Intelligent dialogue knowledge base optimization method based on multi-path recall fusion and self-updating

Abstract

The invention discloses an intelligent dialogue knowledge base optimization method based on multi-path recall fusion and self-updating, and relates to the technical field of artificial intelligence. The method comprises the following steps that A, a multi-path parallel recall engine is constructed, a knowledge base is retrieved in parallel through a keyword matching module, a semantic vector matching module and a user portrait matching module, and the keyword matching module positions keyword segments through inverted indexes; according to the method, a three-way parallel recall engine for keyword matching, semantic vector matching and user portrait matching is constructed, multi-dimensional accurate retrieval of user questions is achieved, a keyword matching module adopts an inverted index to rapidly position keyword segments, and coverage of basic semantics is ensured; the semantic vector matching module calculates cosine similarity through a Sension-BERT model, and captures deep semantic association between a user question and a knowledge base fragment; the user portrait matching module generates a retrieval weight based on a user identity dynamic loading rule base.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence, and particularly relates to an intelligent dialogue knowledge base optimization method based on multi-path recall fusion and self-updating. BACKGROUND

[0002] At present, intelligent dialogue applications based on large models are rapidly popularized in colleges and enterprises, but the intelligent dialogue systems based on retrieval augmented generation (RAG) generally have low information retrieval accuracy, mainly due to the following two aspects: first, in the intelligent dialogue process, due to the difference of user identity background and the diversity of query problems, the system administrator is difficult to cover all potential query contents, which may cause the problem that the user's query cannot be completely matched with the knowledge, and meanwhile, the knowledge base maintenance lacks long-term and timely demand tracking mechanism and knowledge base updating strategy, finally resulting in the problems of incomplete knowledge coverage and insufficient answer quality. Second, the diversity of knowledge base content, the knowledge base contains conflicting information, the key content is too scattered or missing, and the context logic is destroyed due to blockage.

[0003] Therefore, the application provides an intelligent dialogue knowledge base optimization method based on multi-path recall fusion and self-updating to solve the above problems. SUMMARY

[0004] The application aims to provide an intelligent dialogue knowledge base optimization method based on multi-path recall fusion and self-updating, which solves the problem of low information retrieval accuracy of the intelligent dialogue system in the prior art by constructing a multi-path parallel recall engine and a knowledge base self-updating cooperation.

[0005] To solve the above technical problems, the application is implemented by the following technical scheme.

[0006] The application is an intelligent dialogue knowledge base optimization method based on multi-path recall fusion and self-updating, comprising the following steps:

[0007] A, constructing a multi-path parallel recall engine:

[0008] The keyword matching module, the semantic vector matching module and the user portrait matching module are used to retrieve the knowledge base in parallel, wherein the keyword matching module uses an inverted index to locate the keyword fragment;

[0009] The semantic vector matching module uses a Sentence-BERT model to calculate the cosine similarity between the user's question and the knowledge base fragment;

[0010] The user portrait matching module dynamically loads a preset rule base based on the user's identity to generate an identity-related retrieval weight;

[0011] B, multi-path recall result fusion:

[0012] The RRF algorithm is used to weight and fuse the three-way recall results, and output a Top-N candidate set;

[0013] C. Dynamic reordering:

[0014] The fused candidate set is input into a cross-encoder reordering model (Cross-Encoder) to generate a ranking sequence of the final answer sources;

[0015] D. QA knowledge base self-updating trigger:

[0016] Self-updating includes passive trigger (positive evaluation by the user) and active trigger (high-frequency question monitoring),

[0017] The former automatically generates QA pairs, and the latter needs to be updated after being reviewed by the administrator, and the update process is automatically started:

[0018] When the user gives a positive evaluation on the answer generated by the non-QA knowledge base (such as the large model generation), a new QA pair is automatically generated; if the existing QA pair is updated, it needs to be reviewed by the answer comparison module for consistency;

[0019] The consistency is reviewed by the answer comparison evaluation module (using a large model to compare the quality of new and old answers); after the review is passed, the QA knowledge base is inserted and a timestamp is marked.

[0020] The application further sets that the data in step A includes converting PDF, Word, PPT and Excel documents into Markdown structured text when processing, and deleting all formats of the text;

[0021] A timestamp metadata is added to each knowledge base entry;

[0022] The file containing the calculation rules is semantically blocked according to the # level title, and the PDF, Word, PPT and Excel are converted into Markdown structured text and a timestamp is added, which improves the structured degree and version management capability of the knowledge base content, and reduces the retrieval error caused by format difference.

[0023] The application further sets that when the semantic vector matching module is working in step A, multiple rewritten questions are generated in combination with the user's identity and the current time when the user's question has ambiguity;

[0024] The rewritten questions are input into the multi-way recall engine, and the recall results are merged and reordered, and the rewritten questions are generated in combination with the user's identity and the time in the semantic vector matching module, which solves the recall deviation caused by ambiguity and improves the matching accuracy in complex problem scenarios.

[0025] The application further sets that the multi-way recall priority strategy in step B is:

[0026] Prior search QA knowledge base, if the similarity exceeds the threshold value theta, terminate other recall;

[0027] The QA knowledge base adopts a three-column structure for storage: standard question, answer, and question variant, the QA knowledge base is searched preferentially, and a similarity threshold value is set, so that the high-frequency question is responded quickly, and the three-column structure storage (standard question, answer, and question variant) enhances the expansibility and reusability of the knowledge base.

[0028] The application further sets that the RRF fusion algorithm formula in the step B is:

[0029] Wherein k is the number of recall engines, r i is the ranking of the candidate segment in the i-th path recall, and beta is a harmonic parameter, the application of the RRF algorithm effectively fuses the multi-path recall results, avoids the limitation of a single recall path, and improves the diversity and accuracy of the candidate answer.

[0030] The application further sets that the QA knowledge base self-updating process in the step D includes:

[0031] For the question of the user positive evaluation, 3-5 kinds of question variants are automatically generated;

[0032] A standardized answer is generated through an answer integration optimization module (a large model merges multiple source answers);

[0033] When updating the QA knowledge base, a valid time stamp is marked, the latest version is used as the reference in the case of conflict, question variants and standardized answers are automatically generated in the self-updating process, multiple source answers are integrated and optimized, and the comprehensiveness and authority of the knowledge base content are ensured.

[0034] The application further sets that the multimedia processing mode of the data is:

[0035] The picture is converted into Base64 encoding and embedded in Markdown;

[0036] The video generates an abstract text through a multi-modal large model, and is stored in the format of

video abstract

[0037] The original position relationship label is reserved when the picture-text mixed content is blocked, the multimedia processing mode realizes the structured storage of unstructured data such as pictures and videos, the original position relationship label is reserved, and the retrieval efficiency and context integrity of the multi-modal content are improved.

[0038] The application further sets that in the step A, the MCP mechanism is used for structured data query when data is searched:

[0039] Build prompt word template: "Database table structure: {schema}, please convert the question into SQL:"

[0040] The SQL execution result and the text knowledge base recall result are jointly input into the large model to generate a final answer, and the structured data query MCP mechanism combines the SQL execution result and the text recall result, supports precise data retrieval in complex business scenarios, and for example, joint application of database query and document knowledge.

[0041] The application further provides that the operation process of the answer comparison and evaluation module in step D is: inputting the old answer A and the new answer B into the large model evaluation agent;

[0042] The evaluation prompt word is: "As a domain expert, compare answer A: {A} and answer B: {B}, select the better version and explain the reason";

[0043] When the preferred rate of B in the evaluation result is greater than 80%, trigger the knowledge base update, the answer comparison and evaluation module compares the new and old answers through the large model, quantifies the evaluation result and sets the preferred rate threshold, and ensures the controllability of the quality of the knowledge base update.

[0044] The application further provides that the high-frequency question monitoring mechanism needs to be set during data retrieval in step A:

[0045] The frequency of occurrence of the user question that does not match the QA knowledge base is counted;

[0046] When the weekly frequency exceeds the threshold N, automatically generate a knowledge base update task and push the administrator, the high-frequency question monitoring mechanism actively triggers the knowledge base update task by counting the frequency of occurrence of the unmatched question, realizes the demand-driven dynamic maintenance, and reduces the cost of manual intervention.

[0047] The application has the following beneficial effects.

[0048] 1、The application realizes multi-dimensional accurate retrieval of user questions by constructing a three-way parallel recall engine of keyword matching, semantic vector matching and user portrait matching, wherein the keyword matching module adopts an inverted index to quickly locate a keyword fragment, ensuring the coverage of basic semantics; the semantic vector matching module calculates the cosine similarity through a Sentence-BERT model to capture the deep semantic association between the user question and the knowledge base fragment; the user portrait matching module dynamically loads a rule library based on the user identity to generate a retrieval weight, solving the matching deviation problem caused by user background differences in traditional systems; after the three-way recall results are weighted and fused through an RRF algorithm, dynamic sorting is performed through a cross-encoder reordering model, which significantly improves the accuracy and relevance of the candidate answers, effectively solving the core problem of insufficient information retrieval accuracy of existing RAG systems.

[0049] 2、The application aims at the pain point of lack of long-term tracking mechanism in knowledge base maintenance, and designs an automatic QA knowledge base self-updating process, which includes passive triggering (user positive evaluation) and active triggering (high-frequency question monitoring). The former automatically generates QA pairs, and the latter needs to be updated by the administrator after review. The system automatically generates standardized QA pairs from original questions and answers, compares the quality of new and old answers through a large model, and inserts the knowledge base and marks the timestamp after the review is passed. At the same time, the system counts the frequency of occurrence of unmatched questions through the high-frequency question monitoring mechanism, triggers the administrator to intervene in the update task, and this closed-loop mechanism not only realizes the dynamic expansion of the knowledge base, but also combines multiple source answers through the answer integration optimization module to ensure the consistency and timeliness of the knowledge base content, effectively avoiding the problem of incomplete knowledge coverage and insufficient answer quality caused by the lag of manual maintenance in traditional systems. BRIEF DESCRIPTION OF DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed to be used in the description of the embodiments.

[0051] Figure 1 Workflow diagram of the intelligent dialogue knowledge base optimization method based on multi-path recall fusion and self-updating.

[0052] Figure 2 Data processing flowchart in the intelligent dialogue knowledge base optimization method based on multi-path recall fusion and self-updating.

[0053] Figure 3 Multi-path recall workflow diagram in the intelligent dialogue knowledge base optimization method based on multi-path recall fusion and self-updating.

[0054] Figure 4 QA knowledge base self-updating flowchart in the intelligent dialogue knowledge base optimization method based on multi-path recall fusion and self-updating.

[0055] Figure 5 Structured data query flowchart in the intelligent dialogue knowledge base optimization method based on multi-path recall fusion and self-updating. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the application will be described below with reference to the drawings in the embodiments of the application. The described embodiments are only a part of the embodiments of the application, not all embodiments.

[0057] Embodiment one

[0058] Please refer to Figures 1-5 The application is an intelligent dialogue knowledge base optimization method based on multi-path recall fusion and self-updating, which includes the following steps:

[0059] A, construct a multi-path parallel recall engine:

[0060] Retrieving the knowledge base in parallel through the keyword matching module, the semantic vector matching module, and the user portrait matching module, wherein the keyword matching module uses an inverted index to locate keyword fragments;

[0061] The semantic vector matching module uses a Sentence-BERT model to calculate the cosine similarity of the user's question and the knowledge base fragments;

[0062] The user portrait matching module dynamically loads a pre-set rule base based on the user's identity to generate identity-related search weights;

[0063] B. Multi-path recall result fusion:

[0064] The RRF algorithm is used to weight and fuse the three-way recall results to output a Top-N candidate set;

[0065] C. Dynamic reordering:

[0066] The fused candidate set is input into the Cross-Encoder reordering model to generate a ranking sequence of the final answer sources;

[0067] D. QA knowledge base self-updating trigger:

[0068] Self-updating includes passive triggering (user positive evaluation) and active triggering (high-frequency question monitoring),

[0069] The former automatically generates QA pairs, and the latter needs to be updated after being reviewed by the administrator, and the update process is automatically started:

[0070] When the user gives a positive evaluation to the answer generated by the non-QA knowledge base (such as a large model), a new QA pair is automatically generated; if an existing QA pair is updated, it needs to be reviewed for consistency by the answer comparison module;

[0071] The consistency is reviewed by the answer comparison evaluation module (using a large model to compare the quality of new and old answers); after the review is passed, the QA knowledge base is inserted and a timestamp is marked.

[0072] Specifically, the keyword matching module in step A uses inverted index technology to establish a mapping relationship between keywords and knowledge base fragments, and through word segmentation processing, the user's question is disassembled into a keyword set to quickly locate knowledge base fragments containing relevant keywords, achieving fast matching of basic semantics;

[0073] The semantic vector matching module encodes the user's question and the knowledge base fragments into fixed-length semantic vectors based on the Sentence-BERT model, quantifies the semantic correlation degree by calculating the cosine similarity, and captures deep semantic relationships that keyword matching cannot identify, such as synonym replacement or sentence conversion;

[0074] The user portrait matching module dynamically loads the preset rule library according to the user identity information (such as role, authority, historical behavior, etc.), generates a search weight related to the user identity, and preferentially recalls the knowledge base content that is more suitable for the user's demand, for example, preferentially displaying industry-specific solutions for enterprise customers;

[0075] In step B, the RRF algorithm is used to weight and fuse the three-way recall results, to balance the accuracy of keyword matching, the depth of semantic matching, and the personalized needs of user portraits by reconciling the ranking differences of different recall paths, and to output a fusion result set containing Top-N candidate answers;

[0076] In step C, the fused candidate set is input into the Cross-Encoder reordering model, which optimizes the answer sorting strategy through end-to-end training, and considers the context relevance, logical coherence, and user intent fit, to generate a priority sequence of the final answer source;

[0077] In step D, when the user gives a positive evaluation to the answer generated by the non-QA knowledge base (such as the large model generated answer), the system automatically starts the update process: the original question and the answer are standardized as a QA pair, and the consistency of the new and old answers is reviewed by the large model comparison evaluation module. After passing the review, it is inserted into the QA knowledge base and marked with a timestamp, ensuring the timeliness and accuracy of the knowledge base content;

[0078] The QA knowledge base, as a standardized storage module for high-frequency questions, needs to be initialized through a three-column structure (standard question + answer + variant) before first use, and supports threshold-based priority retrieval, which is different from the general document block-based dialogue knowledge base.

[0079] Embodiment Two

[0080] Please refer to Figures 1-5 On the basis of Embodiment One, the data is processed in Step A, including converting PDF, Word, PPT, and Excel documents into Markdown structured text, and deleting all formats in the text;

[0081] Add a timestamp metadata to each knowledge base entry;

[0082] Semantically block the file containing the calculation rules according to the # level headings;

[0083] When the user's question is ambiguous, the semantic vector matching module generates multiple rewritten questions in combination with the user's identity and the current time in Step A;

[0084] The rewritten questions are input into the multi-way recall engine, and the recall results are merged and reordered;

[0085] The multi-way recall priority strategy in Step B is:

[0086] Prior search QA knowledge base, if the similarity exceeds the threshold θ, terminate other recall;

[0087] The QA knowledge base uses a three-column structure to store: standard questions, answers, and question variants;

[0088] The RRF fusion algorithm formula in step B is: Where k is the number of recall engines, r i is the ranking of the candidate segment in the i-th recall, and β is the harmonic parameter;

[0089] The QA knowledge base self-updating process in step D includes:

[0090] For questions with positive user evaluations, automatically generate 3-5 question variants;

[0091] Generate standardized answers through the answer integration optimization module (large model merging multiple source answers);

[0092] When updating the QA knowledge base, mark the effective timestamp, and in case of conflict, use the latest version;

[0093] The multimedia processing method of data is:

[0094] Convert pictures to Base64 encoding and embed them in Markdown;

[0095] Generate summary text for videos through a multi-modal large model, and store it in the format

video summary

[0096] When mixing content blocks, keep the original position relationship labels;

[0097] In step A, when searching for data, use the structured data query MCP mechanism:

[0098] Build a prompt word template: "Database table structure: {schema}, please convert the question to SQL:";

[0099] The SQL execution result and the text knowledge base recall result are jointly input into the large model to generate the final answer;

[0100] The operation process of the answer comparison evaluation module in step D: input old answer A and new answer B into the large model evaluation agent;

[0101] Evaluation prompt: "As a domain expert, compare answers A: {A} and answer B: {B}, select the better version and explain the reason";

[0102] When the preferred rate of B in the evaluation result is >80%, trigger knowledge base update;

[0103] High-frequency problem monitoring mechanism needs to be set when data retrieval in step A:

[0104] Statistics the frequency of user questions that do not match the QA knowledge base;

[0105] When the weekly frequency exceeds the threshold N, automatically generate a knowledge base update task and push it to the administrator.

[0106] Specifically: By converting PDF, Word, PPT, and Excel to Markdown structured text and adding a timestamp, the degree of structuring and version management of the knowledge base content is improved, reducing retrieval errors caused by format differences. In the semantic vector matching module, the user's identity and time are combined to generate rewritten questions, solving the recall bias caused by ambiguity and improving the matching accuracy in complex problem scenarios. Prior retrieval of the QA knowledge base and setting of a similarity threshold ensure quick response to high-frequency problems. The three-column structure storage (standard question, answer, and question variant) enhances the extensibility and reusability of the knowledge base. The application of RRF algorithm effectively integrates multiple recall results, avoiding the limitations of a single recall path and improving the diversity and accuracy of candidate answers. In the self-update process, question variants and standardized answers are automatically generated, and through the integration and optimization of multi-source answers, the comprehensiveness and authority of the knowledge base content are ensured. The multi-media processing method realizes the structured storage of unstructured data such as pictures and videos, retains the original position relationship label, and improves the retrieval efficiency and context integrity of multi-modal content. The structured data query MCP mechanism combines SQL execution results with text recall results, supporting precise data retrieval in complex business scenarios, such as the joint application of database queries and document knowledge. The answer comparison and evaluation module compares new and old answers through a large model, quantifies the evaluation results, and sets a preferred rate threshold to ensure the quality controllability of knowledge base updates. The high-frequency problem monitoring mechanism actively triggers knowledge base update tasks by counting the frequency of unmatched questions, achieving demand-driven dynamic maintenance and reducing manual intervention costs.

[0107] The working principle of the present application is: user questions are retrieved from the knowledge base through three parallel engines of keyword matching, semantic vector matching, and user portrait matching, obtaining candidate answers from three levels of basic semantics, deep semantics, and personalized needs. The three recall results are weighted and fused by RRF algorithm, then input into the cross-encoder model for dynamic reordering to generate a priority sequence. The final answer is generated by a large model combined with structured data query results. User feedback (such as positive evaluation) triggers the QA knowledge base self-update process, and the system realizes dynamic expansion and quality optimization of the knowledge base through automatic generation of standardized QA pairs, large model comparison and review, timestamp marking, and high-frequency problem monitoring, forming a closed-loop management mechanism of "retrieval-generation-feedback-update".

[0108] The above disclosed preferred embodiments of the present application are only used to help explain the present application, the preferred embodiments do not describe all the details, and the present application is not limited to the specific embodiments described, the description of these embodiments is selected and described in the present specification in order to better explain the principles and practical application of the present application, so that the persons skilled in the art can well understand and utilize the present application.

Claims

1. An intelligent dialogue knowledge base optimization method based on multi-path recall fusion and self-updating, characterized in that: Comprise the following steps: A, construct multi-channel parallel recall engine: Retrieval of knowledge base through keyword matching module, semantic vector matching module, user portrait matching module, wherein the keyword matching module adopts inverted index to locate key word fragments; The semantic vector matching module uses the Sentence-BERT model to calculate the cosine similarity of user questions and knowledge base fragments; The user portrait matching module dynamically loads the pre-set rule base based on user identity to generate identity-related retrieval weights; B, multi-channel recall result fusion: Adopt RRF algorithm to weight and fuse three-channel recall results, and output Top-N candidate set; C, dynamic reordering: Input the fused candidate set into the cross-encoder reordering model (Cross-Encoder) to generate a sorting sequence of the final answer source; D, QA knowledge base self-updating trigger: Self-updating includes passive trigger (user positive evaluation) and active trigger (high-frequency question monitoring), the former automatically generates QA pairs, and the latter needs to be updated after administrator review, and automatically starts the update process: When the user gives a positive evaluation to the answer generated by the non-QA knowledge base (such as the large model), a new QA pair is automatically generated; If the existing QA pair is updated, it needs to pass through the answer comparison module to review the consistency of the new and old; Review the consistency through the answer comparison evaluation module (use the large model to compare the quality of new and old answers); After the review is passed, insert the QA knowledge base and mark the time stamp.

2. The method of claim 1, wherein: The data in step A includes converting PDF, Word, PPT and Excel documents into Markdown structured text, and deleting all formats of the text when processing; Add timestamp metadata to each knowledge base entry; Semantically block the files containing calculation rules according to the # level title.

3. The method of claim 1, wherein: When the semantic vector matching module works in step A, multiple rewritten questions are generated when the user's question is ambiguous, combined with the user's identity and the current time; Input the rewritten questions into the multi-channel recall engine, merge the recall results and reorder them.

4. The method of claim 1, wherein: The multi-channel recall priority strategy in step B is: Prioritize searching the QA knowledge base, and terminate other recalls if the matching similarity exceeds the threshold θ; The QA knowledge base uses a three-column structure to store: standard questions, answers, and question variants.

5. The method of claim 1, wherein: The RRF fusion algorithm formula in step B is: where k is the number of recall engines, r i is the ranking of the candidate segment at the i-th path of recall, and β is a harmonic parameter.

6. The method of claim 1, wherein: The QA knowledge base self-updating process in step D includes: For user positive evaluation questions, automatically generate 3-5 question variants; Generate a standardized answer through the answer integration optimization module (large model to merge multiple source answers); When updating the QA knowledge base, mark the effective timestamp, and use the latest version in case of conflict.

7. The method of claim 2, wherein: The multimedia processing method of the data is: Convert pictures to Base64 encoding and embed them in Markdown; Video abstract text is generated by a multi-modal large model, stored in 【video abstract】{URL} format; When block-mixing content, keep the original position relationship label.

8. The method of claim 1, wherein: In step A, when retrieving data, use the structured data query MCP mechanism: Construct a prompt word template: "Database table structure: {schema}, please convert the question to SQL:"; SQL execution results and text knowledge base recall results are jointly input into a large model to generate the final answer.

9. The method for intelligent dialog knowledge base optimization based on multi-path recall fusion and self-updating of claim 1, wherein: The operation process of the answer comparison evaluation module in step D: input the old answer A and the new answer B into the large model evaluation agent; Evaluation prompt: "As a domain expert, compare answer A: {A} and answer B: {B}, select the better version and explain the reason"; Trigger knowledge base update when the preferred rate of B in the evaluation result is > 80%.

10. The method of claim 1, wherein: High-frequency question monitoring mechanism needs to be set during data retrieval in step A: Statistical frequency of user questions that do not match the QA knowledge base; When the weekly frequency exceeds the threshold N, automatically generate a knowledge base update task and push it to the administrator.

Images