A brand information citability optimization method based on large model retrieval path

Abstract

The application discloses a brand information citability optimization method based on a large model retrieval path, and belongs to the technical field of artificial intelligence and big data. The brand information citability optimization method based on the large model retrieval path comprises the following steps: step 1, obtaining an information set of a target brand, and dividing the information set into multiple knowledge units; a knowledge dependency directed acyclic graph is established, nodes of the knowledge dependency directed acyclic graph are the knowledge units, and edges represent retrieval dependency, traceability dependency and reference dependency relationships; step 2, a multi-stage processing task chain is established for each knowledge unit, the multi-stage processing task chain at least comprises a retrieval task, an analysis task and a verification task, wherein the retrieval task is used for reading brand knowledge data from a network or a storage medium, and the analysis task is used for unpacking, standardizing, structuring or semantic embedding of the knowledge data. The application can effectively improve the authority and accuracy of brand information citation.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and big data technology, specifically to a method for optimizing the citationability of brand information based on large model retrieval paths. Background Technology

[0002] With the rapid popularization of large-scale model technology, generative AI has become a core channel for users to obtain information and filter services. Whether brand information can be accurately retrieved and reliably cited by large-scale models directly determines its "visibility" and market competitiveness within the large-scale model ecosystem. When answering brand-related questions, large-scale models typically retrieve brand information from pre-trained knowledge bases and real-time online search results, and generate answers after credibility assessment. In this process, the accuracy, authority, citation, and traceability of brand information directly affect users' perception and trust in the brand, and are also the core requirements for brands to achieve precise communication and avoid the risk of information distortion within the large-scale model ecosystem.

[0003] Based on the above, the inventors have discovered the following problems: Existing technologies lack a reliable retrieval path scheduling mechanism, making it difficult to guarantee the authority of retrieval sources. Large-scale model brand information retrieval largely relies on crawling the entire internet or pre-trained data, failing to perform hierarchical scheduling and filtering of retrieval path reliability. This prevents the prioritization of high-quality information channels such as official brand sources and authoritative media, and instead easily introduces unofficial, unauthoritative, and even false information. For example, some large-scale models cannot distinguish the priority between official brand websites and ordinary sites during retrieval, leading to the brand's core information being misled by unofficial content, thus affecting the authority and accuracy of information citations.

[0004] Therefore, in view of this, we study and improve the existing structure and its shortcomings, and provide a method for optimizing the citationability of brand information based on the large model retrieval path, in order to achieve a more practical value. Summary of the Invention

[0005] The purpose of this invention is to provide a method for optimizing the citationability of brand information based on large model retrieval paths, in order to solve the problem mentioned in the background art of the lack of a reliable retrieval path scheduling mechanism and the difficulty in guaranteeing the authority of retrieval sources in the existing technology.

[0006] In view of the above problems, the technical solution proposed by the present invention is as follows:

[0007] A method for optimizing the citationability of brand information based on a large model retrieval path includes the following steps: Step 1: Obtain the information set of the target brand and divide the information set into multiple knowledge units; Establish a knowledge dependency directed acyclic graph, where the nodes of the knowledge dependency directed acyclic graph are knowledge units, and the edges represent retrieval dependency, source dependency, and citation dependency relationships.

[0008] Step 2: Establish a multi-stage processing task chain for each knowledge unit. The multi-stage processing task chain includes at least a retrieval task, a parsing task, and a verification task. The retrieval task is used to read brand knowledge data from the network or storage medium. The parsing task is used to unpack, standardize, structure, or semantically embed the knowledge data. The verification task is used to complete the authoritative traceability verification of the parsed data and write it into the knowledge trust registration form.

[0009] Step 3: Obtain the current user's query intent and interaction input to determine the predicted retrieval scope; based on the predicted retrieval scope, determine candidate retrieval units from the semantic index, and extract candidate knowledge subgraphs from the knowledge dependency directed acyclic graph accordingly;

[0010] Step 4: Determine the key credible knowledge set in the candidate knowledge subgraph. The key credible knowledge set is used to satisfy the requirement that the first part of the large model output can be cited or to satisfy the minimum credible presentation of brand information within the predicted retrieval range.

[0011] Step 5: Determine critical path information based on the key trusted knowledge set, and generate task deadlines for the multi-stage processing task chain based on the critical path information. The critical path information is the information corresponding to the dependency chain with the largest remaining processing time in the key trusted knowledge set.

[0012] Step 6: Based on the task deadline and critical path information, determine the dispatch priority according to the task deadline and remaining processing time, and under the constraints of retrieval bandwidth budget, parsing budget and verification budget, sort and dispatch the retrieval task, parsing task and verification task across stages.

[0013] Step 7: When the multi-stage processing task chain cannot be completed within the task deadline, select a downgraded trusted knowledge unit for the corresponding knowledge unit and load the downgraded trusted knowledge unit first to ensure minimum trusted presentation.

[0014] Step 8: After a knowledge unit or a downgraded trusted knowledge unit completes the verification task, the reference output is updated by atomic replacement, so that the large model generation pipeline can switch to the verified knowledge unit without interrupting the output loop.

[0015] Step 9: Collect operational metrics during the processing and adaptively adjust at least one of the following based on the operational metrics: retrieval task concurrency depth, parsing task thread concurrency, verification budget per frame, prediction time range, and prediction retrieval quantity.

[0016] Furthermore, knowledge units include:

[0017] The knowledge units include: Brand Body Knowledge Unit, Product Parameter Knowledge Unit, Qualification and Honors Knowledge Unit, Official Statement Knowledge Unit, Contact Information Knowledge Unit, FAQ Knowledge Unit, Data Indicator Knowledge Unit, and Version Update Knowledge Unit.

[0018] The knowledge unit includes knowledge items with different trust levels; and for each type of knowledge unit, a knowledge type identifier and a trust level identifier are recorded to determine the key trust knowledge set and to select downgraded trust knowledge units.

[0019] The beneficial effects of adopting the above-mentioned further solutions are that by dividing brand information into multiple knowledge units and configuring credibility levels and type identifiers, key credible knowledge can be quickly filtered, citation needs can be flexibly matched, the implementation of downgrade strategies and the control of authority can be facilitated, and the output of brand information can be made more standardized, controllable and traceable.

[0020] Furthermore, when constructing the knowledge-dependent directed acyclic graph, the following steps are included:

[0021] Generate a knowledge unit identifier for each knowledge unit and record the mapping relationship between the knowledge unit identifier and the knowledge storage location;

[0022] Dependency edges are established according to the rule that the parent knowledge unit references the child knowledge unit, so that the product parameter knowledge unit depends on the qualification and honor knowledge unit it references, and the brand body knowledge unit depends on the product parameter knowledge unit and the official statement knowledge unit it references.

[0023] The knowledge-dependent directed acyclic graph is topologically sorted to obtain a topological sequence, and the in-degree information, direct predecessor set, and direct successor set of each knowledge unit are stored. During dispatch execution, only the tasks corresponding to knowledge units with an in-degree of zero and whose budget is satisfied are added to the ready queue.

[0024] The beneficial effects of adopting the above-mentioned further scheme are that by mapping knowledge units, constructing dependency edges and sorting topology, the knowledge dependency relationship and execution order are clarified, only tasks with zero in-degree and sufficient budget are scheduled, avoiding circular dependencies, improving execution efficiency, and ensuring the orderly and stable operation of the task chain.

[0025] Furthermore, the multi-stage processing task chain includes:

[0026] Verification tasks are used to verify the integrity of the acquired knowledge data.

[0027] Standardization tasks are used to parse knowledge data into structured data that can be used by large models;

[0028] Format conversion tasks are used to convert text or vector data into target formats supported by the search engine.

[0029] The source tracing preheating task is used to generate or load authoritative source information and verification cache items corresponding to the knowledge unit before the verification task.

[0030] Record the stage type identifier and the location of intermediate output for each stage of the task, so that the output of the parsing task can be directly used as the input for the verification task.

[0031] The beneficial effects of adopting the above-mentioned further solutions are that the multi-stage task chain covers the entire process of verification, standardization, format conversion, and source tracing and warm-up, and the inputs and outputs of each stage can be connected, improving the degree of knowledge structuring and verification efficiency, and providing high-quality, directly referential knowledge data for large models.

[0032] Furthermore, the semantic index is obtained by semantic segmentation of brand information, and the semantic segmentation adopts keyword clustering, semantic vector clustering, or topic segmentation;

[0033] A mapping table is established between semantic unit identifiers and knowledge unit lists for each semantic unit, wherein the knowledge unit lists include at least a basic trusted knowledge unit list and a highly trusted knowledge unit list;

[0034] The basic trusted knowledge unit is used to meet the minimum trusted presentation, and the high-trust knowledge unit is used to gradually replace the basic trusted knowledge unit when the budget allows.

[0035] The beneficial effects of adopting the above-mentioned further scheme are that, based on semantic clustering, a semantic index is constructed to distinguish between basic and high-trust knowledge units, which can quickly match query intent, prioritize the presentation of the lowest trust knowledge, and automatically upgrade high-trust knowledge when the budget is sufficient, thus balancing real-time performance and authority.

[0036] Furthermore, determining the predicted retrieval range includes:

[0037] Set the prediction time range and the prediction number of searches, and extrapolate the search status based on the current query intent, search history and user interaction behavior to obtain the predicted search status sequence;

[0038] The retrieval range is calculated based on the predicted retrieval status sequence to obtain the predicted retrieval range;

[0039] When the amount of change in user interaction input is detected to exceed a preset threshold, the prediction time range is shortened or the number of prediction retrievals is reduced to reduce invalid prefetching caused by prediction errors.

[0040] The beneficial effects of adopting the above-mentioned further solutions are that the retrieval range is dynamically predicted based on query and interaction behavior, the prediction window is automatically shortened when the user input changes abruptly, invalid prefetching is reduced, resource waste is reduced, and retrieval accuracy and system response efficiency are improved.

[0041] Furthermore, determining the set of key credible knowledge includes:

[0042] The citation contribution is calculated for knowledge units in the candidate knowledge subgraph, and the citation contribution is determined based on at least two of the following: semantic matching degree, relevance to query, authority level, and object importance weight;

[0043] Knowledge units whose citation contribution meets a preset threshold are added to the target set, and the target set is supplemented with knowledge units that directly or indirectly depend on them;

[0044] The key credible knowledge set is further filtered from the target set to obtain the key credible knowledge set, so that the key credible knowledge set at least covers the basic brand knowledge unit or basic credible entry knowledge unit required for the first paragraph to be cited and presented, and covers the citation format knowledge unit corresponding to the output content.

[0045] The beneficial effects of adopting the above-mentioned further solutions are that key credible knowledge is selected according to the contribution of citations and the dependency chain is completed, ensuring that the knowledge required for the first output and citation format is covered, and prioritizing the core content to be citationable and presentable under limited resources.

[0046] Furthermore, determining the critical path information and generating the task deadline includes:

[0047] Based on the topological sequence and the historical statistical or estimated time consumption of the knowledge unit, calculate the remaining processing time required to complete the retrieval task, parsing task and verification task for each knowledge unit.

[0048] In the key trusted knowledge set, the dependency chain is determined based on the reference dependency relationship, and the information corresponding to the dependency chain with the longest remaining processing time is taken as the critical path information.

[0049] Based on at least one of the following: the predicted time when the knowledge unit within the predicted retrieval range meets the citation contribution threshold, the first citation time limit, the allowable verification quantity budget per frame, and the allowable parsing time budget per frame, a task deadline is generated for the knowledge unit in the key trusted knowledge set, and a preset time margin is added to the task deadline.

[0050] The beneficial effects of adopting the above-mentioned further solutions are that, based on the time consumption statistics and dependency chain, the critical path is determined, and a task deadline with margin is generated in combination with the output time limit, so as to accurately control the processing rhythm, avoid overdue, and ensure smooth and stable output of large models.

[0051] Furthermore, determining the dispatch priority and performing cross-stage sorting and dispatch execution includes:

[0052] The dispatch priority is determined by comparing the task deadline with the remaining processing time, so that tasks with earlier deadlines and longer remaining processing time have higher dispatch priority.

[0053] Establish a retrieval ready queue, a parsing ready queue, and a verification ready queue, and set corresponding bandwidth budgets, parsing budgets, and verification budgets for each queue.

[0054] Within each scheduling cycle, tasks with higher dispatch priority are dispatched first. When multiple tasks have the same dispatch priority, tasks located on the dependency chain corresponding to the critical path information are dispatched first.

[0055] When the waiting time of the verification ready queue exceeds the congestion threshold, reduce the parsing budget or bandwidth budget to create back pressure; limit the verification amount per frame to within the verification budget so that the output time jitter of the generation thread does not exceed the preset jitter threshold.

[0056] The beneficial effects of adopting the above-mentioned further solutions are that dispatch priorities are set according to deadlines and critical paths, queue budget management is implemented and back pressure adjustment is supported to prevent queue congestion, output jitter is controlled within the threshold, and system stability and real-time performance are improved.

[0057] Furthermore, steps seven through nine also include:

[0058] An alternative mapping table is established for each knowledge unit. The alternative mapping table contains at least the original knowledge unit identifier, the downgraded trusted knowledge unit identifier, and the trust level information. The downgraded trusted knowledge unit includes at least one of low-precision entry knowledge, simplified parameter knowledge, and basic declaration knowledge.

[0059] When it is predicted that the original knowledge unit cannot complete the verification task within the task deadline, the downgraded trusted knowledge unit is loaded and enabled for output, while the original knowledge unit is kept as a high-trusted knowledge unit to be replaced, so as to trigger a gradual replacement when the budget allows.

[0060] The atomic replacement method includes versioned handle switching or pointer swapping, and performs a consistency check of the knowledge trust registration table before replacement. The consistency check includes at least confirming that the knowledge unit identifier to be switched is in a verified state in the knowledge trust registration table, and confirming that the knowledge unit identifier pointed to by the output reference is consistent with the verified knowledge unit identifier in the knowledge trust registration table.

[0061] The operational metrics are collected, which include at least two or more of the following: cache hit rate, number of task expirations, verification queue waiting time, and output time jitter. Based on the operational metrics, at least one of the following is adjusted: prediction time range, prediction retrieval quantity, retrieval task concurrency depth, parsing task thread concurrency, and verification budget per frame.

[0062] The beneficial effects of adopting the above-mentioned further solutions are that by seamlessly switching between downgraded knowledge units as a fallback and atomic replacement, the output is guaranteed to be uninterrupted and the credibility is not reduced. Combined with adaptive parameter tuning of operating indicators, the system performance, citation quality and resource utilization are continuously optimized.

[0063] Compared with existing technologies, the beneficial effects of this invention are as follows: This brand information referability optimization method based on large model retrieval path achieves structured and relational management of brand information through knowledge unit division and knowledge dependency directed acyclic graph construction. Combined with multi-stage processing task chain, it completes standardized processing of the entire process of knowledge retrieval, parsing, and verification. Based on user intent dynamic retrieval, critical path scheduling, fallback and atomic replacement mechanisms, it improves response speed and reduces output jitter while ensuring that the output of the large model is referable, traceable, and highly reliable. It achieves optimal allocation of resources for retrieval, parsing, and verification, and is suitable for real-time generation and compliant referencing scenarios of brand information. Attached Figure Description

[0064] Figure 1 This is a flowchart of a method for optimizing the citationability of brand information based on a large model retrieval path, as disclosed in an embodiment of the present invention. Detailed Implementation

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] Example 1

[0067] Please see Figure 1 This invention provides a technical solution: a method for optimizing the citationability of brand information based on a large model retrieval path, including step one: obtaining the information set of the target brand and dividing the information set into multiple knowledge units; establishing a knowledge dependency directed acyclic graph, where the nodes of the knowledge dependency directed acyclic graph are knowledge units, and the edges represent retrieval dependency, source dependency and citation dependency relationships;

[0068] Step 2: Establish a multi-stage processing task chain for each knowledge unit. The multi-stage processing task chain includes at least a retrieval task, a parsing task, and a verification task. The retrieval task is used to read brand knowledge data from the network or storage medium. The parsing task is used to unpack, standardize, structure, or semantically embed the knowledge data. The verification task is used to complete the authoritative traceability verification of the parsed data and write it into the knowledge trust registration form.

[0069] Step 3: Obtain the current user's query intent and interaction input to determine the predicted retrieval scope; based on the predicted retrieval scope, determine candidate retrieval units from the semantic index, and extract candidate knowledge subgraphs from the knowledge dependency directed acyclic graph accordingly;

[0070] Step 4: Determine the key credible knowledge set in the candidate knowledge subgraph. The key credible knowledge set is used to satisfy the requirement that the first part of the large model output can be cited or to satisfy the minimum credible presentation of brand information within the predicted retrieval range.

[0071] Step 5: Determine critical path information based on the key trusted knowledge set, and generate task deadlines for the multi-stage processing task chain based on the critical path information. The critical path information is the information corresponding to the dependency chain with the largest remaining processing time in the key trusted knowledge set.

[0072] Step 6: Based on the task deadline and critical path information, determine the dispatch priority according to the task deadline and remaining processing time, and under the constraints of retrieval bandwidth budget, parsing budget and verification budget, sort and dispatch the retrieval task, parsing task and verification task across stages.

[0073] Step 7: When the multi-stage processing task chain cannot be completed within the task deadline, select a downgraded trusted knowledge unit for the corresponding knowledge unit and load the downgraded trusted knowledge unit first to ensure minimum trusted presentation.

[0074] Step 8: After a knowledge unit or a downgraded trusted knowledge unit completes the verification task, the reference output is updated by atomic replacement, so that the large model generation pipeline can switch to the verified knowledge unit without interrupting the output loop.

[0075] Step 9: Collect operational metrics during the processing and adaptively adjust at least one of the following based on the operational metrics: retrieval task concurrency depth, parsing task thread concurrency, verification budget per frame, prediction time range, and prediction retrieval quantity.

[0076] In one embodiment of the present invention, the knowledge unit further includes:

[0077] The knowledge units include: Brand Body Knowledge Unit, Product Parameter Knowledge Unit, Qualification and Honors Knowledge Unit, Official Statement Knowledge Unit, Contact Information Knowledge Unit, FAQ Knowledge Unit, Data Indicator Knowledge Unit, and Version Update Knowledge Unit.

[0078] The knowledge unit includes knowledge items with different trust levels; and for each type of knowledge unit, a knowledge type identifier and a trust level identifier are recorded to determine the key trust knowledge set and select downgraded trust knowledge units. By dividing brand information into multiple types of knowledge units and configuring trust levels and type identifiers, key trust knowledge can be quickly filtered, citation needs can be flexibly matched, downgrade strategies can be implemented and authority can be controlled, making the output of brand information more standardized, controllable and traceable.

[0079] In one embodiment of the present invention, further, when establishing the knowledge-dependent directed acyclic graph, the following steps are included:

[0080] Generate a knowledge unit identifier for each knowledge unit and record the mapping relationship between the knowledge unit identifier and the knowledge storage location;

[0081] Dependency edges are established according to the rule that the parent knowledge unit references the child knowledge unit, so that the product parameter knowledge unit depends on the qualification and honor knowledge unit it references, and the brand body knowledge unit depends on the product parameter knowledge unit and the official statement knowledge unit it references.

[0082] The directed acyclic graph of knowledge dependencies is topologically sorted to obtain a topological sequence. The in-degree information, direct predecessor set, and direct successor set of each knowledge unit are stored. During dispatch execution, only the tasks corresponding to knowledge units with an in-degree of zero and a satisfied budget are added to the ready queue. Through knowledge unit mapping, dependency edge construction, and topological sorting, the knowledge dependency relationship and execution order are clarified. Only tasks with an in-degree of zero and a satisfied budget are scheduled, avoiding circular dependencies, improving execution efficiency, and ensuring the orderly and stable operation of the task chain.

[0083] According to one embodiment of the present invention, the multi-stage processing task chain further includes:

[0084] Verification tasks are used to verify the integrity of the acquired knowledge data.

[0085] Standardization tasks are used to parse knowledge data into structured data that can be used by large models;

[0086] Format conversion tasks are used to convert text or vector data into target formats supported by the search engine.

[0087] The source tracing preheating task is used to generate or load authoritative source information and verification cache items corresponding to the knowledge unit before the verification task.

[0088] Each stage of the task is recorded with a stage type identifier and the location of intermediate outputs, so that the output of the parsing task can be directly used as the input of the verification task. The multi-stage task chain covers the entire process of verification, standardization, format conversion, and source tracing. The inputs and outputs of each stage can be connected, improving the degree of knowledge structuring and verification efficiency, and providing high-quality, directly referential knowledge data for large models.

[0089] In one embodiment of the present invention, the semantic index is further obtained by semantic segmentation of brand information, wherein the semantic segmentation adopts keyword clustering, semantic vector clustering or topic segmentation;

[0090] A mapping table is established between semantic unit identifiers and knowledge unit lists for each semantic unit, wherein the knowledge unit lists include at least a basic trusted knowledge unit list and a highly trusted knowledge unit list;

[0091] The basic trusted knowledge unit is used to meet the minimum trusted presentation, and the high trusted knowledge unit is used to gradually replace the basic trusted knowledge unit when the budget allows. A semantic index is built based on semantic clustering to distinguish between basic trusted and high trusted knowledge units, which can quickly match query intent, prioritize the minimum trusted presentation, and automatically upgrade high trusted knowledge when the budget is sufficient, thus balancing real-time performance and authority.

[0092] In one embodiment of the present invention, determining the predicted retrieval range further includes:

[0093] Set the prediction time range and the prediction number of searches, and extrapolate the search status based on the current query intent, search history and user interaction behavior to obtain the predicted search status sequence;

[0094] The retrieval range is calculated based on the predicted retrieval status sequence to obtain the predicted retrieval range;

[0095] When the amount of change in user interaction input is detected to exceed a preset threshold, the prediction time range is shortened or the number of prediction searches is reduced to reduce invalid prefetching caused by prediction errors. The search range is dynamically predicted based on query and interaction behavior. When user input changes abruptly, the prediction window is automatically shortened to reduce invalid prefetching, reduce resource waste, and improve search accuracy and system response efficiency.

[0096] In one embodiment of the present invention, determining the key trusted knowledge set further includes:

[0097] The citation contribution is calculated for knowledge units in the candidate knowledge subgraph, and the citation contribution is determined based on at least two of the following: semantic matching degree, relevance to query, authority level, and object importance weight;

[0098] Knowledge units whose citation contribution meets a preset threshold are added to the target set, and the target set is supplemented with knowledge units that directly or indirectly depend on them;

[0099] The key credible knowledge set is further filtered from the target set to ensure that it covers at least the basic brand knowledge units or basic credible entry knowledge units required for the first paragraph to be presented, and also covers the citation format knowledge units corresponding to the output content. Key credible knowledge is filtered according to citation contribution and the dependency chain is supplemented to ensure that the knowledge required for the first paragraph output and citation format is covered, and to prioritize the core content to be citationable and presentable under limited resources.

[0100] In one embodiment of the present invention, determining the critical path information and generating the task deadline further includes:

[0101] Based on the topological sequence and the historical statistical or estimated time consumption of the knowledge unit, calculate the remaining processing time required to complete the retrieval task, parsing task and verification task for each knowledge unit.

[0102] In the key trusted knowledge set, the dependency chain is determined based on the reference dependency relationship, and the information corresponding to the dependency chain with the longest remaining processing time is taken as the critical path information.

[0103] Based on at least one of the following: the predicted time when knowledge units within the retrieval range meet the citation contribution threshold, the first segment's citation time limit, the allowable verification quantity budget per frame, and the allowable parsing time budget per frame, a task deadline is generated for the knowledge units in the key trusted knowledge set. A preset time margin is added to the task deadline. The critical path is determined based on time consumption statistics and dependency chains. Combined with the output time limit, a task deadline with margin is generated to accurately control the processing rhythm, avoid exceeding the deadline, and ensure smooth and stable output of the large model.

[0104] In one embodiment of the present invention, further, when determining the dispatch priority and performing cross-stage sorting and dispatch execution, the following steps are included:

[0105] The dispatch priority is determined by comparing the task deadline with the remaining processing time, so that tasks with earlier deadlines and longer remaining processing time have higher dispatch priority.

[0106] Establish a retrieval ready queue, a parsing ready queue, and a verification ready queue, and set corresponding bandwidth budgets, parsing budgets, and verification budgets for each queue.

[0107] Within each scheduling cycle, tasks with higher dispatch priority are dispatched first. When multiple tasks have the same dispatch priority, tasks located on the dependency chain corresponding to the critical path information are dispatched first.

[0108] When the waiting time of the verification ready queue exceeds the congestion threshold, reduce the parsing budget or bandwidth budget to create back pressure; limit the verification amount per frame to within the verification budget, so that the output time jitter of the generation thread does not exceed the preset jitter threshold; set the dispatch priority according to the deadline and critical path; manage the queue budget separately and support back pressure adjustment to prevent queue congestion; control the output jitter within the threshold and improve system stability and real-time performance.

[0109] In one embodiment of the present invention, further, steps seven through nine also include:

[0110] An alternative mapping table is established for each knowledge unit. The alternative mapping table contains at least the original knowledge unit identifier, the downgraded trusted knowledge unit identifier, and the trust level information. The downgraded trusted knowledge unit includes at least one of low-precision entry knowledge, simplified parameter knowledge, and basic declaration knowledge.

[0111] When it is predicted that the original knowledge unit cannot complete the verification task within the task deadline, the downgraded trusted knowledge unit is loaded and enabled for output, while the original knowledge unit is kept as a high-trusted knowledge unit to be replaced, so as to trigger a gradual replacement when the budget allows.

[0112] The atomic replacement method includes versioned handle switching or pointer swapping, and performs a consistency check of the knowledge trust registration table before replacement. The consistency check includes at least confirming that the knowledge unit identifier to be switched is in a verified state in the knowledge trust registration table, and confirming that the knowledge unit identifier pointed to by the output reference is consistent with the verified knowledge unit identifier in the knowledge trust registration table.

[0113] The system collects operational metrics, which include at least two or more of the following: cache hit rate, number of task expirations, verification queue waiting time, and output time jitter. Based on these operational metrics, it adjusts at least one of the following: prediction time range, prediction retrieval quantity, retrieval task concurrency depth, parsing task thread concurrency, and verification budget per frame. Through fallback knowledge units and seamless switching with atomic replacement, it ensures uninterrupted output and no reduction in reliability. Combined with adaptive parameter tuning based on operational metrics, it continuously optimizes system performance, citation quality, and resource utilization.

[0114] Example 2

[0115] This embodiment is applied to a corporate brand's official intelligent customer service model Q&A system.

[0116] Specific implementation steps

[0117] 1. Knowledge Unit Construction and Dependency Graph Establishment

[0118] We collect data from brand websites, official after-sales manuals, qualification certificates, official FAQs, and product parameter tables, and divide them into six knowledge units: brand body, product parameters, qualifications and honors, official statements, FAQs, and contact information; each unit is assigned a unique ID and trust level.

[0119] Dependency edges are established according to "Brand Main Body → Product Parameters → Qualifications and Honors → FAQs". A knowledge dependency directed acyclic graph is constructed and topological sorting is completed. Only units with an in-degree of zero and that meet the bandwidth / computing power budget are entered into the task queue.

[0120] 2. Multi-stage task chain execution

[0121] Perform the following steps sequentially for each knowledge unit:

[0122] (1) Retrieval task: Retrieve data from the official knowledge base and local storage;

[0123] (2) Verification task: Check data integrity and timeliness, and filter out expired claims;

[0124] (3) Standardization task: Convert to a unified JSON structured format;

[0125] (4) Format conversion task: Generate semantic vectors and store them in the retrieval engine;

[0126] (5) Source tracing and preheating task: preload official source links and verify cache;

[0127] (6) Verification task: Complete authoritative verification and write it into the knowledge trust registration form.

[0128] 3. Query intent analysis and prediction of search scope

[0129] The system acquires user input and identifies it as an after-sales FAQ query based on intent recognition. It extrapolates and predicts the search range based on the user's historical interactions and current input. When the user input changes abruptly, the prediction window is automatically shortened to reduce invalid searches.

[0130] 4. Screening of key and credible knowledge

[0131] In the candidate knowledge subgraph, the citation contribution is calculated based on semantic matching degree, authority level, and importance weight. The official quality assurance FAQ, corresponding product parameters, and quality assurance qualifications are selected to form a key credible knowledge set. All dependent units are then added to ensure that the knowledge can be directly cited and presented.

[0132] 5. Critical path and deadline generation

[0133] Based on historical time consumption statistics, the total time spent on retrieval, parsing, and verification of each unit is calculated, and the dependency chain with the longest remaining processing time is identified as the critical path. The task deadline is generated by combining the customer service's first reply time limit, and a time margin is added.

[0134] 6. Cross-stage priority scheduling

[0135] Establish three queues: retrieval, parsing, and verification, and configure bandwidth budget, number of parsing threads, and verification budget per frame for each queue. Dispatch tasks according to the principle of "earlier deadline, higher priority; critical path priority". When the verification queue is congested, backpressure is activated to reduce concurrency and ensure stable output.

[0136] 7. Degrade Trusted Element Enabled

[0137] When computing power is insufficient during peak periods and predictions cannot be completed on schedule, the simplified version of the official FAQ and basic quality assurance summary are automatically loaded as downgraded trusted units, prioritizing official and traceable outputs, while the full version of high-trust units continues to be loaded in the background.

[0138] 8. Atomic substitution output

[0139] After the downgrade / full version unit verification is completed, atomic replacement is performed by switching the versioned handle to update the output references of the large model; before replacement, a consistency check of the trusted registry is performed to ensure that only verified and officially valid knowledge is used.

[0140] 9. Adaptive Optimization

[0141] Real-time data collection of cache hit rate, task expiration count, output latency, and response accuracy; dynamic adjustment of retrieval concurrency, parsing thread count, verification budget per frame, and prediction range to continuously improve customer service response speed and citation accuracy.

[0142] Example 3

[0143] This embodiment is applied to a brand public information big data retrieval platform to provide authoritative brand information queries and prevent the use of false information.

[0144] Specific implementation steps

[0145] 1. Knowledge Unit Construction and Dependency Graph Establishment

[0146] The system collects official brand releases, authoritative media reports, government-published qualifications, annual reports, and official clarification statements to construct knowledge units for brand entities, data metrics, qualifications and honors, official statements, clarification information, and version updates. It marks authoritative sources and release dates, filtering out non-authoritative content from self-media. Dependencies are established according to "Brand Entity → Data Metrics → Qualifications and Honors → Official Statements → Clarification Information," generating a directed acyclic graph and performing topological sorting. Task scheduling only allows units with zero in-degree and meeting budget requirements to execute.

[0147] 2. Multi-stage processing task chain execution:

[0148] Search task: Read from authoritative information source databases and trusted network nodes;

[0149] Integrity verification: Filtering out incomplete or tampered data;

[0150] Standardization: unifying fields, definitions, and expression formats;

[0151] Vector transformation: adapting to semantic retrieval;

[0152] Source tracing and preheating: Preloading authoritative source links and public disclosure addresses;

[0153] Authoritative verification: Verify the source's qualifications and the issuing entity, and record the information in the trust registration form.

[0154] 3. Determining the predictive search scope

[0155] Obtain publicly available user queries (such as "What official qualifications has the brand obtained?" or "Has there been any recent statement?"), and extrapolate the search scope by combining search trends and user behavior; reduce the number of predictions when user queries change abruptly to reduce invalid prefetching and resource consumption.

[0156] 4. Screening of key credible knowledge sets

[0157] Calculate the citation contribution of each unit (matching degree + authority level + importance), and select government-published qualifications, official first release statements, and authoritative data indicators as key knowledge; automatically complete dependent units to ensure that the output is verifiable, traceable, and without broken links.

[0158] 5. Setting the critical path and task deadlines

[0159] Calculate the remaining processing time based on historical processing time, and designate the longest dependency chain, qualification verification + official statement, as the critical path; generate a deadline based on the platform's first screen display time limit, and reserve margin to avoid exceeding the deadline.

[0160] 6. Cross-phase scheduling and budget control

[0161] The system uses a queue-based scheduling mechanism for retrieval, parsing, and verification tasks, with constraints set for bandwidth, computing power, and verification. Priorities are sorted by deadline urgency and critical path. When verification is congested, parsing concurrency is automatically reduced to prevent output jitter and stuttering.

[0162] 7. Execution of the downgrade strategy

[0163] When there are network fluctuations, limited computing power, or high concurrency, the basic qualification summary and simplified version of the core declaration are used to downgrade the trusted unit to ensure the minimum trusted presentation; the full version of the high-trust unit is loaded asynchronously in the background and automatically upgraded when the conditions are met.

[0164] 8. Atomic Substitution and Consistency Verification

[0165] The pointer swapping atomic substitution is used to replace the output reference. Before the substitution, the status of the trusted registry is checked to ensure that only verified, unexpired, and officially valid knowledge is used; and to avoid the output of unsourced, non-authoritative, or erroneous information.

[0166] 9. Operational Indicator Collection and Adaptive Optimization

[0167] Continuously monitor cache hit rate, expiration count, verification queue waiting time, output accuracy, and false information interception rate; dynamically adjust prediction range, retrieval concurrency, number of threads, and verification budget per frame accordingly to achieve long-term stable operation and highly reliable output.

Claims

1. A method for optimizing the citationability of brand information based on large model retrieval paths, characterized in that, The steps include: Step 1: Obtaining the information set of the target brand and dividing the information set into multiple knowledge units; establishing a knowledge dependency directed acyclic graph, where the nodes of the knowledge dependency directed acyclic graph are knowledge units, and the edges represent retrieval dependency, source dependency, and reference dependency relationships; Step 2: Establish a multi-stage processing task chain for each knowledge unit. The multi-stage processing task chain includes at least a retrieval task, a parsing task, and a verification task. The retrieval task is used to read brand knowledge data from the network or storage medium. The parsing task is used to unpack, standardize, structure, or semantically embed the knowledge data. The verification task is used to complete the authoritative traceability verification of the parsed data and write it into the knowledge trust registration form. Step 3: Obtain the current user's query intent and interaction input to determine the predicted retrieval scope; based on the predicted retrieval scope, determine candidate retrieval units from the semantic index, and extract candidate knowledge subgraphs from the knowledge dependency directed acyclic graph accordingly; Step 4: Determine the key credible knowledge set in the candidate knowledge subgraph. The key credible knowledge set is used to satisfy the requirement that the first part of the large model output can be cited or to satisfy the minimum credible presentation of brand information within the predicted retrieval range. Step 5: Determine critical path information based on the key trusted knowledge set, and generate task deadlines for the multi-stage processing task chain based on the critical path information. The critical path information is the information corresponding to the dependency chain with the largest remaining processing time in the key trusted knowledge set. Step 6: Based on the task deadline and critical path information, determine the dispatch priority according to the task deadline and remaining processing time, and under the constraints of retrieval bandwidth budget, parsing budget and verification budget, sort and dispatch the retrieval task, parsing task and verification task across stages. Step 7: When the multi-stage processing task chain cannot be completed within the task deadline, select a downgraded trusted knowledge unit for the corresponding knowledge unit and load the downgraded trusted knowledge unit first to ensure minimum trusted presentation. Step 8: After a knowledge unit or a downgraded trusted knowledge unit completes the verification task, the reference output is updated by atomic replacement, so that the large model generation pipeline can switch to the verified knowledge unit without interrupting the output loop. Step 9: Collect operational metrics during the processing and adaptively adjust at least one of the following based on the operational metrics: retrieval task concurrency depth, parsing task thread concurrency, verification budget per frame, prediction time range, and prediction retrieval quantity.

2. The method for optimizing brand information referrability based on large model retrieval paths according to claim 1, characterized in that, The knowledge unit includes: The knowledge units include: Brand Body Knowledge Unit, Product Parameter Knowledge Unit, Qualification and Honors Knowledge Unit, Official Statement Knowledge Unit, Contact Information Knowledge Unit, FAQ Knowledge Unit, Data Indicator Knowledge Unit, and Version Update Knowledge Unit. The knowledge unit includes knowledge items with different trust levels; and for each type of knowledge unit, a knowledge type identifier and a trust level identifier are recorded to determine the key trust knowledge set and to select downgraded trust knowledge units.

3. The method for optimizing brand information referrability based on large model retrieval paths according to claim 1, characterized in that, When constructing the aforementioned knowledge-dependent directed acyclic graph, the following steps are included: Generate a knowledge unit identifier for each knowledge unit and record the mapping relationship between the knowledge unit identifier and the knowledge storage location; Dependency edges are established according to the rule that the parent knowledge unit references the child knowledge unit, so that the product parameter knowledge unit depends on the qualification and honor knowledge unit it references, and the brand body knowledge unit depends on the product parameter knowledge unit and the official statement knowledge unit it references. The knowledge-dependent directed acyclic graph is topologically sorted to obtain a topological sequence, and the in-degree information, direct predecessor set, and direct successor set of each knowledge unit are stored. During dispatch execution, only the tasks corresponding to knowledge units with an in-degree of zero and whose budget is satisfied are added to the ready queue.

4. The method for optimizing brand information referrability based on large model retrieval paths according to claim 1, characterized in that, The multi-stage processing task chain includes: Verification tasks are used to verify the integrity of the acquired knowledge data. Standardization tasks are used to parse knowledge data into structured data that can be used by large models; Format conversion tasks are used to convert text or vector data into target formats supported by the search engine. The source tracing preheating task is used to generate or load authoritative source information and verification cache items corresponding to the knowledge unit before the verification task. Record the stage type identifier and the location of intermediate output for each stage of the task, so that the output of the parsing task can be directly used as the input for the verification task.

5. The method for optimizing brand information referrability based on large model retrieval paths according to claim 1, characterized in that, The semantic index is obtained by semantic segmentation of brand information, and the semantic segmentation adopts keyword clustering, semantic vector clustering or topic segmentation; A mapping table is established between semantic unit identifiers and knowledge unit lists for each semantic unit, wherein the knowledge unit lists include at least a basic trusted knowledge unit list and a highly trusted knowledge unit list; The basic trusted knowledge unit is used to meet the minimum trusted presentation, and the high-trust knowledge unit is used to gradually replace the basic trusted knowledge unit when the budget allows.

6. The method for optimizing brand information referrability based on large model retrieval paths according to claim 1, characterized in that, Determining the predicted retrieval range includes: Set the prediction time range and the prediction number of searches, and extrapolate the search status based on the current query intent, search history and user interaction behavior to obtain the predicted search status sequence; The retrieval range is calculated based on the predicted retrieval status sequence to obtain the predicted retrieval range; When the amount of change in user interaction input is detected to exceed a preset threshold, the prediction time range is shortened or the number of prediction retrievals is reduced to reduce invalid prefetching caused by prediction errors.

7. The method for optimizing brand information referrability based on large model retrieval paths according to claim 1, characterized in that, Determining the set of key credible knowledge includes: The citation contribution is calculated for knowledge units in the candidate knowledge subgraph, and the citation contribution is determined based on at least two of the following: semantic matching degree, relevance to query, authority level, and object importance weight; Knowledge units whose citation contribution meets a preset threshold are added to the target set, and the target set is supplemented with knowledge units that directly or indirectly depend on them; The key credible knowledge set is further filtered from the target set to obtain the key credible knowledge set, so that the key credible knowledge set at least covers the basic brand knowledge unit or basic credible entry knowledge unit required for the first paragraph to be cited and presented, and covers the citation format knowledge unit corresponding to the output content.

8. The method for optimizing brand information referrability based on large model retrieval paths according to claim 1, characterized in that, Determining the critical path information and generating the task deadline includes: Based on the topological sequence and the historical statistical or estimated time consumption of the knowledge unit, calculate the remaining processing time required to complete the retrieval task, parsing task and verification task for each knowledge unit. In the key trusted knowledge set, the dependency chain is determined based on the reference dependency relationship, and the information corresponding to the dependency chain with the longest remaining processing time is taken as the critical path information. Based on at least one of the following: the predicted time when the knowledge unit within the predicted retrieval range meets the citation contribution threshold, the first citation time limit, the allowable verification quantity budget per frame, and the allowable parsing time budget per frame, a task deadline is generated for the knowledge unit in the key trusted knowledge set, and a preset time margin is added to the task deadline.

9. The method for optimizing brand information referrability based on a large model retrieval path according to claim 1, characterized in that, Determining the dispatch priority and performing cross-stage sorting and dispatch execution includes: The dispatch priority is determined by comparing the task deadline with the remaining processing time, so that tasks with earlier deadlines and longer remaining processing time have higher dispatch priority. Establish a retrieval ready queue, a parsing ready queue, and a verification ready queue, and set corresponding bandwidth budgets, parsing budgets, and verification budgets for each queue. Within each scheduling cycle, tasks with higher dispatch priority are dispatched first. When multiple tasks have the same dispatch priority, tasks located on the dependency chain corresponding to the critical path information are dispatched first. When the waiting time of the verification ready queue exceeds the congestion threshold, reduce the parsing budget or bandwidth budget to create back pressure; limit the verification amount per frame to within the verification budget so that the output time jitter of the generation thread does not exceed the preset jitter threshold.

10. The method for optimizing brand information referrability based on a large model retrieval path according to claim 1, characterized in that, Steps seven through nine also include: An alternative mapping table is established for each knowledge unit. The alternative mapping table contains at least the original knowledge unit identifier, the downgraded trusted knowledge unit identifier, and the trust level information. The downgraded trusted knowledge unit includes at least one of low-precision entry knowledge, simplified parameter knowledge, and basic declaration knowledge. When it is predicted that the original knowledge unit cannot complete the verification task within the task deadline, the downgraded trusted knowledge unit is loaded and enabled for output, while the original knowledge unit is kept as a high-trusted knowledge unit to be replaced, so as to trigger a gradual replacement when the budget allows. The atomic replacement method includes versioned handle switching or pointer swapping, and performs a consistency check of the knowledge trust registration table before replacement. The consistency check includes at least confirming that the knowledge unit identifier to be switched is in a verified state in the knowledge trust registration table, and confirming that the knowledge unit identifier pointed to by the output reference is consistent with the verified knowledge unit identifier in the knowledge trust registration table. The operational metrics are collected, which include at least two or more of the following: cache hit rate, number of task expirations, verification queue waiting time, and output time jitter. Based on the operational metrics, at least one of the following is adjusted: prediction time range, prediction retrieval quantity, retrieval task concurrency depth, parsing task thread concurrency, and verification budget per frame.

Images