer information; If the test result indicates that the corresponding effect feedback index is less than or equal to the historical benchmark value, or if the difference between the corresponding effect feedback index and the historical benchmark value is not statistically significant, then the system will automatically roll back to the previous knowledge package version.

[0090] Using real key performance indicators from the vehicle side as the basis for release, combined with statistical significance testing, ensures the reliability of the judgment conclusions and effectively avoids misjudgment and over-release due to sampling errors. At the same time, by using information such as version number, applicable vehicle model and software version baseline correspondence, digital signature and rollback pointer contained in the knowledge package version snapshot, the accuracy and traceability of rollback operations are ensured. Every knowledge update is tested and verified, and unqualified updates are automatically rolled back, which greatly improves the reliability and controllability of knowledge service iteration.

[0091] Since the above content is a passive response update based on real interactive feedback, a knowledge pre-update is performed before the vehicle software version is released, as follows: In one feasible implementation, the method further includes: Obtain the vehicle software version release plan, and extract the change function points based on the baseline change list of the vehicle software version release plan; Based on the changed function points, determine the corresponding prediction intent clusters, as well as the prediction knowledge entries and cluster fingerprints corresponding to each prediction intent cluster; The prediction knowledge entries and cluster fingerprints corresponding to each prediction intent cluster are distributed to the vehicle along with the vehicle software version in the vehicle software version release plan.

[0092] In one feasible implementation, based on the changed function point, the corresponding prediction intent cluster, and the prediction knowledge entry and cluster fingerprint corresponding to each prediction intent cluster are determined, including: The modified function points are assigned to the corresponding function domains, and each function domain is queried to see if there is a corresponding intent cluster. If no corresponding intent cluster exists for any functional domain, multiple prediction question statements are generated based on the functional description of the corresponding changed functional point, and then processed by the intent cluster. Figure 1 After consistency refeedback verification, the corresponding predicted intent cluster is obtained, and the corresponding predicted knowledge entries and cluster fingerprints are generated based on the predicted intent cluster; If any functional domain has a corresponding intent cluster, then the intent cluster is incrementally updated based on the changed functional point; the incremental update includes adding new question variations and corresponding knowledge entries to the intent cluster, and updating the cluster fingerprint of the intent cluster.

[0093] In one feasible implementation, after distributing the prediction knowledge entries and cluster fingerprints corresponding to each prediction intent cluster to the vehicle along with the vehicle software version in the vehicle software version release plan, the method further includes: Based on the cluster fingerprint corresponding to each predicted intent cluster, collect the vehicle-side performance feedback indicators for each predicted intent cluster. If the semantic similarity between the actual problem statement of any predicted intent cluster and the predicted problem statement of the corresponding predicted intent cluster is greater than or equal to a preset similarity threshold, then the corresponding predicted intent cluster will be converted into an intent cluster. If the semantic similarity between the actual problem statement of any predicted intent cluster and the predicted problem statement of the corresponding predicted intent cluster is less than a preset similarity threshold, or if no actual sample appears for a consecutive preset number of rounds, the corresponding predicted intent cluster will be taken offline and archived.

[0094] The intelligent cockpit dynamic knowledge service method based on OnStar provided in this application embodiment uses a vehicle-side interaction perception module to perform structured data embedding for question-and-answer interaction data and combines a confidence-based dual-threshold diversion strategy for reporting and filtering, effectively reducing the bandwidth consumption of the vehicle-cloud link. Simultaneously, the structured data embedding does not contain the original corpus, ensuring user privacy data compliance from the source. The cloud-based demand mining module uses semantic clustering to group massive, scattered question-and-answer events into intent clusters and generates cross-version fixed cluster fingerprints, providing stable semantic anchors for subsequent effect feedback. This solves the technical problem of unreliable attribution due to measurement object drift caused by model version iteration. Lightweight encoding of cluster fingerprints based on vehicle-side resource constraints enables deployment in vehicle-side environments with limited storage capacity and computing power, achieving local intent attribution judgment on the vehicle side. Attribution is completed without the original corpus leaving the vehicle, meeting regulatory requirements for in-vehicle data processing and contributing to improved user interaction experience. The knowledge operation and verification module automatically generates knowledge update work orders based on priority, ensuring that high-frequency issues are resolved first. It automatically generates and verifies test case sets based on the question phrasing of each intent cluster, ensuring that test cases are semantically consistent with real-world scenarios. This provides verifiable quality control for knowledge revisions, significantly improving the success rate of deployment. The distribution module sends knowledge package version snapshots and cluster fingerprints to the vehicle, enabling knowledge updates to be deployed gradually in a controlled, gray-scale manner, reducing the risk of large-scale releases. The vehicle-side intelligent agent updates the vehicle's knowledge base based on the knowledge package version snapshot and determines the answer to the user's current question locally on the vehicle, ensuring that knowledge updates take effect immediately on the vehicle without waiting for cloud interaction, resulting in faster response times. Simultaneously, it completes intent attribution determination based on cluster fingerprints without uploading original corpora, changing reported data from raw questions to structured feedback indicators, significantly reducing the amount of reported data and substantially lowering operating costs. The effect feedback verification module uses cluster fingerprinting to statistically compare the effect feedback indicators of the corresponding intent cluster with historical benchmark values, forming a complete closed-loop link. This ensures that every knowledge update is automatically verified by real key performance indicators on the vehicle side, transforming knowledge services from passive manual maintenance to data-driven closed-loop autonomy. This enables the entire knowledge update process to be traceable, measurable, and controllable, improving the iteration efficiency and reliability of knowledge services.

[0095] It should be noted that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for the device embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiments. The device embodiments described above are merely illustrative, and the units described as separate components may or may not be physically separate. The components indicated as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment solution according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0096] The above description is merely one specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A dynamic knowledge service system for intelligent cockpits based on OnStar, characterized in that, The system includes: a vehicle-side interaction perception module, a cloud-based demand mining module, a knowledge operation and verification module, a distribution module, a vehicle-side intelligent agent, and an effect feedback verification module; The vehicle-side interaction perception module is used to perform structured data collection based on the question-and-answer data during the interaction process, obtain structured data collection data, and filter the structured data collection data according to the confidence level dual threshold diversion strategy to obtain the events to be optimized and report them to the cloud-based demand mining module; the structured data collection data does not contain the original corpus. The cloud-based demand mining module is used to perform semantic clustering on the reported events to be optimized, obtain multiple intent clusters, generate a corresponding cluster fingerprint for each intent cluster, and output the cluster fingerprint after lightweight encoding according to the vehicle-side resource constraints; the cluster fingerprint is used to uniquely identify the semantic identity of the corresponding intent cluster. The knowledge operation and verification module is used to generate corresponding knowledge update work orders based on each intent cluster, revise the knowledge items corresponding to each knowledge update work order according to the priority of the knowledge update work orders, generate corresponding test case sets based on the question of each intent cluster, verify the corresponding revised knowledge items using the test case sets, and package the verified revised knowledge items into the corresponding knowledge package version snapshot; the knowledge update work order includes the question of the corresponding intent cluster, sample size, and knowledge item location information. The distribution module is used to send the snapshot of each knowledge package version and the cluster fingerprint corresponding to each intent cluster after lightweight encoding to the vehicle-side intelligent agent; The vehicle-side intelligent agent is used to update the vehicle-side knowledge base based on the snapshots of each knowledge package version, determine the answer to the user's current question locally on the vehicle-side based on the updated vehicle-side knowledge base, and, without uploading the original corpus, determine the intent attribution of the current question and answer content based on the cluster fingerprints corresponding to each intent cluster after lightweight encoding, and generate corresponding effect feedback indicators based on the judgment results and report them to the effect feedback verification module. The effect feedback verification module is used to statistically compare the effect feedback indicators of the corresponding intent cluster with the historical benchmark value based on the fingerprint of each cluster, and determine the actual effect of the corresponding knowledge package version snapshot based on the comparison results.

2. The system according to claim 1, characterized in that, The system further includes: a prediction module; the prediction module includes: The change function point extraction unit is used to obtain the vehicle software version release plan and extract the change function points based on the baseline change list of the vehicle software version release plan. The prediction unit is used to determine the corresponding prediction intent cluster, as well as the prediction knowledge entry and cluster fingerprint corresponding to each prediction intent cluster, based on the changed function point. The prediction result distribution unit is used to distribute the prediction knowledge entries and cluster fingerprints corresponding to each prediction intention cluster to the vehicle along with the vehicle software version in the vehicle software version release plan.

3. The system according to claim 2, characterized in that, The prediction unit is used for: The modified function points are assigned to the corresponding function domains, and each function domain is queried to see if there is a corresponding intent cluster. If no corresponding intent cluster exists in any functional domain, multiple prediction question statements are generated based on the functional description of the corresponding changed functional point. After the intent consistency is verified, the corresponding prediction intent cluster is obtained, and the corresponding prediction knowledge item and cluster fingerprint are generated based on the prediction intent cluster. If any functional domain has a corresponding intent cluster, then the intent cluster is incrementally updated based on the changed functional point; the incremental update includes adding new question variants and corresponding knowledge entries to the intent cluster, and updating the cluster fingerprint of the intent cluster.

4. The system according to claim 2, characterized in that, The prediction module further includes: a prediction result verification unit; The prediction result verification unit is used to collect the vehicle-side performance feedback indicators for each prediction intent cluster based on the cluster fingerprint corresponding to each prediction intent cluster. If the semantic similarity between the actual problem statement of any predicted intent cluster and the predicted problem statement of the corresponding predicted intent cluster is greater than or equal to a preset similarity threshold, then the corresponding predicted intent cluster will be converted into an intent cluster. If the semantic similarity between the actual problem statement of any predicted intent cluster and the predicted problem statement of the corresponding predicted intent cluster is less than a preset similarity threshold, or if no actual sample appears for a consecutive preset number of rounds, the corresponding predicted intent cluster will be taken offline and archived.

5. The system according to claim 1, characterized in that, The effect feedback verification module is used for: The effect feedback index of each intent cluster was statistically significant compared with the historical baseline, and the corresponding test results were obtained. If the corresponding performance feedback indicator is greater than the historical benchmark value, and the corresponding test result indicates that the difference between the corresponding performance feedback indicator and the historical benchmark value is statistically significant, then the corresponding knowledge package version snapshot is valid. The knowledge package version snapshot includes the content of the knowledge entries, version number, correspondence between applicable vehicle models and software version baselines, digital signature, and rollback pointer information; If the test result indicates that the corresponding effect feedback index is less than or equal to the historical benchmark value, or if the difference between the corresponding effect feedback index and the historical benchmark value is not statistically significant, then the system will automatically roll back to the previous knowledge package version.

6. The system according to claim 1, characterized in that, The vehicle-side interactive perception module is used for: If the confidence value of the answer is less than or equal to the first preset threshold, the corresponding question and answer data is reported as the event to be optimized; if the confidence value of the answer is greater than the first preset threshold and less than the second preset threshold, a clarification follow-up question is initiated to the user. If the user's reply to the clarification follow-up question indicates that the original question has been answered, it is not reported; otherwise, the corresponding question and answer data is reported as the event to be optimized. If the confidence value of the answer is greater than or equal to the second preset threshold, then a local anonymous count is performed and no report is submitted.

7. A method for providing dynamic knowledge services for intelligent cockpits based on OnStar, characterized in that, The method is applied to the OnStar-based intelligent cockpit dynamic knowledge service system as described in claim 1, and the method includes: The vehicle-side interaction perception module performs structured data collection based on the question-and-answer data during the interaction process, obtains structured data collection data, and filters the structured data collection data according to the confidence level dual threshold diversion strategy to obtain the events to be optimized and report them to the cloud-based demand mining module; the structured data collection data does not contain the original corpus. The cloud-based demand mining module performs semantic clustering on the reported events to be optimized, obtains multiple intent clusters, and generates a corresponding cluster fingerprint for each intent cluster. The cluster fingerprint is then output after lightweight encoding based on vehicle-side resource constraints. The cluster fingerprint is used to uniquely identify the semantic identity of the corresponding intent cluster. The knowledge operation and verification module generates corresponding knowledge update work orders based on each intent cluster, revises the knowledge items corresponding to each knowledge update work order according to the priority of the knowledge update work orders, generates corresponding test case sets based on the question of each intent cluster, uses the test case sets to verify the corresponding revised knowledge items, and packages the verified revised knowledge items into the corresponding knowledge package version snapshot; the knowledge update work order includes the question of the corresponding intent cluster, sample size, and knowledge item location information. The distribution module sends the version snapshots of each knowledge package and the cluster fingerprints corresponding to each intent cluster after lightweight encoding to the vehicle-side intelligent agent. The vehicle-side intelligent agent updates the vehicle-side knowledge base based on the snapshots of each knowledge package version, determines the answer to the user's current question locally on the vehicle-side based on the updated vehicle-side knowledge base, and performs intent attribution judgment on the current question and answer content based on the cluster fingerprints corresponding to each intent cluster after lightweight encoding without uploading the original corpus. Based on the judgment result, it generates corresponding effect feedback indicators and reports them to the effect feedback verification module. The effect feedback verification module statistically compares the effect feedback indicators of the corresponding intent cluster with the historical benchmark value based on the fingerprint of each cluster, and determines the actual effect of the corresponding knowledge package version snapshot based on the comparison results.

8. The method according to claim 7, characterized in that, The method further includes: Obtain the vehicle software version release plan, and extract the change function points based on the baseline change list of the vehicle software version release plan; Based on the changed function points, determine the corresponding prediction intent clusters, as well as the prediction knowledge entries and cluster fingerprints corresponding to each prediction intent cluster; The prediction knowledge entries and cluster fingerprints corresponding to each prediction intent cluster are distributed to the vehicle along with the vehicle software version in the vehicle software version release plan.

9. The method according to claim 7, characterized in that, The step of determining the corresponding prediction intent cluster, and the prediction knowledge entry and cluster fingerprint corresponding to each prediction intent cluster, based on the changed function point, includes: The modified function points are assigned to the corresponding function domains, and each function domain is queried to see if there is a corresponding intent cluster. If no corresponding intent cluster exists in any functional domain, multiple prediction question statements are generated based on the functional description of the corresponding changed functional point. After the intent consistency is verified, the corresponding prediction intent cluster is obtained, and the corresponding prediction knowledge item and cluster fingerprint are generated based on the prediction intent cluster. If any functional domain has a corresponding intent cluster, then the intent cluster is incrementally updated based on the changed functional point; the incremental update includes adding new question variants and corresponding knowledge entries to the intent cluster, and updating the cluster fingerprint of the intent cluster.

10. The method according to claim 7, characterized in that, After distributing the prediction knowledge entries and cluster fingerprints corresponding to each prediction intent cluster to the vehicle along with the vehicle software version in the vehicle software version release plan, the method further includes: Based on the cluster fingerprint corresponding to each predicted intent cluster, collect the vehicle-side performance feedback indicators for each predicted intent cluster. If the semantic similarity between the actual problem statement of any predicted intent cluster and the predicted problem statement of the corresponding predicted intent cluster is greater than or equal to a preset similarity threshold, then the corresponding predicted intent cluster will be converted into an intent cluster. If the semantic similarity between the actual problem statement of any predicted intent cluster and the predicted problem statement of the corresponding predicted intent cluster is less than a preset similarity threshold, or if no actual sample appears for a consecutive preset number of rounds, the corresponding predicted intent cluster will be taken offline and archived.

Images