Intelligent quality inspection method and device based on RAG and dual-enhancement mechanism

Abstract

The invention provides an intelligent quality inspection method and device based on RAG and a dual-enhancement mechanism, and the method comprises the steps: determining a retrieval text based on a transcriptional text of voice data to be subjected to quality inspection and an initial quality inspection result; based on the semantics of the retrieval text and the scene label of the voice data, retrieving from a preset knowledge base to obtain matched entries and the relevancy of the matched entries; whether an IDK enhancement mechanism is triggered or not is judged based on the matching items, if not, a direct preference optimization DPO model is called to conduct preference matching degree scoring on the initial quality inspection result, and the DPO matching degree of the initial quality inspection result is obtained; and determining the accuracy and confidence of the initial quality inspection result based on the correlation degree of the matched items, the confidence of the large model, the DPO matching degree and the accuracy of the historical similar cases, and determining a final quality inspection result of the voice data based on the accuracy and confidence. According to the method and the device provided by the invention, the result accuracy, the illusion resistance and the service adaptation efficiency of intelligent quality inspection are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an intelligent quality inspection method and apparatus based on RAG and a dual reinforcement mechanism. Background Technology

[0002] As regulatory requirements for the consumer finance industry become increasingly stringent, intelligent quality inspection has gradually become a core means of post-loan quality control. Its typical technical approach is an automated judgment scheme combining "large-scale models + automatic voice recognition." This scheme aims to replace traditional manual sampling inspections, improving inspection efficiency and coverage.

[0003] Currently, the mainstream technologies in the industry to ensure the accuracy of intelligent quality inspection results mainly fall into three categories: First, there is the single-rule verification technology based on manually preset rules. Although this method can identify some explicit violations, it is difficult to cope with complex semantic biases, and the actual accuracy guarantee rate is generally less than 70%. Second, there is the pure large model iterative optimization technology, which relies on manually labeled error samples to fine-tune the model. However, this method still suffers from the problem of model "illusion"—when the model lacks knowledge of specific quality inspection items, the proportion of generating fictitious judgment basis exceeds 15%. At the same time, the fine-tuning process heavily relies on a large amount of labeled data, resulting in slow business response. After policy iteration, it often takes more than 10 days to complete the model update, which cannot adapt to the rapidly changing compliance requirements of the preservation business. Third, there is the simplified retrieval augmented generation (RAG) technology, which introduces external knowledge bases to assist in generation, but its effect on improving the accuracy of quality inspection is still limited.

[0004] In summary, existing technologies have not systematically addressed the core requirements of intelligent quality inspection in terms of "high accuracy, low illusion, and fast adaptation," and there is an urgent need to provide an intelligent quality inspection method with higher accuracy. Summary of the Invention

[0005] This invention provides an intelligent quality inspection method and device based on RAG and dual reinforcement mechanism to solve the shortcomings of the existing intelligent quality inspection technology in terms of both accuracy and adaptability.

[0006] This invention provides an intelligent quality inspection method based on RAG and a dual reinforcement mechanism, comprising: Based on the transcribed text of the speech data to be inspected and the initial quality inspection results, the retrieval text is determined. Based on the semantics of the retrieved text and the scene tags of the voice data, matching entries and the relevance of the matching entries are retrieved from a preset knowledge base; Based on the matching entries, it is determined whether the IDK enhancement mechanism is triggered. If it is not triggered, the trained Direct Preference Optimization (DPO) model is called to score the preference matching degree of the initial quality inspection result, and the DPO matching degree of the initial quality inspection result is obtained. The DPO model is trained based on the sample preference text pairs in historical quality inspection cases. Based on the relevance of the matching entries, the confidence of the large model, the matching degree of the DPO, and the accuracy of historical similar cases, the accuracy confidence of the initial quality inspection result is determined, and the final quality inspection result of the speech data is determined based on the accuracy confidence.

[0007] According to the intelligent quality inspection method based on RAG and dual reinforcement mechanism provided by the present invention, the step of retrieving matching entries from a preset knowledge base based on the semantics of the retrieved text and the scene tags of the voice data, and the relevance of the matching entries, includes: Candidate entries are selected from the entries based on the semantic similarity between the searched text and each entry in the knowledge base; Based on the scene tags of the voice data, target entries are obtained by filtering from the candidate entries; The target entries are reordered, and the matching entries and their relevance are determined based on the reordering results.

[0008] According to the intelligent quality inspection method based on RAG and dual reinforcement mechanism provided by the present invention, the step of reordering the target items, determining the matching items based on the reordering results, and the relevance of the matching items includes: Based on the reordering model, the relevance, timeliness, and scenario matching degree of the target item are obtained; The relevance, timeliness, and scenario matching of the target item are weighted and summed to obtain the comprehensive score of the target item; Based on the comprehensive score, the target items are reordered, and a preset number of target items at the top of the ranking and their relevance are used as the matching items, along with the relevance of the matching items.

[0009] According to the intelligent quality inspection method based on RAG and dual reinforcement mechanism provided by the present invention, the step of determining whether the IDK reinforcement mechanism is triggered based on the matching entry includes: If the overall score of the matched item is less than the overall score threshold, or the confidence level of the large model is less than the confidence threshold, or the expiration time of the matched item is less than the current time, then the IDK reinforcement mechanism is determined to be triggered; otherwise, it is determined not to be triggered.

[0010] According to the intelligent quality inspection method based on RAG and dual reinforcement mechanism provided by the present invention, after determining that the IDK reinforcement mechanism is triggered, the method further includes: The final quality inspection result of the voice data is determined to be pending review, and the pending review result is prohibited from being sent back to the business system; Add knowledge gap information to the database and push corresponding reminder messages to the knowledge base maintenance personnel's terminals.

[0011] According to the intelligent quality inspection method based on RAG and dual reinforcement mechanism provided by the present invention, the step of calling the trained Direct Preference Optimization (DPO) model to score the preference matching degree of the initial quality inspection result, and obtaining the DPO matching degree of the initial quality inspection result, includes: Based on the DPO model, calculate the initial matching score of the initial quality inspection results; If the initial matching score is less than the matching score threshold, the initial quality inspection result is adjusted and the adjusted matching score is calculated until the adjusted matching score is greater than or equal to the matching score threshold or the number of adjustments meets the preset number. The final adjusted matching score is used as the DPO matching score of the initial quality inspection result.

[0012] According to the intelligent quality inspection method based on RAG and dual enhancement mechanism provided by the present invention, determining the final quality inspection result of the speech data based on the accuracy confidence level includes: If the accuracy confidence level falls within the first confidence level range, a final quality inspection result consistent with the initial quality inspection result is generated, the matching entry is used as the reference, and the result is sent back to the business system. If the accuracy confidence level falls within the second confidence level range, then the final quality inspection result and review recommendations are generated and sent back to the business system; If the accuracy confidence level falls within the third confidence level range, then the final quality inspection result is determined to be pending review.

[0013] This invention also provides an intelligent quality inspection device based on RAG and a dual-enhancement mechanism, comprising: The text determination unit is used to determine the retrieval text based on the transcribed text of the speech data to be inspected and the initial quality inspection results; The knowledge retrieval unit is used to retrieve matching entries and the relevance of the matching entries from a preset knowledge base based on the semantics of the retrieved text and the scene tags of the voice data. The preference scoring unit is used to determine whether the IDK reinforcement mechanism is triggered based on the matching item. If it is not triggered, the trained Direct Preference Optimization (DPO) model is called to score the preference matching degree of the initial quality inspection result and obtain the DPO matching degree of the initial quality inspection result. The DPO model is trained based on the sample preference text pairs in historical quality inspection cases. The result determination unit is used to determine the accuracy confidence of the initial quality inspection result based on the relevance of the matching entries, the large model confidence, the DPO matching degree, and the accuracy of historical similar cases, and to determine the final quality inspection result of the speech data based on the accuracy confidence.

[0014] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the intelligent quality inspection method based on RAG and dual reinforcement mechanism as described above.

[0015] The present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the intelligent quality inspection method based on RAG and dual reinforcement mechanism as described above.

[0016] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the intelligent quality inspection method based on RAG and dual reinforcement mechanism as described above.

[0017] The intelligent quality inspection method and apparatus based on RAG and dual reinforcement mechanisms provided by this invention significantly improve the accuracy, anti-illusion capability, and business adaptation efficiency of intelligent quality inspection results by integrating semantic retrieval, dual reinforcement mechanisms, and multi-dimensional confidence assessment. This solution achieves rapid response to changes in business rules without frequent model fine-tuning, significantly shortening the adaptation cycle after policy iteration, while ensuring high accuracy and low false positive rate. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention 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 some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is one of the flowcharts of the intelligent quality inspection method based on RAG and dual reinforcement mechanism provided by the present invention.

[0020] Figure 2 This is the second flowchart of the intelligent quality inspection method based on RAG and dual reinforcement mechanism provided by the present invention.

[0021] Figure 3 This is a schematic diagram of the intelligent quality inspection device based on RAG and dual reinforcement mechanism provided by the present invention.

[0022] Figure 4This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0024] Current technologies have not systematically addressed the core requirements of intelligent quality inspection in terms of "high accuracy, low illusion, and rapid adaptation." Specific problems include: a lack of control over the unknown aspects of the large model generation process, leading to susceptibility to misjudgments in knowledge blind spots; insufficient retrieval accuracy of traditional RAG technology, making it difficult to accurately match the semantic relationships between policy clauses and specific quality inspection items; high manual maintenance costs and significant delays during business rule iterations, resulting in obvious compliance risks; and the lack of quantifiable evaluation of the reliability of output results, making it impossible to effectively distinguish between high-reliability and low-reliability conclusions, leading to wasted resources for manual review and limited overall quality inspection efficiency.

[0025] To address the aforementioned issues, this invention proposes an intelligent quality inspection method based on RAG and a dual-enhancement mechanism. In this method, firstly, based on the transcribed text of the speech data to be inspected and the initial quality inspection results, a retrieval text is determined. Then, based on the semantics of the retrieval text and the scene tags of the speech data, matching entries and their relevance are retrieved from a pre-defined knowledge base. Next, based on the matching entries, it is determined whether the IDK enhancement mechanism is triggered. If not, a trained Direct Preference Optimization (DPO) model is invoked to score the preference matching degree of the initial quality inspection results, obtaining the DPO matching degree of the initial quality inspection results. The DPO model is trained based on sample preference text pairs from historical quality inspection cases. Finally, based on the relevance of the matching entries, the large model confidence, and the DPO matching degree, the accuracy confidence of the initial quality inspection results is determined, and the final quality inspection result of the speech data is determined based on the accuracy confidence.

[0026] The method provided in this invention, firstly, utilizes a retrieval mechanism based on semantics and scene tags to accurately match relevant policy and rule entries from the knowledge base, effectively overcoming the low relevance problem of traditional keyword matching methods, thus providing a reliable basis for quality inspection judgment. Secondly, by using an IDK reinforcement mechanism to suppress the generation of "illusions" in knowledge blind spots, combined with a preference matching score based on DPO training, it can better align with the decision preferences of human experts in historical cases, reducing misjudgments caused by subjective bias. Finally, by comprehensively considering retrieval relevance, large model confidence, and DPO matching degree to generate an accuracy confidence score, it not only achieves the quantification of the credibility of quality inspection results but also automatically distinguishes between high-confidence and low-confidence results, thereby reducing the pressure of manual review and improving quality inspection efficiency.

[0027] In summary, the method provided in this invention significantly improves the accuracy, anti-illusion capability, and business adaptation efficiency of intelligent quality inspection results by integrating semantic retrieval, reinforcement learning preference alignment, and multi-dimensional confidence assessment. This solution achieves rapid response to changes in business rules without requiring frequent model fine-tuning, significantly shortening the adaptation cycle after policy iterations, while ensuring high accuracy and low false positive rate.

[0028] The embodiments of this invention can be applied to scenarios requiring intelligent quality inspection, such as quality control scenarios for post-loan preservation. The executing entity of this method can be an electronic device such as a terminal device, computer, server, server cluster, or specially designed intelligent quality inspection equipment, or it can be an intelligent quality inspection device installed in the electronic device. The intelligent quality inspection device can be implemented through software, hardware, or a combination of both.

[0029] In the description of the embodiments of the present invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of the embodiments of the present invention, "multiple" means two or more, unless otherwise explicitly specified.

[0030] Figure 1 This is one of the flowcharts of the intelligent quality inspection method based on RAG and dual reinforcement mechanism provided by the present invention, such as... Figure 1 As shown, the method includes the following steps 110-140.

[0031] Step 110: Determine the search text based on the transcribed text of the speech data to be inspected and the initial quality inspection results.

[0032] Specifically, the voice data to be inspected refers to the audio recordings that need to be checked for compliance, such as voice recordings of customer service conversations. The transcription of the voice data into text can be achieved using automatic speech recognition technology, converting the audio data into written content. This transcription can be accomplished by accessing the data retrieval interface of the business system.

[0033] The initial quality inspection results can be the quality inspection conclusions and judgment criteria output by the large model after preliminary analysis of the transcribed text. For example, the initial quality inspection result may be "Violation: suspected false promises, and text content involving violations in the transcribed text".

[0034] The search text is a collection of key information fragments extracted from the initial quality control results and transcribed text for subsequent knowledge base retrieval. Its purpose is to provide precise query content for the next step of knowledge base retrieval, avoiding the direct use of large amounts of redundant and ineffective dialogue text, thereby significantly improving retrieval efficiency and accuracy. Natural language processing techniques are used to extract key entities and phrases from the initial quality control results and transcribed text. The extracted keywords are then combined with specific content from the dialogue to form the search text.

[0035] Step 120: Based on the scene tags of the semantic and speech data of the retrieved text, retrieve matching entries and the relevance of the matching entries from the preset knowledge base.

[0036] Specifically, this step utilizes text retrieval to quickly locate and recall the most relevant specific rule clauses or historical precedents from a massive and complex knowledge base, and assigns relevance scores to these entries. The aim is to provide objective and quantitative policy basis for quality inspection judgments, rather than solely relying on the implicit knowledge of large models.

[0037] Prior to this, a knowledge base can be built in advance. The knowledge base construction process is as follows: (1) Document parsing: call the ApacheTika tool to parse policy PDFs and rule Excels, extract structured text, and remove redundant content; (2) Vector conversion: use the bge-large-en-v1.5 model to convert structured text into 768-dimensional vectors, and set vector generation parameters; (3) Vector import: import “text content-vector-metadata” into the Milvus vector database, establish a “scenario tag” index, and support filtering and retrieval by tag. Understandably, the knowledge base stores all compliance policies, rules and regulations, verbal norms, historical precedents, and other knowledge that match the post-loan preservation business.

[0038] Scene tags for voice data are labels used to categorize the metadata of the voice calls to be examined, such as "credit card M2 stage communication" or "personal loan first-time phone communication." These tags provide business context to narrow down the search scope.

[0039] Matching entries are rules or cases retrieved from the knowledge base that are semantically and business-specifically relevant to the currently searched text. Matching entries can be retrieved using vectorized retrieval. Relevance can be a numerical score representing the degree of relevance between the retrieved matching entries and the searched text; a higher score indicates stronger relevance.

[0040] In some possible implementations, step 120 specifically includes: Step 121: Based on the semantic similarity between the retrieved text and each entry in the knowledge base, candidate entries are selected from each entry; Step 122: Based on the scene labels of the voice data, select the target item from the candidate items; Step 123: Reorder the target entries, determine the matching entries and their relevance based on the sorting results.

[0041] Specifically, semantic similarity refers to the degree of similarity in meaning between two texts, rather than simple keyword matching. For example, "Don't expose my address book" and "Don't leak my contact information" have high semantic similarity.

[0042] The bge-large-en-v1.5 model was used to convert the retrieved text into a 768-dimensional vector. Vector similarity retrieval was performed in the Milvus vector database, using Euclidean distance to calculate similarity and identifying the top K semantically most similar entries as candidate entries. For example, the number of candidate entries returned was set to 10. Semantic similarity retrieval solves the problem of low relevance in traditional keyword matching, and can retrieve rules with different expressions but related meanings, such as "do not harass family members" and "prohibit communication with third parties unrelated to the debt."

[0043] Considering that semantically similar entries may not be applicable in the current business scenario, this embodiment performs secondary filtering on candidate entries based on scene tags of voice data to obtain target entries in order to ensure that the retrieved rule entries match the actual business background of the current call (such as product type, overdue stage). This greatly improves the business relevance and accuracy of the entries and avoids misjudgments of "correct rules but wrong scenarios".

[0044] For example, using "scene_tag" retrieved from the security system as a filtering condition, entries matching "scene_tag" are selected as target entries from the 10 candidate entries returned by semantic retrieval. If fewer than 3 target entries are returned, the criteria are broadened to "similar scene tags".

[0045] Subsequently, more complex and precise rules are used to further refine the scoring and ranking of the initially selected target items, i.e., the target items are re-ranked, and matching items are determined based on the ranking results. The matching item can be the highest-ranked and most relevant item after re-ranking, and it is the authoritative basis ultimately selected to assist in decision-making. Re-ranking can be achieved through cross-encoder re-ranking or feature fusion ranking, and this embodiment of the invention does not specifically limit the specific methods used.

[0046] In other embodiments, the target entries are reordered, and matching entries are determined based on the reordering results, as well as the relevance of the matching entries. Specifically, step 123 includes: Step 123-1: Based on the re-ranking model, obtain the relevance, timeliness, and scenario matching degree of the target item; Step 123-2: Weight the relevance, timeliness, and scenario matching of the target item and sum them to obtain the comprehensive score of the target item; Step 123-3: Re-rank the target items based on the comprehensive score, and take the top-ranked target items and their relevance as matching items, as well as the relevance of the matching items.

[0047] Specifically, the re-ranking of target items can be achieved using a reranking model. For example, a reranking model could be the cross-encoder / ms-marco-MiniLM-L-6-v2 model. The target items are input into the cross-encoder / ms-marco-MiniLM-L-6-v2 model, which outputs a relevance score, a timeliness score, and a scene matching score for each target item.

[0048] Relevance refers to the semantic relevance score between the text content of the target entry and the searched text. Timeliness is a quantitative indicator of whether a knowledge base entry is outdated; newly released policies and clauses that are in effect have high timeliness, while old policies that have been repealed or updated have low timeliness. Scene matching refers to the degree of match between the scene tags of the target entry itself and the scene tags of the current voice data. Relevance, timeliness, and scene matching can be specifically represented by scores from 0 to 100.

[0049] This step involves in-depth mining and quantification of the multidimensional value of each target item, moving beyond the single dimension of semantic relevance to a comprehensive evaluation from three key perspectives: content relevance, policy timeliness, and business scenario fit. This provides rich and accurate input signals for the final decision.

[0050] Then, based on the weights of relevance, timeliness, and scenario matching, a weighted sum is performed to obtain the final score, which integrates the three major factors of content, time, and scenario—the comprehensive score of the target item. By adjusting the weights, different business needs can be flexibly adapted. For example, during periods of strict regulation, the weight of timeliness can be increased to ensure that the system prioritizes the latest and strictest policies.

[0051] For example, the overall score is calculated as "Relevance × 60% + Timeliness × 30% + Scenario Matching × 10%", and the top 3 items with the highest overall scores are used as the final RAG search criteria, i.e., the matching items. Matching items are those with the most relevant content, the latest policies, and those that best fit the current business scenario. The "Relevance Score" and "Overall Score" of each matching item are recorded in the `rag_log` table.

[0052] This embodiment improves the accuracy and credibility of subsequent quality inspection judgments by using multi-dimensional quantitative evaluation and configurable weighted fusion to elevate target items into optimal, business-oriented authoritative evidence. Step 130: Determine whether the IDK reinforcement mechanism is triggered based on the matching entries. If not, call the trained Direct Preference Optimization (DPO) model to score the preference matching degree of the initial quality inspection results and obtain the DPO matching degree of the initial quality inspection results. The DPO model is trained based on the sample preference text pairs in historical quality inspection cases.

[0053] Specifically, in order to further enhance the search results, an IDK enhancement mechanism is introduced in this step. Here, the main purpose is to determine whether to trigger the IDK enhancement mechanism by judging whether the matched entry is a valid entry.

[0054] If the overall score of the matched item is less than the overall score threshold, or the confidence level of the large model is less than the confidence threshold, or the expiration time of the matched item is less than the current time, then the IDK reinforcement mechanism is determined to be triggered; otherwise, it is determined not to be triggered.

[0055] Here, the overall scoring threshold and confidence threshold are determined based on a large number of historical cases. For example, the overall scoring threshold can be set to 60 points, and the confidence threshold can be set to 70 points.

[0056] If the overall score of a matching entry is less than 60 points, it means that the entry is either not relevant enough, outdated, or does not match the scenario, and is not worth using as a reliable basis for subsequent decision-making, i.e., there is no effective retrieval basis.

[0057] The confidence score of the large model refers to the degree of certainty the large model itself has about its output results, which can be obtained from the response of the large model interface that generates the initial quality inspection results. If the confidence score of the large model is less than 70 points, it indicates that the model itself is uncertain in its judgment, and the reliability of its results is questionable. Therefore, the entries retrieved based on this result lose their meaning for verification.

[0058] The expiration time of a matching entry is the date and time marked on each rule or policy entry in the knowledge base, indicating when the rule began to expire. If the expiration time of a matching entry is less than the current time, it means that the rule or policy has expired and is invalid, and there is no valid basis for its expiration.

[0059] Understandably, the IDK enhancement mechanism is triggered when any of the above conditions are met. After the IDK enhancement mechanism is triggered, the method also includes: The final quality inspection result of the voice data is determined to be pending review, and the pending review result is prohibited from being sent back to the business system; knowledge blind spot information is added to the database, and corresponding reminder information is pushed to the terminal of the knowledge base maintenance personnel.

[0060] When the IDK enhancement mechanism is triggered, it indicates that the system is unable to make a high-confidence automatic judgment due to its own capability limitations or insufficient data. It will proactively declare "I don't know" and hand over the decision-making power to a human, while recording the unknown problem for future improvement.

[0061] At this point, the final quality inspection result for the voice data is "Pending Review," meaning it requires manual confirmation and cannot be determined at this time. For quality inspection results marked "Pending Review," it is prohibited to send them back to the business system for evaluation purposes; instead, they should be routed to a dedicated pool of manual review work orders.

[0062] Knowledge blind spot information refers to key metadata captured during this quality inspection process that prevents the system from making a judgment; it is usually a structured problem description. For example, a knowledge blind spot record table can be created in the database, and knowledge blind spot information can be added to this table. At the same time, corresponding reminder information can be pushed to the terminals of knowledge base maintenance personnel to achieve closed-loop management and proactive maintenance.

[0063] This embodiment utilizes the flywheel effect of "blind spot discovery - manual correction - knowledge base update - capability enhancement" to enable the system to quickly adapt to business changes and gradually reduce future reliance on manual review.

[0064] When the IDK reinforcement mechanism is not triggered, the trained Direct Preference Optimization (DPO) model is called to score the preference matching degree of the initial quality inspection results, and the DPO matching degree of the initial quality inspection results is obtained.

[0065] Preference matching score compares the initial quality inspection results with the retrieved valid rules, simulating the decision-making process of human experts, and assesses whether the initial results align with expert preferences in terms of "compliance assurance" and "conformity of precedents." The aim is to perform a secondary calibration of the model results from the perspective of human preferences.

[0066] By learning from a large number of expert preferences, the DPO model can keenly identify unreasonable and non-compliant biases generated by the model, and directly correct semantic-level judgment biases, thus significantly reducing false positives and false negatives.

[0067] The DPO model is trained based on sample preference text pairs from historical quality inspection cases. 5000 "high-quality results" and 5000 "incorrect results" are selected from historical quality inspection cases, forming 5000 preference pairs (high-quality result text, incorrect result text). The preference pair texts are then cleaned (removing meaningless characters), segmented (using jieba segmentation), and encoded (using BPE encoding) to convert them into a format that the model can input. A large language model, LLM (13B parameters), is used as the base model, with pre-trained weights loaded. The training epochs are 5, the learning rate is 2e-6, the batch size is 16, the optimizer is AdamW, and the loss function is DPO loss. After each training epoch, the "preference alignment accuracy" (the proportion of high-quality results selected by the model) is calculated on the validation set (1000 preference pairs). Training stops when the accuracy is ≥92%, and the fine-tuned DPO model weights are saved, resulting in the DPO model. The initial quality inspection results are input into the trained Direct Preference Optimization (DPO) model, which outputs the DPO matching score.

[0068] In other embodiments, step 130 involves calling a trained Direct Preference Optimization (DPO) model to score the preference matching degree of the initial quality inspection results, thereby obtaining the DPO matching degree of the initial quality inspection results. Specifically, this includes: Step 131: Based on the DPO model, calculate the initial matching score of the initial quality inspection results; Step 132: If the initial matching score is less than the matching score threshold, adjust the initial quality inspection result and calculate the adjusted matching score until the adjusted matching score is greater than or equal to the matching score threshold or the number of adjustments meets the preset number of times. Step 133: Use the final adjusted matching score as the DPO matching score of the initial quality inspection result.

[0069] Specifically, after the initial quality inspection results are generated, the DPO model is called to score the preference matching degree of the result text and output the initial matching degree score, which can be set to 0-100 points. The higher the score, the more it matches the business preference.

[0070] The matching score threshold here can be set to 80 points. For example, if the initial matching score is <80 points, the initial judgment statement will be automatically adjusted to make the result more consistent with the characteristics of the preferred dataset. After adjustment, the matching score is recalculated until the matching score is ≥80 points or the number of adjustments reaches 3 (to avoid infinite loops). At this point, the matching score calculation is terminated, and the matching score after the last adjustment is taken as the final DPO matching score. Since the scoring is based on the perspective of a "human expert," the output DPO matching score is one of the core indicators for measuring the credibility of the results.

[0071] Step 140: Based on the relevance of the matching entries, the confidence of the large model, the matching degree of DPO, and the accuracy of similar historical cases, determine the accuracy confidence of the initial quality inspection results, and determine the final quality inspection results of the speech data based on the accuracy confidence.

[0072] Specifically, this embodiment integrates signals from four dimensions: policy basis, model self-awareness, expert preferences, and historical similar cases. Through weighted fusion, it calculates a comprehensive and reliable overall confidence score, completely solving the problem of "no quantitative standard for result credibility." Ultimately, based on this score, automated graded processing of results is achieved, thereby realizing the refined operational goal of automating high-credibility results and manually processing low-credibility results.

[0073] The average comprehensive score of multiple matching items can be calculated based on the relevance of the matching items; the confidence of the large model is obtained from the response of the large model interface that generates the initial quality inspection results; the DPO matching degree is obtained through step 130; and the accuracy of historical similar cases can be obtained by querying the database.

[0074] The rationality of different weight combinations was verified through A / B testing, and the final weights were determined: Average overall score of matched items: weighted at 40%; Large model confidence (llm_confidence): weight 30%; DPO matching score (final matching score): weight 20%; Accuracy of similar historical cases (accuracy of quality inspection results for the same scenario and judgment type in the past 3 months): weight 10%.

[0075] The confidence level of accuracy can be expressed by the formula: Accuracy confidence = (average overall score of matched items × 40% + large model confidence × 30% + DPO matching degree × 20% + historical accuracy × 10%) / 100 × 100 (rounded to the nearest integer); Example: The average overall score of the matched items = 90 points, the confidence level of the large model = 85 points, the DPO matching degree = 92 points, and the historical accuracy = 95 points. Then the accuracy confidence level = (90×0.4+85×0.3+92×0.2+95×0.1) = 36+25.5+18.4+9.5 = 89.4 → 89 points.

[0076] The calculation results of accuracy confidence are stored in the "result_confidence" table, and "result_id" is associated with the scores of each parameter to support subsequent traceability.

[0077] Based on this, the final quality inspection result of the voice data is determined according to the accuracy confidence level, including: If the accuracy confidence level is within the first confidence level range, a final quality inspection result consistent with the initial quality inspection result is generated, the matching item is used as the reference, and the result is sent back to the business system. If the accuracy confidence level falls within the second confidence level range, the final quality inspection result and review recommendations are generated and sent back to the business system. If the accuracy confidence level falls within the third confidence level range, then the final quality inspection result is determined to be pending review.

[0078] Specifically, multiple confidence intervals can be pre-defined, and different subsequent processes can be automatically executed based on the interval in which the calculated accuracy confidence level falls. For example, the confidence level can be divided into three intervals: a high confidence interval (confidence level ≥ 90 points) and a low confidence interval (confidence level < 70 points).

[0079] If the accuracy confidence level is in the high confidence range (confidence level ≥ 90 points), a "final_conclusion" (consistent with the initial conclusion, with additional RAG references) will be automatically generated and synchronized to the business system through the result feedback interface, marked as "can be directly used for salesperson evaluation".

[0080] If the accuracy confidence level falls within the medium confidence range (70 points ≤ confidence level < 90 points), then a "final_conclusion" is generated, along with a "review_suggestion", and sent back to the business system, marked as "requires selective review".

[0081] If the accuracy confidence level is less than 70 points, it will be marked according to the IDK processing logic, and the final quality inspection result will be determined as pending review. It will not be sent back to the business system assessment module, but will only be pushed to the quality inspector's workbench, including a knowledge blind spot prompt.

[0082] It should be noted that the "manual review correction rate" of the results at each level can be calculated weekly. If the correction rate is greater than 5%, the confidence parameter weights are adjusted. The preference dataset is updated monthly, and the DPO model is fine-tuned to ensure that the model adapts to business changes.

[0083] Based on the above embodiments, Figure 2 This is the second flowchart of the intelligent quality inspection method based on RAG and dual reinforcement mechanism provided by the present invention, as shown below. Figure 2 As shown, the method includes: S101, Agent Framework and Core Module Configuration. This section describes how to build an intelligent quality inspection agent framework on an intelligent agent platform and configure three core modules: (1) Dialogue flow management module: Configure process nodes by dragging and dropping through the platform's visual interface, define the closed-loop process logic of "quality inspection result generation → RAG retrieval and verification → double reinforcement adjustment → confidence judgment → result output", and set the data transmission format between nodes.

[0084] (2) Knowledge base module: Import structured knowledge and build a three-layer knowledge system of "policy-rules-case".

[0085] (3) Long-term memory module: Stores historical quality inspection standards, RAG retrieval logs, and dual reinforcement parameters (such as IDK trigger thresholds and DPO preference data). Create three tables in MySQL: history_standard, rag_log, and reinforcement_params.

[0086] S102, Business System Interface Integration. Develop standardized HTTP / HTTPS interfaces to enable bidirectional data interaction between the Agent and the business system. (1) Data retrieval: Automatically retrieve basic data such as audio transcription text, customer identifier (after anonymization), and voice time period. Input parameters include customer_id (customer identifier), transcript_text (transcribed text), collection_time (collection time), and scene_tag (scene tag); output parameters include pull_status (success / failure) and data_id (unique data identifier).

[0087] (2) Result Feedback: Synchronize the quality inspection results containing "accuracy confidence level" and "judgment basis (knowledge base reference entries)", and support business systems to view the original policies / rules in conjunction with them. Output parameters include result_id, customer_id (customer identifier), final_conclusion (final quality inspection conclusion), accuracy_confidence (accuracy confidence level), reference_entries (referenced knowledge base entries, including ID and content), review_suggestion (review suggestion, exclusive to medium-confidence results), and result_level (result level: high / medium / pending review).

[0088] Interface security configuration: Token authentication (the token is automatically updated every 15 days and is periodically synchronized between the business system and the Agent), data transmission encryption (HTTPS protocol), and interface access frequency limit (≤100 times / second) are used to prevent data leakage or interface abuse.

[0089] S103, a multi-dimensional search strategy. Breaking away from traditional keyword search, it adopts a dual-dimensional search approach combining semantics and context. Search results are reordered, with the top 3 entries used as verification criteria, and entries with a relevance of ≥85% marked as "highly reliable evidence."

[0090] S104, IDK enhancement mechanism implementation. When RAG search finds no relevant entries, the confidence level of the large model in judging the quality inspection item is <70%, and there is no corresponding policy / rule in the knowledge base, the IDK mechanism is triggered; the quality inspection item is marked as pending review, no definitive judgment result is generated, and knowledge blind spots are recorded, such as "new 'student customer preservation compliance' not covered", and pushed to the knowledge base update list.

[0091] S105, DPO reinforcement mechanism implementation. After the initial quality inspection results are generated, the DPO model is called to score the preference matching degree of the result text. If the matching degree is <80 points, the judgment statement of the initial result is automatically adjusted to make the result more consistent with the characteristics of the preference dataset. After adjustment, the matching degree is recalculated to meet the conditions.

[0092] S106, Accuracy Confidence Calculation. Construct a multi-dimensional confidence assessment model. Input parameters include "RAG retrieval relevance (weight 40%), large model judgment confidence (weight 30%), DPO enhanced matching degree (weight 20%), and historical similar case accuracy (weight 10%)", and output a confidence score of 0-100.

[0093] S107, Result Classification and Output. Confidence score ≥ 90: Marked as "Highly Confidential Result" and directly sent back to the business system application; 70 score ≤ Confidence score < 90: Marked as "Mediumly Confidential Result" and an "Assigned Review Point" is added; Confidence score < 70: The IDK mechanism is triggered, marked as "Pending Manual Review," and direct application is prohibited.

[0094] The method provided in this invention, through the collaborative design of "RAG precise retrieval + dual enhancement mechanism + confidence closed loop," achieves four core breakthroughs compared to existing technologies, with the following beneficial effects: 1. Significantly reduced illusions in large models and greatly improved accuracy: The IDK enhancement mechanism can 100% avoid fictitious judgments due to "knowledge blind spots". Combined with DPO enhancement to align with business preferences, it significantly reduces the misjudgment rate of quality inspection results; RAG multi-dimensional retrieval and re-ranking improve the accuracy of matching judgment criteria, which is significantly higher than the accuracy of traditional keyword retrieval, avoiding misjudgments caused by bias in criteria.

[0095] 2. Improved business adaptation efficiency and reduced compliance risks: When policies / rules are updated, only knowledge base entries need to be added and imported. There is no need to re-annotate a large amount of data to fine-tune the large model. The adaptation cycle is shortened from 10 days to 1 hour, ensuring that quality inspection standards are synchronized with business. The long-term memory module stores historical quality inspection standards and supports cross-version result comparison, avoiding compliance traceability risks caused by standard iteration.

[0096] 3. Optimized workload and improved quality inspection efficiency: The confidence level classification process ensures that the proportion of highly reliable results reaches more than 75%, eliminating the need for manual review and reducing the workload compared to pure manual verification; "Medium-reliable results" are accompanied by "suggested review points" to guide quality inspectors to focus on key question items, reducing the review time for a single result from 10 minutes to 3 minutes, thus improving the overall quality inspection efficiency.

[0097] 4. Quantifiable reliability and enhanced traceability of results: Accuracy confidence provides a quantitative reliability standard for quality inspection results, solving the problem of no basis for judging the quality of existing technical results, and facilitating the application of business systems according to confidence level; all quality inspection results are associated with "RAG retrieval basis (knowledge base entry ID)" and "dual reinforcement adjustment record", supporting full-chain traceability and meeting the regulatory requirement of "traceable quality inspection results".

[0098] The intelligent quality inspection device based on RAG and dual reinforcement mechanism provided by the present invention will be described below. The intelligent quality inspection device based on RAG and dual reinforcement mechanism described below can be referred to in correspondence with the intelligent quality inspection method based on RAG and dual reinforcement mechanism described above.

[0099] Figure 3 This is a schematic diagram of the intelligent quality inspection device based on RAG and dual reinforcement mechanism provided by the present invention, as shown below. Figure 3 As shown, the device includes: The text determination unit 310 determines the search text based on the transcribed text of the speech data to be inspected and the initial quality inspection results; The knowledge retrieval unit 320 retrieves matching entries and the relevance of the matching entries from a preset knowledge base based on the semantics of the retrieved text and the scene tags of the voice data. Preference scoring unit 330 determines whether the IDK enhancement mechanism is triggered based on the matching item. If it is not triggered, it calls the trained Direct Preference Optimization (DPO) model to score the preference matching degree of the initial quality inspection result and obtains the DPO matching degree of the initial quality inspection result. The DPO model is trained based on the sample preference text pairs in historical quality inspection cases. The result determination unit 340 determines the accuracy confidence of the initial quality inspection result based on the relevance of the matching entries, the confidence of the large model, the matching degree of the DPO, and the accuracy of historical similar cases, and determines the final quality inspection result of the voice data based on the accuracy confidence.

[0100] Based on the above embodiments, the knowledge retrieval unit is specifically used for: Candidate entries are selected from the entries based on the semantic similarity between the searched text and each entry in the knowledge base; Based on the scene tags of the voice data, target entries are obtained by filtering from the candidate entries; The target entries are reordered, and the matching entries and their relevance are determined based on the reordering results.

[0101] Based on the above embodiments, the knowledge retrieval unit is specifically used for: Based on the reordering model, the relevance, timeliness, and scenario matching degree of the target item are obtained; The relevance, timeliness, and scenario matching of the target item are weighted and summed to obtain the comprehensive score of the target item; Based on the comprehensive score, the target items are reordered, and a preset number of target items at the top of the ranking and their relevance are used as the matching items, along with the relevance of the matching items.

[0102] Based on the above embodiments, the preference scoring unit is specifically used for: If the overall score of the matched item is less than the overall score threshold, or the confidence level of the large model is less than the confidence threshold, or the expiration time of the matched item is less than the current time, then the IDK reinforcement mechanism is determined to be triggered; otherwise, it is determined not to be triggered.

[0103] Based on the above embodiments, the device further includes a result determination unit for: The final quality inspection result of the voice data is determined to be pending review, and the pending review result is prohibited from being sent back to the business system; Add knowledge gap information to the database and push corresponding reminder messages to the knowledge base maintenance personnel's terminals.

[0104] Based on the above embodiments, the preference scoring unit is specifically used for: Based on the DPO model, calculate the initial matching score of the initial quality inspection results; If the initial matching score is less than the matching score threshold, the initial quality inspection result is adjusted and the adjusted matching score is calculated until the adjusted matching score is greater than or equal to the matching score threshold or the number of adjustments meets the preset number. The final adjusted matching score is used as the DPO matching score of the initial quality inspection result.

[0105] Based on the above embodiments, the result determination unit is specifically used for: If the accuracy confidence level falls within the first confidence level range, a final quality inspection result consistent with the initial quality inspection result is generated, the matching entry is used as the reference, and the result is sent back to the business system. If the accuracy confidence level falls within the second confidence level range, then the final quality inspection result and review recommendations are generated and sent back to the business system; If the accuracy confidence level falls within the third confidence level range, then the final quality inspection result is determined to be pending review.

[0106] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4 As shown, the electronic device may include: a processor 410, a communications interface 420, a memory 430, and a communications bus 440, wherein the processor 410, the communications interface 420, and the memory 430 communicate with each other through the communications bus 440. The processor 410 can call logic instructions in the memory 430 to execute an intelligent quality inspection method based on RAG and dual reinforcement mechanisms. This method includes: determining a retrieval text based on the transcribed text of the speech data to be inspected and the initial quality inspection result; retrieving matching entries from a preset knowledge base based on the semantics of the retrieval text and the scene tags of the speech data, as well as the relevance of the matching entries; determining whether the IDK reinforcement mechanism is triggered based on the matching entries; if not triggered, calling a trained Direct Preference Optimization (DPO) model to score the preference matching degree of the initial quality inspection result, obtaining the DPO matching degree of the initial quality inspection result; the DPO model being trained based on sample preference text pairs from historical quality inspection cases; determining the accuracy confidence of the initial quality inspection result based on the relevance of the matching entries, the large model confidence, the DPO matching degree, and the accuracy of similar historical cases; and determining the final quality inspection result of the speech data based on the accuracy confidence.

[0107] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0108] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the intelligent quality inspection method based on RAG and dual reinforcement mechanism provided by the above methods. The method includes: determining a retrieval text based on the transcribed text of the speech data to be inspected and the initial quality inspection result; retrieving matching entries and the relevance of the matching entries from a preset knowledge base based on the semantics of the retrieval text and the scene tags of the speech data; determining whether the IDK reinforcement mechanism is triggered based on the matching entries; if not triggered, calling a trained Direct Preference Optimization (DPO) model to score the preference matching degree of the initial quality inspection result to obtain the DPO matching degree of the initial quality inspection result, wherein the DPO model is trained based on sample preference text pairs in historical quality inspection cases; determining the accuracy confidence degree of the initial quality inspection result based on the relevance of the matching entries, the large model confidence, the DPO matching degree, and the accuracy of similar historical cases, and determining the final quality inspection result of the speech data based on the accuracy confidence degree.

[0109] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the intelligent quality inspection method based on RAG and dual reinforcement mechanisms provided by the above methods. The method includes: determining a retrieval text based on the transcribed text of the speech data to be inspected and the initial quality inspection result; retrieving matching entries and the relevance of the matching entries from a preset knowledge base based on the semantics of the retrieval text and the scene tags of the speech data; determining whether the IDK reinforcement mechanism is triggered based on the matching entries; if not triggered, calling a trained Direct Preference Optimization (DPO) model to score the preference matching degree of the initial quality inspection result to obtain the DPO matching degree of the initial quality inspection result, wherein the DPO model is trained based on sample preference text pairs in historical quality inspection cases; determining the accuracy confidence degree of the initial quality inspection result based on the relevance of the matching entries, the large model confidence, the DPO matching degree, and the accuracy of similar historical cases, and determining the final quality inspection result of the speech data based on the accuracy confidence degree.

[0110] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown 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 according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0111] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0112] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent quality inspection method based on RAG and double reinforcement mechanism, characterized in that, The method comprises the following steps: Based on the transcription text of the voice data to be quality inspected and the initial quality inspection result, a retrieval text is determined; Based on the semantics of the retrieval text and the scene label of the voice data, a matching item is retrieved from a preset knowledge base, and the relevance of the matching item is determined; Based on the matching item, it is determined whether to trigger an IDK reinforcement mechanism. If not, a trained direct preference optimization (DPO) model is called to perform preference matching degree scoring on the initial quality inspection result to obtain the DPO matching degree of the initial quality inspection result. The DPO model is trained based on sample preference text in historical quality inspection cases; Based on the relevance of the matching item, the large model confidence, the DPO matching degree, and the accuracy of historical similar cases, the accuracy confidence of the initial quality inspection result is determined, and the final quality inspection result of the voice data is determined based on the accuracy confidence.

2. The intelligent quality inspection method based on RAG and double reinforcement mechanism according to claim 1, characterized in that, The method comprises the following steps: Based on the semantics of the retrieval text and the scene label of the voice data, a matching item is retrieved from a preset knowledge base, and the relevance of the matching item is determined; Based on the semantic similarity between the retrieval text and each item in the knowledge base, candidate items are selected from the items; Based on the scene label of the voice data, target items are selected from the candidate items; 3. The intelligent quality inspection method based on RAG and double reinforcement mechanism according to claim 2, characterized in that, The target items are reordered, and the matching items are determined based on the sorting results, and the relevance of the matching items is determined. The method comprises the following steps: Based on a reordering model, the relevance, timeliness, and scene matching degree of the target items are obtained; The relevance, timeliness, and scene matching degree of the target items are weighted and summed to obtain the comprehensive score of the target items; 4. The intelligent quality inspection method based on RAG and double reinforcement mechanism according to claim 1, characterized in that, The target items are reordered based on the comprehensive score, and the top preset number of target items and their relevance are selected as the matching items and the relevance of the matching items. The method comprises the following steps:

5. The intelligent quality inspection method based on RAG and double reinforcement mechanism according to claim 4, characterized in that, If the comprehensive score of the matching item is less than a comprehensive score threshold, or the large model confidence is less than a confidence threshold, or the invalidation time of the matching item is less than the current time, it is determined that the IDK reinforcement mechanism is triggered; otherwise, it is determined that the IDK reinforcement mechanism is not triggered. After determining whether to trigger the IDK reinforcement mechanism, the method further comprises the following steps: The final quality inspection result of the voice data is determined to be pending review, and the pending review result is prohibited from being returned to the business system; 6. The RAG and double reinforcement mechanism based intelligent quality inspection method according to any one of claims 1 to 5, characterized in that, Knowledge blind area information is added to the database, and corresponding reminder information is pushed to the knowledge base maintenance personnel terminal. The method comprises the following steps: Based on the DPO model, the initial matching degree score of the initial quality inspection result is calculated; If the initial matching degree score is less than a matching degree score threshold, the initial quality inspection result is adjusted and the adjusted matching degree score is calculated until the adjusted matching degree score is greater than or equal to the matching degree score threshold or the adjustment times meet a preset number of times. The last adjusted matching degree score is taken as a DPO matching degree of the initial quality inspection result.

7. The RAG and double reinforcement mechanism based intelligent quality inspection method according to any one of claims 1 to 5, characterized in that, The final quality inspection result of the voice data is determined based on the accuracy confidence, including: If the accuracy confidence belongs to a first confidence interval, a final quality inspection result consistent with the initial quality inspection result is generated, the matching item is taken as a reference basis, and is returned to a business system; If the accuracy confidence belongs to a second confidence interval, a final quality inspection result and a review suggestion are generated, and are returned to the business system; If the accuracy confidence belongs to a third confidence interval, the final quality inspection result is determined to be pending review.

8. An intelligent quality inspection device based on RAG and double reinforcement mechanism, characterized in that, including: The text determination unit is configured to determine a search text based on a transcription text of the voice data to be inspected and the initial quality inspection result; The knowledge search unit is configured to search for a matching item and a relevance of the matching item from a preset knowledge base based on semantics of the search text and a scene label of the voice data; The preference score unit is configured to determine whether an IDK reinforcement mechanism is triggered based on the matching item, and if not, to call a trained direct preference optimization (DPO) model to perform preference matching degree scoring on the initial quality inspection result to obtain a DPO matching degree of the initial quality inspection result, the DPO model being trained based on sample preference texts in historical quality inspection cases; The result determination unit is configured to determine an accuracy confidence of the initial quality inspection result based on the relevance of the matching item, a large model confidence, the DPO matching degree, and an accuracy of a historical similar case, and to determine a final quality inspection result of the voice data based on the accuracy confidence.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and running on the processor, characterized in that, The processor executes the computer program to implement the intelligent quality inspection method based on RAG and double reinforcement mechanisms according to any one of claims 1 to 7. 10.A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the intelligent quality inspection method based on RAG and double reinforcement mechanisms according to any one of claims 1 to 7.

Images