A method for dynamic updating and precise delivery of seedling knowledge base based on adaptive RAG
By constructing a multi-level vector index library and dynamic user profiles, and combining the context of seedling cultivation scenarios with feedback verification, the problems of single criteria for knowledge fragment selection and untimely updates in seedling cultivation knowledge services have been solved, thus achieving accurate delivery and improved adaptability of seedling cultivation knowledge.
Patent Information
- Application Number
- CN202611308993.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-27
- Publication Date
- 2026-11-17
AI Technical Summary
In existing seedling knowledge services, the criteria for selecting knowledge fragments are singular, there is a lack of suppression of conflicting knowledge in different scenarios, user feedback is difficult to reflect the actual application effect, and knowledge updates are difficult to link with vector index refresh, resulting in knowledge retrieval and push not adapting to the continuously changing conditions of the seedling stage.
The method for dynamic updating and precise push of seedling knowledge base based on adaptive RAG constructs a multi-level vector index library. It combines dynamic user profiles and scene context, and integrates gating fusion similarity, scene relevance, knowledge freshness, source authority, feedback credibility, and conflict penalty factor to reorder knowledge fragments. It collects user feedback and actual application effects to update knowledge, refresh knowledge weights and indexes.
It enables reliable retrieval based on dynamic user profiles and the current seedling cultivation scenario, and feedback-driven knowledge updates improve the adaptability and reliability of seedling cultivation knowledge services, reduce the probability of outdated or inapplicable knowledge entering high-priority results, and optimize the knowledge base maintenance process.
Smart Images

Figure CN123115138A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of smart agricultural information processing technology, and in particular to a method for dynamic updating and precise delivery of seedling knowledge base based on adaptive RAG. Background Technology
[0002] Seedling raising is a crucial step in agricultural production. Environmental control, seedling management, pest and disease control, and water and fertilizer management during the seedling stage directly impact seedling survival and subsequent production. With increasing agricultural informatization, seedling technology data has expanded from traditional manuals and expert experience to various formats including agricultural research papers, technical guidance documents, instructional videos, seedling images, and user-generated seedling records. Data from different sources differ in content structure, presentation, and update frequency. Furthermore, seedling knowledge is strongly correlated with crop type, seedling stage, geographical conditions, environmental parameters, and pest and disease status. The same management measures may have different applicability conditions on different crops or at different seedling stages. Therefore, relying solely on fixed knowledge bases or keyword searches is insufficient to continuously provide knowledge content relevant to the current seedling raising context.
[0003] Retrieval-enhanced generation technology retrieves external knowledge before the language model generates answers, and can constrain the generation process using knowledge base content. It has already been applied to agricultural knowledge question answering and knowledge recommendation. Among existing publicly available solutions, CN121614587A relates to agricultural intelligent question answering based on knowledge graphs and enhanced generation; CN119226454A relates to retrieval-enhanced generation and large-scale agricultural knowledge question answering; CN117370506A relates to multimodal, multi-turn dialogue agricultural intelligent question answering; and CN117540026B relates to agricultural knowledge recommendation combining user attributes and historical behavior. Furthermore, CN121390297A relates to spatiotemporal knowledge graphs and hierarchical prompt generation; CN119782476A relates to agricultural question answering processes and answer result evaluation; and CN122287922A relates to farmland multimodal data fusion and decision attribution. The aforementioned technologies have been able to utilize knowledge graphs, vector knowledge bases, multimodal data, language models, user attributes, or feedback evaluations to complete agricultural knowledge retrieval, question-and-answer generation, or knowledge recommendation, providing a corresponding technological foundation for agricultural knowledge services.
[0004] However, seedling knowledge is not only highly context-dependent but also changes as the seedling stage progresses, environmental conditions change, new knowledge is added, and practical application feedback emerges. Existing agricultural knowledge question-answering or recommendation methods, after obtaining candidate knowledge, typically rely mainly on semantic similarity, knowledge graph relationships, or user preferences for selection. This fails to simultaneously reflect the degree of matching between candidate knowledge and the user's long-term seedling characteristics, the degree of adaptability to the current seedling scenario, the knowledge's publication time and maintenance status, the reliability of the knowledge source, and the credibility of historical application feedback. When semantically similar knowledge corresponds to different crops, different seedling stages, different environmental parameter ranges, or different pest and disease states, relying solely on semantic relevance can easily lead to outdated or unsuitable knowledge entering the generation context or gaining higher recommendation priority.
[0005] On the other hand, existing knowledge base updates typically focus on adding knowledge content, adjusting weights, or periodically rebuilding indexes. User feedback is also mostly used to update user preferences or evaluate response results. There is a lack of processing methods that integrate knowledge credibility, scenario conflict, and index maintenance processes to address practical application results such as changes in survival rate, yield, pest and disease incidence, and environmental parameter compliance after users adopt recommended content. When a piece of knowledge consistently shows discrepancies in application effects across different users or seedling scenarios, if only the recommendation weight is reduced without simultaneously adjusting knowledge confidence, cross-modal associations, and the corresponding vector index, the knowledge may still repeatedly enter the candidate set during subsequent retrievals. Furthermore, rebuilding the entire vector index for each knowledge change increases the computational load during continuous knowledge base maintenance.
[0006] Therefore, existing technologies still need to address the problem that seedling knowledge retrieval, feedback verification, and knowledge base updates are relatively independent of each other. This would enable the selection of candidate knowledge to simultaneously consider user status, current seedling scenario, knowledge timeliness, knowledge source, and historical feedback. This would allow for the identification of knowledge that conflicts with the current scenario or has low credibility, and enable practical application feedback to further participate in the updating of knowledge status and related indexes. This would adapt to the needs of knowledge retrieval and proactive push under continuously changing conditions in the seedling stage. Summary of the Invention
[0007] The purpose of this invention is to provide a method for dynamic updating and precise push of seedling knowledge base based on adaptive RAG. This method can solve the problems of single knowledge fragment selection criteria, lack of suppression of conflicting knowledge in scenarios, difficulty in reflecting actual application effects in user feedback, and difficulty in linking knowledge updates with vector index refresh in existing seedling knowledge services. It can achieve reliable retrieval based on dynamic user profiles and current seedling scenarios, feedback-driven knowledge updates, and precise push, thereby improving the adaptability and reliability of seedling knowledge services.
[0008] To achieve the above objectives, the present invention provides the following solution:
[0009] A method for dynamic updating and precise delivery of a seedling knowledge base based on adaptive RAG, comprising:
[0010] A multi-level vector index library was constructed based on the seedling knowledge base;
[0011] Dynamic user profiles are constructed based on users' historical seedling data, query records, and behavioral preferences, and user profile vectors are formed.
[0012] A scenario context is formed based on the user's current seedling crop, seedling time, and environmental parameters; the scenario context includes the current seedling stage determined based on the current seedling crop and seedling time;
[0013] By integrating query vectors generated from user queries and user profile vectors, relevant knowledge fragments can be retrieved from a multi-level vector index library;
[0014] The gating fusion similarity between relevant knowledge fragments and dynamic user profiles, the scene relevance between relevant knowledge fragments and scene context, the knowledge freshness and source authority of relevant knowledge fragments, the feedback credibility corresponding to relevant knowledge fragments, and the conflict penalty factor of relevant knowledge fragments relative to scene context are determined to form a comprehensive recommendation score, and the relevant knowledge fragments are re-ranked according to the comprehensive recommendation score.
[0015] Based on the reordered relevant knowledge fragments, structured context information is constructed. User queries and structured context information are input into a domain-adaptive language model to generate answers.
[0016] Collect user feedback and the actual application effect after adopting the answer, update the feedback credibility corresponding to the relevant knowledge fragments used to construct structured context information, and determine the application effect deviation. Based on the updated feedback credibility, application effect deviation, conflict penalty factor, and the time difference between the release time of the corresponding knowledge and the current time, a knowledge update trigger value is formed.
[0017] When the knowledge update trigger value meets the preset update conditions, the knowledge weight and knowledge confidence of the relevant knowledge fragments used to construct structured context information are updated, the incremental vector is recalculated for the knowledge affected by the update, and the corresponding level index is refreshed.
[0018] Based on the updated seedling knowledge base, dynamic user profiles, and the current seedling stage, a personalized knowledge recommendation list is generated, and the recommended content in the personalized knowledge recommendation list is pushed to users.
[0019] Preferably, a multi-level vector index library is constructed based on the seedling knowledge base, including:
[0020] Acquire textual, video, and image knowledge from the seedling knowledge base;
[0021] Based on crop type, seedling stage, pest and disease type and environmental parameters, establish cross-modal association relationships between text knowledge, video knowledge and image knowledge to form a unified multi-source knowledge representation;
[0022] Convert text knowledge, video knowledge, and image knowledge into their respective vector representations;
[0023] Set up hierarchical index nodes according to domain, crop type, seedling stage, and knowledge type;
[0024] Based on the unified multi-source knowledge representation corresponding to the domain, crop type, seedling stage and knowledge type, the vector representation is written into the corresponding hierarchical index node to form a multi-level vector index library.
[0025] Preferably, the query vector and user profile vector formed by user queries are merged to retrieve relevant knowledge fragments in a multi-level vector index library, including:
[0026] Based on dynamic user profiles, scenario contexts, and the release time and update frequency of seedling knowledge, the retrieval scope in the multi-level vector index library is determined;
[0027] By integrating query vectors and user profile vectors, a personalized query representation can be formed;
[0028] Based on personalized query representation, vector retrieval is performed within the retrieval scope to obtain relevant knowledge fragments.
[0029] Preferably, the following factors are determined: the gated fusion similarity between relevant knowledge fragments and dynamic user profiles; the scene relevance between relevant knowledge fragments and scene context; the knowledge freshness and source authority of relevant knowledge fragments; the feedback credibility corresponding to relevant knowledge fragments; and the conflict penalty factor of relevant knowledge fragments relative to scene context. These factors are used to form a comprehensive recommendation score, and the relevant knowledge fragments are then re-ranked according to the comprehensive recommendation score, including:
[0030] For each relevant knowledge fragment, determine the gating fusion similarity, scene relevance, knowledge novelty, source authority, feedback credibility, and conflict penalty factor.
[0031] We weighted the similarity of gating fusion, scene relevance, knowledge novelty, source authority and feedback credibility respectively, and used the weighted conflict penalty factor as a deduction item to form a comprehensive evaluation result.
[0032] The comprehensive evaluation results are normalized to form a comprehensive recommendation score;
[0033] The relevant knowledge segments were reordered based on their overall recommendation scores.
[0034] Preferably, the determination of gated fusion similarity includes:
[0035] Obtain the normalized values of relevant knowledge fragments and dynamic user profiles across each profile feature dimension;
[0036] Based on the normalized values of relevant knowledge fragments and dynamic user profiles on the corresponding profile feature dimensions, as well as the weights of the corresponding profile feature dimensions, the weighted directional similarity is determined.
[0037] Based on the differences between relevant knowledge fragments and dynamic user profiles in the corresponding profile feature dimensions and the difference attenuation coefficient of the corresponding profile feature dimensions, determine the profile feature difference attenuation term;
[0038] The gating coefficient is determined based on the interaction between the normalized values of relevant knowledge fragments on each profile feature dimension and the normalized values of the dynamic user profile.
[0039] By adjusting the proportions of weighted directional similarity and portrait feature difference attenuation terms in similarity calculation using gating coefficients, a gated fusion similarity is formed.
[0040] Preferably, the determination of scenario relevance, knowledge novelty, and source authority includes:
[0041] Based on the differences between relevant knowledge fragments and scene context in terms of seedling time, region, seedling stage, environmental parameters, and pest and disease status, as well as the weight of each scene dimension, the scene relevance is determined.
[0042] The freshness of knowledge is determined by the time difference between the publication time of the relevant knowledge fragment and the current time, as well as the update frequency of the relevant knowledge fragment.
[0043] The authority of a source is determined based on the type of knowledge source, the qualifications of the publishing entity, the status of expert review, and the stability of historical feedback for the relevant knowledge fragments.
[0044] Preferably, the determination of feedback credibility and conflict penalty factor includes:
[0045] Obtain the satisfaction score, adoption mark, and actual application effect achievement value corresponding to the relevant knowledge fragments;
[0046] The satisfaction scores, adoption markers, and actual application effect achievement values are weighted and normalized to determine the credibility of the feedback.
[0047] Based on the degree of conflict between relevant knowledge fragments and the context of the scenario regarding crop type, seedling stage, environmental parameters, and pest and disease status, determine the conflict risk value;
[0048] When the conflict risk value exceeds the preset conflict threshold, the conflict penalty factor is determined according to the degree to which the conflict risk value exceeds the preset conflict threshold.
[0049] When the conflict risk value does not exceed the preset conflict threshold, the conflict penalty factor is set to zero.
[0050] Preferably, user feedback and the actual application effect after adopting the answer are collected; the feedback credibility corresponding to the relevant knowledge fragments used to construct structured context information is updated; and the application effect deviation is determined. A knowledge update trigger value is formed based on the updated feedback credibility, application effect deviation, conflict penalty factor, and the time difference between the publication time of the corresponding knowledge and the current time, including:
[0051] Collect satisfaction feedback, adoption feedback, and follow-up feedback to form user feedback;
[0052] Collect at least one of the following data after users adopt the answer: survival rate, yield, changes in pest and disease incidence, and compliance of environmental parameters, to form the actual application effect;
[0053] Based on user feedback and actual application results, update the feedback credibility of relevant knowledge fragments used to construct structured context information;
[0054] Determine the application effect deviation based on the difference between the expected application results and the actual application effects corresponding to the answers;
[0055] Determine the time difference between the publication time of the relevant knowledge fragments used to construct structured context information and the current time;
[0056] The updated feedback credibility, application effect deviation, conflict penalty factor, and time difference are compared with the corresponding preset trigger thresholds.
[0057] The knowledge update trigger value is generated based on the number of times the corresponding trigger condition is met.
[0058] Preferably, when the knowledge update trigger value meets the preset update conditions, the knowledge weights and knowledge confidence levels corresponding to the relevant knowledge fragments used to construct structured context information are updated, and incremental vector recalculation is performed on the knowledge affected by the update and the corresponding level index is refreshed, including:
[0059] Based on the updated feedback credibility and application effect deviation, adjust the knowledge weight and knowledge confidence of the relevant knowledge fragments;
[0060] Update the cross-modal association relationships between relevant knowledge fragments based on the knowledge verification results;
[0061] Knowledge whose knowledge weight, knowledge confidence, or cross-modal association changes are identified as knowledge affected by the update;
[0062] Identify adjacent knowledge that is related to the knowledge affected by the update;
[0063] Perform incremental vector recalculation on the knowledge affected by the update and adjacent knowledge;
[0064] Based on the vector representation recalculated from the incremental vector, refresh the hierarchical indexes corresponding to the domain, crop type, seedling stage, and knowledge type of the knowledge affected by the update and adjacent knowledge;
[0065] Keep the hierarchical indexes that are not affected by the update unchanged.
[0066] Preferably, a personalized knowledge recommendation list is generated based on the updated seedling knowledge base, dynamic user profiles, and the current seedling stage, and recommended content from the personalized knowledge recommendation list is pushed to the user, including:
[0067] Match relevant knowledge, technical points, and precautions corresponding to the current seedling stage from the updated seedling knowledge base;
[0068] Determine the gating fusion similarity, scene relevance, knowledge freshness, source authority, feedback credibility, and conflict penalty factor corresponding to the relevant knowledge obtained from the matching, and form a corresponding comprehensive recommendation score;
[0069] Based on dynamic user profiles and comprehensive recommendation scores, the relevant knowledge, technical points and precautions obtained from the matching are sorted in a personalized way to form a personalized knowledge recommendation list.
[0070] Push recommended content from the personalized knowledge recommendation list.
[0071] The present invention discloses the following beneficial effects:
[0072] (1) This invention does not merely apply retrieval enhancement generation technology to seedling knowledge services. Instead, it establishes a continuous data processing relationship between dynamic user profiles, scenario context, trusted reordering, feedback verification, and knowledge updates, taking into account the characteristic that the applicable conditions of knowledge in the seedling scenario continuously change with crop type, seedling stage, environmental parameters, and actual application results. The query vector generated by the user query is fused with the user profile vector and then enters a multi-level vector index library to complete knowledge retrieval. This ensures that the retrieval process is simultaneously constrained by the user's long-term seedling characteristics and current seedling needs, reducing retrieval bias caused by knowledge matching based solely on the semantics of a single query.
[0073] (2) This invention comprehensively determines gated fusion similarity, scene relevance, knowledge freshness, source authority, feedback credibility, and conflict penalty factor for relevant knowledge fragments retrieved, and forms a comprehensive recommendation score. Gated fusion similarity is used to evaluate the degree of matching between relevant knowledge fragments and dynamic user profiles; scene relevance is used to evaluate the degree of adaptation between relevant knowledge fragments and the current scene context; knowledge freshness is used to reflect the time status and maintenance status of knowledge; source authority is used to reflect the reliability of the knowledge source; feedback credibility is used to reflect the degree of support for knowledge from historical feedback and actual application results; and conflict penalty factor is used to suppress knowledge that is inconsistent with the current seedling cultivation scenario. The above evaluation factors jointly participate in the re-ranking of relevant knowledge fragments, which can reduce the probability of ranking outdated, low-credibility, or knowledge that does not match the current seedling cultivation stage or environmental conditions into high-priority results based solely on high semantic similarity.
[0074] (3) This invention uses gated fusion similarity to evaluate the matching relationship between relevant knowledge fragments and dynamic user profiles. Compared with using only a single vector similarity for ranking, it can simultaneously reflect the degree of directional consistency between knowledge features and user profile features, as well as the attenuation effect caused by differences in key features. Furthermore, it adjusts the influence of different evaluation parts according to the interaction relationship between knowledge features and user profile features. When there are significant differences between relevant knowledge fragments and users in terms of crop preferences, seedling stages, or focus, the matching degree of the corresponding knowledge fragments can be reduced, thereby improving the adaptability between relevant knowledge fragments and users' actual seedling needs.
[0075] (4) This invention constructs the reordered relevant knowledge fragments into structured context information, and inputs the user query and the structured context information into the domain-adaptive language model, so that the answer generation is constrained by seedling knowledge that has undergone credibility evaluation and scenario screening. Compared with the method of directly using the language model to generate answers or retrieving knowledge based solely on semantic similarity, the knowledge entering the answer generation stage has already undergone user matching, scenario matching, timeliness evaluation, source evaluation, feedback evaluation, and conflict suppression. This can reduce the interference of knowledge that is irrelevant or inapplicable to the current seedling problem on the answer generation process and improve the consistency between the answer and the current seedling scenario.
[0076] (5) After the answer is adopted by the user, this invention continues to collect user feedback and actual application effects, and updates the feedback credibility of the relevant knowledge fragments based on user feedback and actual application effects, while determining the application effect deviation between the expected application results and the actual application results. Thus, knowledge evaluation no longer stops at user satisfaction, clicks, or adoption, but can verify the applicability of knowledge by combining actual seedling results such as survival rate, yield, changes in pest and disease incidence, and compliance of environmental parameters. For knowledge with high user ratings but whose actual application results do not meet expectations, this invention avoids continuously increasing the recommendation priority based solely on subjective evaluations, thereby improving the reliability of feedback information in knowledge maintenance.
[0077] (6) This invention further forms a knowledge update trigger value based on the updated feedback credibility, application effect deviation, conflict penalty factor, and the time difference between the publication time of the corresponding knowledge and the current time, and uses the knowledge update trigger value to control the knowledge maintenance process. When the knowledge update trigger value meets the preset update conditions, the knowledge weight and knowledge confidence corresponding to the relevant knowledge fragment are updated, so that the feedback credibility, actual application deviation, current scenario conflict, and knowledge time status can jointly affect the usage status of knowledge in subsequent retrieval and recommendation. Compared with adjusting the recommendation weight only based on a single user feedback or updating the knowledge base uniformly according to a fixed period, this can reduce the long-term retention of low-credibility, large application effect deviation, scenario incompatibility, or reduced timeliness knowledge in the original recommendation state.
[0078] (7) After the preset update conditions are met, the present invention performs incremental vector recalculation on the knowledge affected by the update and refreshes the corresponding level index, so that the knowledge evaluation results can be further applied to the subsequent vector retrieval process. After the knowledge state changes, it is not necessary to perform indiscriminate vector recalculation and index reconstruction on all knowledge. Instead, the update is completed for the knowledge affected by the update and the corresponding level index. This can reduce unnecessary repetitive calculations in the continuous maintenance of the knowledge base. At the same time, the updated knowledge weights and knowledge confidence are reflected in the subsequent knowledge retrieval results in a timely manner, realizing the linkage between knowledge state update and retrieval base update.
[0079] (8) After updating the knowledge status and corresponding hierarchical index, this invention combines the updated seedling knowledge base, dynamic user profile, and current seedling stage to form a personalized knowledge recommendation list. The recommended content is no longer determined solely by the user's historical behavior or static knowledge matching results, but is based on the aforementioned knowledge retrieval, reliable reordering, answer application, actual effect verification, and knowledge update results, and is adjusted according to changes in the current seedling stage, thereby expanding the seedling knowledge service from a single passive question and answer to a continuous and accurate push for the seedling process.
[0080] In summary, this invention constructs a reliable re-ranking mechanism for relevant knowledge fragments by gating and fusing similarity, scene relevance, knowledge freshness, source authority, feedback credibility, and conflict penalty factors. It further refines the knowledge evaluation state based on user feedback and practical application effects. Furthermore, it links knowledge weights, knowledge confidence, incremental vector recalculation, and corresponding level index refresh through knowledge update trigger values. Finally, the updated knowledge state is used for personalized knowledge recommendations in the current seedling stage. This forms a closed-loop processing procedure from knowledge retrieval, credibility evaluation, answer generation, practical application verification, knowledge state update, index refresh to subsequent recommendations. This reduces the probability of outdated, low-credibility, or conflicting knowledge from the current seedling scenario continuously entering high-priority retrieval and recommendation results, improving the adaptability and reliability of dynamic updates and accurate delivery of seedling knowledge. Attached Figure Description
[0081] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0082] Figure 1 A flowchart of the method provided in an embodiment of the present invention;
[0083] Figure 2 A flowchart illustrating the closed-loop process of dynamic updating and precise push of the seedling knowledge base provided in this embodiment of the invention;
[0084] Figure 3 This is an architecture diagram of a seedling knowledge base dynamic update and precise push system based on adaptive RAG provided in an embodiment of the present invention.
[0085] Figure 4 A flowchart of multi-source knowledge processing provided for embodiments of the present invention;
[0086] Figure 5 A flowchart for generating enhanced search results provided in this embodiment of the invention;
[0087] Figure 6 The feedback learning flowchart is provided for embodiments of the present invention. Detailed Implementation
[0088] 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.
[0089] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0090] Figure 1 The method flowchart provided in the embodiments of the present invention is as follows: Figure 1 As shown, this embodiment provides a method for dynamic updating and precise push of a seedling knowledge base based on adaptive RAG, including:
[0091] Step 100: Construct a multi-level vector index library based on the seedling knowledge base;
[0092] Step 200: Construct a dynamic user profile based on the user's historical seedling data, query records, and behavioral preferences, and form a user profile vector;
[0093] Step 300: Form a scenario context based on the user's current seedling crop, seedling time, and environmental parameters; the scenario context includes the current seedling stage determined based on the current seedling crop and seedling time;
[0094] Step 400: Merge the query vector and user profile vector formed by user queries, and retrieve relevant knowledge fragments in the multi-level vector index library;
[0095] Step 500: Determine the gating fusion similarity between relevant knowledge fragments and dynamic user profiles, the scene relevance between relevant knowledge fragments and scene context, the knowledge freshness and source authority of relevant knowledge fragments, the feedback credibility corresponding to relevant knowledge fragments, and the conflict penalty factor of relevant knowledge fragments relative to scene context, form a comprehensive recommendation score, and re-rank relevant knowledge fragments according to the comprehensive recommendation score;
[0096] Step 600: Construct structured context information based on the reordered relevant knowledge fragments, input the user query and structured context information into the domain-adaptive language model, and generate the answer;
[0097] Step 700: Collect user feedback and the actual application effect after adopting the answer, update the feedback credibility corresponding to the relevant knowledge fragments used to construct structured context information, and determine the application effect deviation. Based on the updated feedback credibility, application effect deviation, conflict penalty factor, and the time difference between the release time of the corresponding knowledge and the current time, a knowledge update trigger value is formed.
[0098] Step 800: When the knowledge update trigger value meets the preset update conditions, update the knowledge weight and knowledge confidence corresponding to the relevant knowledge fragments used to construct structured context information, perform incremental vector recalculation on the knowledge affected by the update and refresh the corresponding level index;
[0099] Step 900: Based on the updated seedling knowledge base, dynamic user profiles, and the current seedling stage, a personalized knowledge recommendation list is generated, and the recommended content in the personalized knowledge recommendation list is pushed to the user.
[0100] This embodiment organizes seedling cultivation knowledge scattered across text, video, images, and expert experience into a continuously maintainable knowledge resource. After seedling cultivation knowledge enters the knowledge base, it forms a vector index. Historical seedling cultivation data and interaction data are used to create dynamic user profiles. User queries, user profiles, and the current seedling cultivation scenario jointly define the retrieval scope and participate in the reliable reordering of relevant knowledge fragments. User feedback and actual seedling cultivation results generated after an answer is adopted re-influence the knowledge state and index state. The updated knowledge continues to participate in proactive recommendations, thus ensuring a continuous connection between retrieval, generation, feedback, maintenance, and push notifications.
[0101] Figure 2 A flowchart illustrating the closed-loop process of dynamic updating and precise push of the seedling knowledge base provided in this embodiment of the invention. Figure 2 The diagram illustrates the functional steps of multimodal seedling knowledge acquisition and preprocessing, knowledge vectorization and index construction, user profile construction, retrieval enhancement generation, feedback learning and knowledge updating, proactive recommendation, and data flow between users. Step 1: Multimodal Seedling Knowledge Acquisition and Preprocessing and Step 2: Knowledge Vectorization and Index Construction together form the foundation for knowledge preparation and index construction in Step 100; Step 3: User Profile Construction corresponds to Step 200; Step 4: Retrieval Enhancement Generation covers scenario determination, personalized retrieval, reliable re-ranking, and answer generation from Steps 300 to 600; Step 5: Feedback Learning and Knowledge Updating corresponds to Steps 700 and 800; and Step 6: Proactive Recommendation corresponds to Step 900. After preprocessing, multimodal seedling knowledge enters the knowledge vectorization and index construction stage, forming a multi-level vector index library. This multi-level vector index library provides the retrieval foundation for the retrieval enhancement generation stage and receives knowledge base and index update information formed by the feedback learning and knowledge updating stages. The dynamic user profiles generated during the user profile building phase are then passed to the retrieval enhancement generation phase and the proactive recommendation phase. User queries are processed in the retrieval enhancement generation phase, where answers are obtained. Feedback from the query and answer usage then enters the feedback learning and knowledge update phase. This phase also returns updated profile information to the user profile building phase and changes in knowledge and index status to the knowledge vectorization and index building phase. The proactive recommendation phase generates recommended content based on the dynamic user profiles and the current development stage and pushes it to the user. Figure 2 Solid lines represent the main processing flow and push path, while dashed lines represent the feedback path for updating the knowledge base and index, updating user profiles, querying and feedback, and proactive recommendations based on profiles and stages.
[0102] The seedling knowledge base in step 100 uses multi-source seedling knowledge data as its knowledge source. This multi-source seedling knowledge data includes various types of data such as literature, videos, images, and expert experience. Text data sources include seedling technology manuals, agricultural journal articles, expert technical blogs, government agricultural guidance documents, seedling technology documents, papers, and expert experience; video data sources include agricultural technology promotion videos, expert lecture recordings, farmer practical operation videos, seedling teaching videos, expert lecture videos, and practical operation videos; image data sources include seedling monitoring images, pest and disease sample images, seedling environment images, seedling condition images, pest and disease images, and environmental images. The above raw data undergoes cleaning, labeling, and structuring to remove duplicate or invalid content, supplement knowledge source and seedling condition identification information, and ensure that different data types can enter the subsequent knowledge extraction process according to unified fields.
[0103] When processing text data, word segmentation tools are used to segment seedling technology manuals, agricultural journal articles, expert technical blogs, government agricultural guidance documents, seedling technology documents, papers, and expert experience. Then, a Named Entity Recognition (NER) model is used to identify entities such as crop names, pest and disease names, and environmental parameters. Finally, a relation extraction model is used to extract the relationships between entities, forming structured textual knowledge. Crop names represent the seedling object, pest and disease names represent the risk objects in the seedling process, and environmental parameters represent environmental conditions involved in the seedling process, such as temperature, humidity, light, ventilation, and water and fertilizer. The relationships between entities describe the correspondence between crop types, seedling stages, environmental parameters, pest and disease types, management measures, and key technical points.
[0104] The formation of video knowledge employs a different data path than text knowledge. For agricultural technology extension videos, expert lecture recordings, farmer practical operation videos, seedling cultivation instruction videos, and other similar videos, key video frames that represent the operation process or changes in seedling conditions are first extracted. The video audio is then converted into text, and a video content summary is generated based on the audio-to-text conversion. These key video frames, audio-to-text conversions, and video content summaries are used together to extract key knowledge points from the video, thereby obtaining video knowledge that represents the seedling cultivation operation process, expert explanations, practical steps, and precautions.
[0105] Image data primarily reflects seedling status, pest and disease status, and the state of the seedling environment. A target detection model is used to identify seedlings and pest / disease areas in seedling monitoring images, pest and disease sample images, seedling environment images, seedling condition images, pest and disease images, and environmental images. Then, an image classification model is used to determine the seedling status and pest / disease type. The results of seedling identification, pest / disease area identification, seedling status determination, and pest / disease type determination collectively form image knowledge, enabling this knowledge to characterize seedling growth, the location of pests / diseases, the type of pests / diseases, and the state of the seedling environment.
[0106] After textual, video, and image knowledge are formed, cross-modal associations are established using crop type, seedling stage, pest / disease type, environmental parameters, and knowledge type as association fields. For example, textual knowledge on temperature and humidity control during tomato seedling emergence, video knowledge explaining tomato seedling management, and image knowledge of tomato seedling status during seedling emergence are associated. For the topic of pest / disease control, textual knowledge on pest / disease control, video knowledge on pest / disease control operations, and image knowledge on pest / disease area identification are associated. Thus, knowledge from different data sources involving the same crop, the same seedling stage, the same pest / disease type, or the same environmental conditions can be organized in the same multi-source knowledge representation.
[0107] Figure 4 A flowchart of multi-source knowledge processing provided for embodiments of the present invention. Figure 4 The left dashed box represents the text knowledge processing flow, with input being "seedling cultivation documents / papers / expert experience," which sequentially undergoes "word segmentation," "named entity recognition (NER)," and "relation extraction" to obtain "structured text knowledge." The middle dashed box represents the video knowledge processing flow, with input being "seedling cultivation teaching / lectures / practical videos." One branch extracts "keyframes" to form "image features," while the other branch converts "speech to text" before entering "content summary generation." Image features and content summaries together form "video key knowledge points." The right dashed box represents the image knowledge processing flow, with input being "seedling condition / pests and diseases / environmental images," which undergo "target detection" and "feature extraction." Target detection results are used to "identify seedlings / pest and disease areas," while feature extraction results are used to "determine seedling condition / pest and disease type." Together, they form "image knowledge points." Structured text knowledge, video key knowledge points, and image knowledge points are processed together. Figure 4 The lower section, "Knowledge Fusion: Establishing Cross-Modal Relationships," further develops a "Unified Multimodal Knowledge Representation." The unified multimodal knowledge representation in the diagram corresponds to the unified multi-source knowledge representation described in this embodiment, which is used for subsequent knowledge vectorization, index construction, and retrieval enhancement generation.
[0108] After a unified multi-source knowledge representation is formed, knowledge from different modalities is converted into vector representations. Textual knowledge is semantically encoded using a pre-trained language model; in one specific implementation, a pre-trained model such as Sentence-BERT is used to encode the text into a 1024-dimensional semantic vector. The vector representation of video knowledge utilizes both keyframes and content summaries. Video keyframes are encoded into image vectors using a visual model such as ResNet, and the video summary text is encoded into text vectors using a text encoder. The two are then concatenated to form a bimodal vector. Image knowledge can be encoded into a text-aligned vector space using a visual-language pre-trained model such as CLIP.
[0109] The multi-level vector index library is structured hierarchically according to domain, crop type, seedling stage, and knowledge type. The first level can be divided by domain into vegetable seedling cultivation, fruit tree seedling cultivation, and flower seedling cultivation; the second level can be divided by crop type into tomato, cucumber, and watermelon; the third level can be divided by seedling stage into sowing period, emergence period, seedling stage, and mature seedling stage; and the fourth level can be divided by knowledge type into environmental control, water and fertilizer management, and pest and disease control. The vector representations corresponding to each knowledge are written into the hierarchical index nodes that match their domain, crop type, seedling stage, and knowledge type, forming the multi-level vector index library. In one specific implementation, a near-nearest neighbor ANN index is constructed using vector retrieval libraries such as FAISS to support the rapid acquisition of candidate knowledge within the target level.
[0110] The multi-level vector index library is not static after its initial creation. When knowledge content, knowledge weight, knowledge confidence, or cross-modal relationships change, the vector representation of the affected knowledge can be recalculated, and the corresponding level index nodes can be refreshed. Cache optimization is performed on frequently accessed knowledge, while low-value knowledge that has not been accessed for a long time is cleaned up. When subsequent feedback verification has established knowledge confidence and conflict states, frequently accessed knowledge with high feedback credibility can be cached first, while knowledge that has not been accessed for a long time, has low knowledge confidence, or consistently high conflict penalties is included in the low-value knowledge cleanup scope.
[0111] Step 200 is used to create a dynamic user profile that changes with user behavior. Historical seedling data includes basic user information and historical seedling records. The basic information includes planting region, facility type, and planting scale, while the historical seedling records include seedling crop, seedling time, environmental parameter settings, survival rate, and yield. Query records and behavioral preferences include query time, query content, click records, and dwell time. After parsing the above data, historical data features, query preference features, and behavioral preference features are formed respectively.
[0112] Historical data features, query preference features, and behavioral preference features are integrated to form user profile features, which are then encoded into user profile vectors using a user profile model. The user profile vector includes dimensions for crop preference, technical capability, focus areas, and time patterns. The crop preference dimension represents the user's planting preferences for different crops; the technical capability dimension represents the user's seedling cultivation skills; the focus area dimension represents the key technical points the user focuses on, such as environmental control and pest and disease prevention; and the time pattern dimension represents the user's seedling cultivation time patterns and seasonal preferences.
[0113] Dynamic user profiles are continuously updated with new interaction data. After a user generates new queries, clicks, dwell times, satisfaction feedback, adoption behaviors, or actual application data, the new data continues to influence the user profile vector. In one specific implementation, an exponentially weighted moving average (EWMA) method is used to fuse historical profiles and the latest behaviors, allowing the profile to reflect recent query needs and changes in the development stage while retaining long-term preferences. The latest user behaviors and feedback obtained during the feedback learning process are also fed back into the user profile update process, forming... Figure 2 The image update path is shown.
[0114] The scenario context required in step 300 is formed around the current actual seedling conditions. After obtaining the user's current seedling crop and seedling start time, the current seedling stage is determined based on the growth cycle of the current seedling crop and the seedling start time. The seedling stage includes sowing period, emergence period, seedling stage, and mature seedling stage. The scenario context also includes current environmental parameters, including temperature, humidity, light, ventilation, and water and fertilizer conditions. Depending on the available data, regional and pest / disease status can also be added. Regional information reflects the differences in seedling conditions in different regions, and pest / disease status reflects whether the current seedlings are affected by specific pests / diseases and their corresponding categories. The above information limits the search scope in step 400 and participates in the determination of scenario relevance and conflict penalty factors in step 500.
[0115] After receiving the user query in step 400, the system first identifies the user's intent and semantics. Intent identification determines whether the current query belongs to the corresponding type within knowledge query, problem diagnosis, or solution recommendation. Semantic parsing extracts query features such as crop type, seedling stage, environmental parameters, pest and disease types, and management measures. For colloquial, abbreviated, or omitted expressions, the query expression can be expanded through query rewriting to convert it into a searchable expression, and a query vector is formed based on the query understanding results.
[0116] The retrieval strategy varies depending on the knowledge status, user status, and scenario status. The timeliness of seedling knowledge is determined by the knowledge release time and update frequency. Dynamic user profiles are used to determine user preferences, and scenario context provides constraints such as the current seedling crop, seedling time, seedling stage, region, and environmental parameters. Based on the above information, the retrieval scope in the multi-level vector index is determined, and the evaluation weights corresponding to gating fusion similarity, scenario relevance, knowledge freshness, source authority, feedback credibility, and conflict penalty factor are adjusted during subsequent re-ranking. This allows the same user to use different retrieval scopes and ranking emphases at different seedling stages or under different environmental conditions.
[0117] Subsequently, the query vector and user profile vector are merged to form a personalized query representation, and vector similarity retrieval is performed within the defined hierarchical index range to return Top- Relevant knowledge fragments. Relevant knowledge fragments can come from textual knowledge, video knowledge, and image knowledge. Initial vector retrieval is used to obtain candidate knowledge related to the query semantics from a multi-level vector index library. Step 500 then combines these candidate knowledge with dynamic user profiles and the current scenario to perform a reliable re-ranking.
[0118] Source authority is used to evaluate the reliability of knowledge sources. It can be determined based on the type of knowledge source, the qualifications of the publishing entity, whether it has been reviewed by experts, whether it comes from publicly available agricultural guidance documents, and the stability of historical feedback. The system can pre-set different basic authority scores for government agricultural guidance documents, agricultural research papers, expert-reviewed knowledge, and knowledge uploaded by platform users. Then, it can adjust the basic authority scores based on the credibility of historical feedback to obtain a normalized source authority. If the source of a candidate knowledge fragment is reliable and its historical application feedback is stable, the corresponding source authority and feedback credibility are increased. If the candidate knowledge fragment is inconsistent with the current context in any key scenario dimension such as seedling stage, crop type, pest and disease status, or environmental parameters, the conflict risk value of the corresponding candidate knowledge fragment is increased.
[0119] Step 500 generates a unified comprehensive recommendation score based on the relevant knowledge fragments obtained in step 400. For the... Seedling cultivation knowledge, No. The user and the first The overall recommendation score is determined based on the context of each scene using the following formula:
[0120]
[0121] In the formula, Indicates the first Seedling knowledge for the first The user and in the th A comprehensive recommendation score for each scene context; This indicates the gated fusion similarity between relevant knowledge fragments and dynamic user profiles; This indicates the scene relevance between relevant knowledge fragments and the scene context; Indicates the freshness of knowledge; Indicates the authority of the source; Indicates the credibility of the feedback; Indicates the conflict penalty factor; For bias terms, to Each evaluation item has a weight, which is either a non-negative weight or constrained to be non-negative after training. This is a normalization function used to compress the overall recommendation score to the range of 0 to 1. Because... This indicates a risk of mismatch and is therefore deducted from the overall recommendation score.
[0122] Source authority It can be obtained by weighting the base score based on source type, expert review mark, and historical feedback stability. When the knowledge source is a reviewed agricultural technical document or expert-confirmed knowledge, Take the higher value; when the knowledge source is unapproved uploaded content and historical feedback fluctuates significantly, Take the lower value.
[0123] Gated fusion similarity is used to simultaneously measure the consistency between knowledge features and user profiles in feature directions and the degree of difference decay in key profile dimensions. Its calculation method is as follows:
[0124]
[0125] In the formula, Indicates the first Seedling knowledge in the first Normalized values for each portrait feature dimension Indicates the first The user profile in the first Normalized values for each portrait feature dimension and The values range from 0 to 1; Indicates the first Each portrait feature dimension weight and ; Indicates the first The difference attenuation coefficient of each portrait feature dimension and ; Indicates the number of dimensions of the portrait features; To prevent positive smoothing terms with a denominator of zero; The gating coefficient and The first term in the above formula is a weighted directional similarity, used to measure the directional consistency between knowledge features and user profiles across different profile feature dimensions; the second term is a profile feature difference attenuation term, used to reduce similarity when key profile dimensions differ significantly. The profile feature dimensions correspond to the user profile features in step 200, including crop preferences, technical capabilities, focus areas, and time patterns, and are combined with the current seedling stage to reflect the matching between knowledge features and user status.
[0126] Gating coefficient The weighted directional similarity and the proportion of the profile feature difference attenuation term are adaptively adjusted based on the interaction strength between knowledge features and user profiles, and determined according to the following formula:
[0127]
[0128] In the formula, For gating bias terms, For the first Gating weights for each portrait feature dimension This is a normalization function. When the knowledge features and user profiles match well across the main profile feature dimensions, the influence of the directional similarity term is increased; when there are significant differences between the two in key seedling stages, crop preferences, or areas of focus, the suppression effect of the profile feature difference attenuation term is increased.
[0129] Scene relevance We employ a weighted distance calculation with normalized intervals to measure the degree of matching between candidate knowledge fragments and the current scene context in dimensions such as seedling time, region, seedling stage, environmental parameters, and pest and disease status.
[0130]
[0131] In the formula, Indicates the first Seedling knowledge in the first The values in each scene dimension Indicates the first The scene context in the first The values in each scene dimension and They represent the first The maximum and minimum values for each scene dimension. Indicates the first Each scene dimension weight and , Indicates the number of scene dimensions. This is a positive smoothing term. (Using...) The function truncates the lower bound of scene relevance, making it suitable for scenarios with significant scene differences. Not less than 0. Scene dimensions include seedling time, region, seedling stage, environmental parameters, and pest and disease status. If some scene dimensions do not have valid values, differences can be calculated only for scene dimensions with valid values, and normalization can be performed using the weights corresponding to the valid scene dimensions to prevent missing scene items from directly expanding the scene distance.
[0132] Knowledge freshness Taking into account both the knowledge release time and the knowledge update frequency, the following formula is used to determine:
[0133]
[0134] In the formula, Indicates the first The difference between the time the seedling knowledge was published and the current time, in days; Represents the time decay coefficient and ; Indicates the impact coefficient of update frequency and ; Indicates the first The unit cycle update frequency of seedling cultivation knowledge; This represents the maximum unit cycle update frequency in the candidate knowledge set; This is a positive smoothing term. The exponential decay term reflects the decrease in freshness of knowledge as it becomes outdated over time, while the update frequency term improves the freshness of continuously maintained knowledge. The effect of the denominator on the update frequency is normalized to make... Under normal value ranges, it remains between 0 and 1. For continuously updated data sources, its continuous update status can be reflected in... In the statistical process, to distinguish between knowledge that is continuously maintained and knowledge that has been discontinued for a long time, the structure of the above knowledge freshness formula is not changed.
[0135] Feedback credibility Explicit feedback, implicit adoption, and actual application results are all incorporated into the evaluation to measure the impact of user feedback and actual seedling cultivation outcomes on the first... Level of support for the knowledge points:
[0136]
[0137] In the formula, This represents the user satisfaction rating. The actual ratings in the feedback data are on a scale of 1 to 5. The feedback credibility calculation variable is... The value range is from 0 to 5; This indicates an adoption flag with a value of 0 or 1; This represents the actual application effect meeting the standard value, and the value ranges from 0 to 1. , and Weights for non-negative feedback dimensions; This is a positive smoothing term. Normalization is used to prevent feedback credibility from exceeding limits due to changes in feedback dimension weights. In addition to the above quantification terms, feedback source, feedback time, feedback completeness, and the consistency between feedback content and actual application results are also used in the knowledge verification process to determine whether the feedback data is stable, and participate in the aggregation of feedback credibility under different users, different seedling stages, and different environmental conditions.
[0138] Conflict Punishment Factor This is used to indicate the degree of mismatch between candidate knowledge fragments and the current seedling scenario. When the crop type, seedling stage, environmental parameter range, or pest and disease status corresponding to the candidate knowledge fragment is inconsistent with the current scenario, even if the candidate knowledge fragment has a high semantic similarity to the user query, the overall recommendation score is reduced by increasing the conflict penalty factor. The conflict penalty factor is determined according to the following formula:
[0139]
[0140] In the formula, Indicates the first The knowledge item is relative to the first The normalized conflict risk value for each scenario context, with a value ranging from 0 to 1; Indicates the threshold for triggering conflict calculations. When No more than hour, When the value is 0, candidate knowledge fragments are not deducted due to scene conflicts; when Exceed hour, It increases as the risk of conflict increases. It can be obtained by weighted sum of conflicts in crop type, seedling stage, environmental parameters, and pest and disease status.
[0141] The comprehensive recommendation score for each relevant knowledge fragment is calculated based on the aforementioned evaluation metrics, and then re-ranked accordingly. Knowledge with higher comprehensive recommendation scores and lower conflict penalty factors is prioritized for step 600 to form structured contextual information, making it less likely for semantically relevant knowledge that clearly conflicts with the current crop, current seedling stage, or current environmental conditions to enter the high-priority knowledge set.
[0142] Step 600 organizes the reordered relevant knowledge fragments into structured context information according to logical relationships. The structured context information includes the problem background, relevant knowledge points, operational steps, and precautions. The problem background explains the seedling object, seedling stage, and problem scenario corresponding to the user's query; relevant knowledge points provide the seedling knowledge basis related to the user's query; operational steps provide executable seedling management measures; and precautions highlight matters that need to be avoided or emphasized during the seedling process.
[0143] The user query and structured context information are used as input to the domain-adaptive language model, which is then guided by prompt word templates to generate a structured answer. The answer includes a direct response, knowledge source citations, and related recommendations. The direct response addresses the user query, the knowledge source citations explain the relevant knowledge fragments upon which the answer is based, and the related recommendations provide extended seedling cultivation knowledge or subsequent management suggestions relevant to the current query.
[0144] Figure 5 The flowchart for retrieval enhancement generation provided in the embodiments of the present invention. Figure 5 It is used to represent the main data flow relationships in the process of retrieval enhancement generation, but does not draw each evaluation metric within knowledge reordering as an independent calculation box. Figure 5The process, from top to bottom, includes "User Query," "Query Understanding: Intent Recognition and Semantic Parsing," "Generate Query Vector," "User Profile Vector," "Fusion with User Profile Vector," "Personalized Retrieval: Similarity Retrieval in a Multi-Level Vector Index," "Return Top-K Relevant Knowledge Fragments," "Knowledge Reordering," "Context Construction," "Input Large Language Model (with Prompt Word Templates)," and "Generate Structured Answers." In the knowledge reordering box, "Semantic Similarity / Timeliness / Authority / User Preference" are simplified annotations of the ordering criteria: semantic similarity represents the degree of knowledge relevance in the initial vector retrieval stage; user preference is represented by dynamic user profiles; timeliness and authority correspond to knowledge freshness and source authority, respectively. In this embodiment, during the actual reordering process, step 500 also determines gated fusion similarity, scene relevance, feedback credibility, and conflict penalty factor. These factors together form a comprehensive recommendation score. The reordered relevant knowledge fragments enter the context construction stage, forming structured context information according to the problem background, knowledge points, operation steps, and precautions. Finally, the user query, structured context information, and prompt word template are provided to the domain-adaptive language model to obtain a structured answer that includes direct answers, source citations, and related recommendations.
[0145] Step 700 collects multi-dimensional feedback data after the answer is generated. Explicit feedback includes user satisfaction ratings, likes, and dislikes, with satisfaction ratings using a scale of 1 to 5. Implicit feedback includes whether users adopted the suggestions and whether they asked follow-up questions. Application feedback includes changes in survival rate, yield, pest and disease incidence, and environmental parameter compliance after users followed the suggestions. The latest user behavior and feedback are also used to dynamically update user profiles.
[0146] Application effect deviation is determined based on the difference between the expected effect of the answer or recommended content and the actual seedling results. When the recommended content involves improving the survival rate, the application effect deviation can be obtained from the normalized difference between the target survival rate and the actual survival rate; when the recommended content involves pest and disease control, the application effect deviation can be obtained from the normalized difference between the expected pest and disease incidence rate and the actual pest and disease incidence rate; depending on the specific recommended content, the application effect deviation can also be obtained by normalizing at least one of the differences in yield before and after the recommendation or the differences in environmental parameter compliance.
[0147] User feedback and actual seedling results are used to verify the accuracy and applicability of the knowledge used to generate the answers. New knowledge that passes verification can be added to the seedling knowledge base and have its corresponding vector representation generated; existing knowledge that passes verification has its corresponding knowledge weight increased and its knowledge relationships updated; existing knowledge with application effect deviations, insufficient feedback credibility, or conflicts with the current scenario has its knowledge weight reduced, a conflict marker added, or marked as pending review. Updating knowledge relationships includes adjusting cross-modal relationships between textual, video, and image knowledge, as well as adjusting knowledge connections between different seedling stages, different crop types, different pest and disease types, and different environmental parameters.
[0148] The credibility of feedback generated by the same knowledge under different users, different seedling stages, and different environmental parameters can be aggregated. When the aggregated feedback credibility is high and the application effect deviation is low, the confidence of the corresponding knowledge is increased; when the feedback credibility is low, the application effect deviation is high, or the conflict risk is consistently higher than the corresponding threshold, the knowledge confidence is decreased and a pending review flag is triggered. For knowledge whose content or cross-modal associations have changed significantly, the knowledge verification results can also be used to reconstruct the cross-modal association relationships between affected knowledge.
[0149] To determine whether knowledge needs to be updated, demoted, reviewed, or trigger an index refresh, a knowledge update trigger value is calculated. :
[0150]
[0151] In the formula, Indicates an indicator function; Indicates the feedback credibility threshold; Indicates the first The knowledge in the first The deviation in application performance for each user can be obtained by normalizing at least one of the following: the difference in survival rate before and after recommendation, the difference in yield, the difference in the incidence of pests and diseases, or the difference in the compliance of environmental parameters. Indicates the threshold for deviation in application effect; Indicates the threshold for triggering conflict penalties; This indicates the threshold for triggering knowledge decay over time. Still indicating the first The difference between the publication time of a knowledge item and the current time. The value of represents the number of update trigger conditions that have been met for the current knowledge.
[0152] when When the preset update threshold is reached, perform knowledge weight adjustment, knowledge confidence update, incremental vector recalculation, or hierarchical index refresh on the corresponding knowledge; when When the preset review threshold is reached, the corresponding knowledge is added to the review queue and its recommendation priority is reduced. Feedback credibility, application effect deviation, conflict penalty, and knowledge time status no longer only affect the current answer evaluation, but can directly influence the update judgment of knowledge status and index status.
[0153] When performing incremental vector recalculation in step 800, knowledge nodes whose knowledge content, knowledge weight, knowledge confidence, or cross-modal associations have changed are identified as knowledge nodes affected by the update, and adjacent associated nodes that have associations with these knowledge nodes are also identified. Adjacent associated nodes can be nodes connected through cross-modal associations between text knowledge, video knowledge, and image knowledge, or nodes that form knowledge connection relationships through seedling stage, crop type, pest and disease type, or environmental parameters.
[0154] Only the knowledge nodes affected by the update and their adjacent related nodes are recalculated in vector representation, and then the hierarchical index nodes corresponding to the domain, crop type, seedling stage, and knowledge type of these nodes are refreshed; index nodes unaffected by this knowledge state change remain unchanged. In addition to incremental updates driven by knowledge update trigger values, periodic knowledge vector and index maintenance can also be performed. Knowledge accessed frequently and with high feedback confidence can be included in the cache optimization scope, while knowledge that has not been accessed for a long time, has low knowledge confidence, or consistently high conflict penalties can be cleaned up as low-value knowledge. Feedback data can also be used to adjust personalized recommendation strategies, enabling subsequent recommendations to better suit the user's seedling stage, technical level, and focus.
[0155] Figure 6 The feedback learning flowchart is provided for embodiments of the present invention. Figure 6 The main processing relationship between multi-dimensional feedback collection, knowledge verification, knowledge update, index update and recommendation strategy optimization is used as the display object. Therefore, feedback credibility, application effect deviation, knowledge update trigger value, knowledge confidence and hierarchical index incremental refresh are not set as independent flowcharts. Figure 6The top "Multi-dimensional Feedback Collection" is divided into four paths: "Explicit Feedback: Satisfaction rating, likes / dislikes," "Implicit Feedback: Whether adopted, whether follow-up questions asked," "Application Feedback: Actual data such as survival rate and yield," and "Dynamically updated user profile." Explicit, implicit, and application feedback enter the "Knowledge Verification: Comparison with actual seedling results" stage. In this knowledge verification stage, step 700 determines the credibility of the feedback, the deviation in application effect, and the risk of conflict. When the verification results meet the knowledge retention conditions, the system enters the "Knowledge Update: Add new knowledge / increase weight / update associations" path, which can also improve the knowledge confidence. When the verification results have deviations, insufficient feedback credibility, or a high risk of conflict, the system enters the "Reduce the weight of old knowledge / mark as pending review" path. After the knowledge state changes and the knowledge update trigger value meets the preset update conditions, the "Index Update: Recalculate Vectors, Update Index Library, Cache Optimization, Clean Up Low-Value Knowledge" step in the diagram represents the index maintenance process described in step 800. Under incremental update conditions, recalculating vectors and updating the index library only apply to the knowledge nodes affected by the update and their corresponding hierarchical indexes; unaffected hierarchical indexes remain unchanged. The updated index enters the vector index library along the "Update Index" path. "Recommendation Strategy Optimization: Adjust Recommendation Algorithm and Parameters Based on Feedback Data" adjusts the ranking emphasis for the next recommendation based on feedback data and acts on the active recommendation module along the "Affect Next Recommendation" path.
[0156] Step 900 first determines the current seedling stage based on the growth cycle of the user's current crop and the start time of seedling cultivation. Then, it matches relevant knowledge, technical points, and precautions corresponding to the current seedling stage from the updated seedling knowledge base. The matched content may also include key technical points, common problem warnings, and management suggestions for the current seedling stage. Personalized sorting is performed based on the user's technical capabilities and priorities in the dynamic user profile. For example, for users concerned with environmental control, knowledge related to temperature, humidity, light, and ventilation is prioritized; for users concerned with pest and disease control, knowledge related to pest and disease identification, control measures, and warnings is prioritized.
[0157] For candidate knowledge in the updated seedling knowledge base, step 500 can be followed to redetermine the gating fusion similarity, scenario relevance, knowledge freshness, source authority, updated feedback credibility, and conflict penalty factor, and recalculate the comprehensive recommendation score. Knowledge with a high overall recommendation score and a low conflict penalty factor will be prioritized for inclusion in the personalized knowledge recommendation list as relevant knowledge, technical points, and precautions. When the conflict penalty factor exceeds the corresponding conflict threshold or the feedback credibility is lower than the corresponding feedback threshold, the corresponding knowledge can be marked as pending review and its recommendation priority will be reduced.
[0158] In one specific implementation, recommended content is provided to users via app push notifications, SMS, or WeChat, with the push time selectable in the morning or evening. Recommended content includes relevant knowledge, technical points, and precautions, and may also include warnings about common problems and management suggestions. User queries, satisfaction feedback, adoption behaviors, follow-up inquiries, and actual seedling results generated after receiving the recommended content continue to step 700, thus forming the next round of knowledge verification, index maintenance, and recommendation strategy adjustment.
[0159] Based on the above-described method implementation, this embodiment also provides a dynamic update and precise push system for the seedling knowledge base based on adaptive RAG, used to execute the aforementioned method. Figure 3 This is a system architecture diagram provided in an embodiment of the present invention. Figure 3 The system is divided into a data acquisition layer, a knowledge processing layer, and an interaction service layer. Users can query, receive returned answers, receive proactive push notifications, and receive feedback through the interaction service layer.
[0160] The data acquisition layer comprises a web crawler submodule, an API interface submodule, and a user upload submodule. The web crawler submodule collects publicly available seedling cultivation text, video, and image data from the internet; the API interface submodule accesses external agricultural knowledge data or seedling cultivation data; and the user upload submodule receives user-uploaded images, videos, seedling cultivation records, and experience data related to the seedling cultivation process. The data generated by these three submodules is then jointly transmitted to the knowledge processing layer.
[0161] The knowledge processing layer includes a knowledge preprocessing module, a text processing submodule, a video processing submodule, an image processing submodule, a knowledge fusion submodule, a vectorization module, a text encoder, a video encoder, an image encoder, and an index building module. The knowledge preprocessing module is responsible for cleaning, labeling, and structuring the raw knowledge data; the text processing submodule performs word segmentation, named entity recognition, and relation extraction; the video processing submodule performs keyframe extraction, speech-to-text conversion, and content summarization; the image processing submodule performs object detection, feature extraction, crop condition assessment, and pest and disease type identification; the knowledge fusion submodule establishes cross-modal relationships between textual, video, and image knowledge; the vectorization module converts structured knowledge into vector representations; and the text encoder, video encoder, and image encoder each generate knowledge vectors for their respective modalities.
[0162] The index building module includes a vector indexing submodule, a hierarchical classification submodule, and a cache optimization submodule. The hierarchical classification submodule establishes hierarchical index nodes according to domain, crop type, seedling stage, and knowledge type. The vector indexing submodule establishes corresponding vector indexes based on the vector representations of different modalities of knowledge. The cache optimization submodule maintains the cache of frequently accessed knowledge and cleans up low-value knowledge in conjunction with the knowledge update process.
[0163] The interactive service layer includes a retrieval module, a generation module, a user profiling module, a feedback learning module, and a recommendation module. Within each module, there are sub-modules for query understanding, similarity calculation, result reordering, context construction, prompt word engineering, answer generation, data collection, user profile generation, user profile update, feedback collection, knowledge verification, index update, stage identification, knowledge matching, recommendation generation, and push.
[0164] The retrieval module uses the query understanding submodule to identify the intent and semantics of user queries. The similarity calculation submodule merges the query vector and the user profile vector provided by the user profile module and retrieves relevant knowledge fragments from the multi-level vector index library. The result re-ranking submodule forms a comprehensive recommendation score based on gating fusion similarity, scene relevance, knowledge freshness, source authority, feedback credibility, and conflict penalty factor, and completes the ranking of relevant knowledge fragments. Figure 3 In the process, the user query enters the retrieval module along the "query" path, and the dynamic user profile participates in the formation of personalized query representation along the path from the user profile module to the retrieval module.
[0165] The generation module includes a context construction submodule, a prompt word engineering submodule, and an answer generation submodule. The context construction submodule constructs structured context information from reordered relevant knowledge fragments according to the question background, relevant knowledge points, operational steps, and precautions. The prompt word engineering submodule forms prompt word templates to guide the generation of structured answers using a domain-adaptive language model. The answer generation submodule generates answers containing direct answers, knowledge source citations, and relevant recommendations based on user queries and structured context information, and follows the... Figure 3 The "Return to Answer" path is provided to the user.
[0166] The user profiling module comprises a data acquisition submodule, a profile generation submodule, and a profile update submodule. The data acquisition submodule collects users' historical seedling data, query records, behavioral preferences, and feedback data. The profile generation submodule encodes user-related data into user profile vectors. The profile update submodule updates the dynamic user profile based on the user's latest behavior and feedback. These user profile vectors are then passed to the retrieval and recommendation modules for personalized retrieval and personalized recommendations. Figure 3 The user behavior data shown is processed through the data acquisition submodule, the profile generation submodule, and the profile update submodule to form a profile update path.
[0167] The feedback learning module includes a feedback collection submodule, a knowledge verification submodule, and an index update submodule. The feedback collection submodule collects explicit feedback, implicit feedback, and application feedback; the knowledge verification submodule compares user feedback and actual application effects with actual seedling cultivation results to determine feedback credibility, application effect deviations, and conflict risks, and verifies the accuracy and applicability of the knowledge; when the knowledge update trigger value meets the preset update conditions, the index update submodule performs incremental vector recalculation and hierarchical index refresh on the affected knowledge nodes. Figure 3 In the process, user feedback enters the feedback learning module. The feedback learning module forms an updated index path through the feedback collection submodule, knowledge verification submodule, and index update submodule, and then passes the updated index to the vector index submodule in the knowledge processing layer.
[0168] The recommendation module includes a stage identification submodule, a knowledge matching submodule, a recommendation generation submodule, and a push submodule. The stage identification submodule determines the current seedling stage based on the user's current seedling crop and seedling time. The knowledge matching submodule matches relevant knowledge, technical points, and precautions corresponding to the current seedling stage from the updated seedling knowledge base. The recommendation generation submodule combines dynamic user profiles to form a personalized knowledge recommendation list. The push submodule proactively provides the recommended content from the personalized knowledge recommendation list to the user.
[0169] Figure 3 The collaborative relationships between the three layers are also illustrated. The data acquisition layer transmits the acquired knowledge data to the knowledge preprocessing module; the knowledge preprocessing module calls the text processing submodule, video processing submodule, and image processing submodule respectively, and the processing results enter the knowledge fusion submodule to form fused knowledge; the fused knowledge is vectorized by the vectorization module, text encoder, video encoder, and image encoder to form a vector representation; the index building module forms a vector index, hierarchical classification, and cache optimization results based on the vector representation. In the interactive service layer, the retrieval module calls the vector index to perform retrieval, the feedback learning module updates the index through the index update submodule, and the recommendation module completes proactive recommendations based on user profiles and the current seedling stage. User queries, returned answers, proactive push notifications, user feedback, user behavior data, and updated indexes together constitute the core of the interactive service layer. Figure 3 The data exchange relationship between the interactive service layer, the knowledge processing layer, and the user.
[0170] Through the above implementation methods, multi-source seedling knowledge is formed into vector representations and multi-level vector indexes starting from collection, knowledge extraction, and cross-modal association. User historical seedling data and behavioral data form dynamic user profiles, and user queries and the current seedling scenario jointly limit the retrieval scope. Relevant knowledge fragments are re-ranked based on gated fusion similarity, scenario relevance, knowledge freshness, source authority, feedback credibility, and conflict penalty factors before participating in answer generation. Feedback after the actual application of the answers and seedling results further form updated feedback credibility, application effect deviation, and knowledge update trigger values. Knowledge that meets the update conditions undergoes linked updates of knowledge weight, knowledge confidence, cross-modal association, and corresponding level indexes. The updated seedling knowledge base continues to be used for subsequent retrieval and proactive recommendation.
[0171] The different embodiments described in this specification can be used interchangeably. For the same technical content that is not repeated in a certain embodiment, the corresponding part above can be referred to. The above specific embodiments are used to illustrate the technical principles and implementation process of the present invention, and are not intended to limit the scope of protection of the present invention. Without departing from the technical concept of the present invention, adaptive adjustments made by those skilled in the art to the specific implementation form according to the actual seedling cultivation object, data source and knowledge base size can all be implemented using the aforementioned technical relationships.
[0172] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.
[0173] Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. Furthermore, those skilled in the art will recognize that, based on the ideas of this invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this invention.
Claims
1. A method for dynamic updating and accurate pushing of a seedling raising knowledge base based on adaptive RAG, characterized in that, include: A multi-level vector index library was constructed based on the seedling knowledge base; Dynamic user profiles are constructed based on users' historical seedling data, query records, and behavioral preferences, and user profile vectors are formed. A scenario context is generated based on the user's current seedling crop, seedling time, and environmental parameters; The scenario context includes the current seedling stage determined based on the current seedling crop and the seedling time; By integrating the query vector formed by user queries and the user profile vector, relevant knowledge fragments are retrieved from the multi-level vector index library; The gating fusion similarity between the relevant knowledge fragment and the dynamic user profile, the scene relevance between the relevant knowledge fragment and the scene context, the knowledge freshness and source authority of the relevant knowledge fragment, the feedback credibility corresponding to the relevant knowledge fragment, and the conflict penalty factor of the relevant knowledge fragment relative to the scene context are determined to form a comprehensive recommendation score, and the relevant knowledge fragments are re-ranked according to the comprehensive recommendation score. Based on the reordered relevant knowledge fragments, structured context information is constructed. The user query and the structured context information are then input into a domain-adaptive language model to generate an answer. Collect user feedback and the actual application effect after adopting the answer, update the feedback credibility corresponding to the relevant knowledge fragment used to construct the structured context information, determine the application effect deviation, and form a knowledge update trigger value based on the updated feedback credibility, the application effect deviation, the conflict penalty factor, and the time difference between the publication time of the corresponding knowledge and the current time. When the knowledge update trigger value meets the preset update conditions, the knowledge weight and knowledge confidence corresponding to the relevant knowledge fragment used to construct the structured context information are updated, and the incremental vector is recalculated and the corresponding level index is refreshed for the knowledge affected by the update. Based on the updated seedling knowledge base, the dynamic user profile, and the current seedling stage, a personalized knowledge recommendation list is generated, and the recommended content in the personalized knowledge recommendation list is pushed to the user.
2. The method according to claim 1, wherein, A multi-level vector index library is constructed based on the seedling knowledge base, including: Acquire text knowledge, video knowledge, and image knowledge from the seedling knowledge base; Based on crop type, seedling stage, pest and disease type, and environmental parameters, establish cross-modal association relationships between the text knowledge, video knowledge, and image knowledge to form a unified multi-source knowledge representation; The text knowledge, the video knowledge, and the image knowledge are respectively converted into corresponding vector representations; Set up hierarchical index nodes according to domain, crop type, seedling stage, and knowledge type; Based on the domain, crop type, seedling stage, and knowledge type corresponding to the unified multi-source knowledge representation, the vector representation is written into the corresponding hierarchical index node to form the multi-level vector index library.
3. The method according to claim 1, wherein, By fusing the query vector generated from user queries and the user profile vector, relevant knowledge fragments are retrieved from the multi-level vector index library, including: Based on the dynamic user profile, the scenario context, and the publication time and update frequency of seedling knowledge, the retrieval scope in the multi-level vector index library is determined; The query vector and the user profile vector are combined to form a personalized query representation; Based on the personalized query representation, vector retrieval is performed within the retrieval range to obtain the relevant knowledge fragments.
4. The method according to claim 1, wherein, The following steps are taken to determine the gated fusion similarity between the relevant knowledge fragment and the dynamic user profile, the scene relevance between the relevant knowledge fragment and the scene context, the knowledge freshness and source authority of the relevant knowledge fragment, the feedback credibility corresponding to the relevant knowledge fragment, and the conflict penalty factor of the relevant knowledge fragment relative to the scene context, forming a comprehensive recommendation score. The relevant knowledge fragments are then re-ranked according to the comprehensive recommendation score, including: For each of the relevant knowledge fragments, the gating fusion similarity, the scene relevance, the knowledge freshness, the source authority, the feedback credibility, and the conflict penalty factor are determined respectively. The gating fusion similarity, scene relevance, knowledge freshness, source authority, and feedback credibility are weighted respectively, and the weighted conflict penalty factor is used as a deduction item to form a comprehensive evaluation result. The comprehensive evaluation results are normalized to form the comprehensive recommendation score; The relevant knowledge fragments are reordered based on the comprehensive recommendation score.
5. The method according to claim 1, wherein, The determination of the gated fusion similarity includes: Obtain the normalized values of the relevant knowledge fragments and the dynamic user profile in each profile feature dimension; Based on the relevant knowledge fragments and the normalized values of the dynamic user profiles on the corresponding profile feature dimensions, as well as the weights of the corresponding profile feature dimensions, the weighted directional similarity is determined. Based on the differences between the relevant knowledge fragments and the dynamic user profile in the corresponding profile feature dimensions, and the difference attenuation coefficient of the corresponding profile feature dimensions, determine the profile feature difference attenuation term; The gating coefficient is determined based on the interaction between the normalized values of the relevant knowledge fragments on each feature dimension of the profile and the normalized values of the dynamic user profile. The gating coefficient is used to adjust the proportions of the weighted directional similarity and the portrait feature difference attenuation term in the similarity calculation to form the gated fusion similarity.
6. The method according to claim 1, wherein, The determination of the scenario relevance, the knowledge freshness, and the source authority includes: The relevance of the scenario is determined based on the differences between the relevant knowledge fragments and the scenario context in terms of seedling time, region, seedling stage, environmental parameters, and pest and disease status, as well as the weight of each scenario dimension. The freshness of the knowledge is determined based on the time difference between the publication time of the relevant knowledge fragment and the current time, as well as the update frequency of the relevant knowledge fragment. The authority of the source is determined based on the type of knowledge source, the qualifications of the publishing entity, the status of expert review, and the stability of historical feedback for the relevant knowledge fragments.
7. The method for dynamic updating and precise push of seedling knowledge base based on adaptive RAG as described in claim 1, characterized in that, The determination of the feedback credibility and the conflict penalty factor includes: Obtain the satisfaction score, adoption mark, and actual application effect achievement value corresponding to the relevant knowledge fragments; The satisfaction score, the adoption marker, and the actual application effect achievement value are weighted and normalized to determine the credibility of the feedback. Based on the degree of conflict between the relevant knowledge fragments and the scenario context in terms of crop type, seedling stage, environmental parameters, and pest and disease status, the conflict risk value is determined. When the conflict risk value exceeds a preset conflict threshold, the conflict penalty factor is determined based on the degree to which the conflict risk value exceeds the preset conflict threshold. When the conflict risk value does not exceed the preset conflict threshold, the conflict penalty factor is set to zero.
8. The method for dynamic updating and precise push of seedling knowledge base based on adaptive RAG as described in claim 1, characterized in that, Collect user feedback and the actual application effect after adopting the answer, update the feedback credibility corresponding to the relevant knowledge fragment used to construct the structured context information, determine the application effect deviation, and form a knowledge update trigger value based on the updated feedback credibility, the application effect deviation, the conflict penalty factor, and the time difference between the publication time of the corresponding knowledge and the current time, including: Collect satisfaction feedback, adoption feedback, and follow-up feedback to form the aforementioned user feedback; The actual application effect is formed by collecting at least one of the following after the user adopts the answer: survival rate, yield, changes in pest and disease incidence, and environmental parameter compliance. Based on the user feedback and the actual application effect, update the feedback credibility corresponding to the relevant knowledge fragments used to construct the structured context information; The application effect deviation is determined based on the difference between the expected application result corresponding to the answer and the actual application effect; Determine the time difference between the publication time of the relevant knowledge fragment used to construct the structured context information and the current time; The updated feedback credibility, application effect deviation, conflict penalty factor, and time difference are compared with the corresponding preset trigger thresholds, respectively. The knowledge update trigger value is formed based on the number of times the corresponding trigger condition is met.
9. The method for dynamic updating and precise push of seedling knowledge base based on adaptive RAG according to claim 1, characterized in that, When the knowledge update trigger value meets the preset update conditions, the knowledge weights and knowledge confidence levels corresponding to the relevant knowledge fragments used to construct the structured context information are updated. Incremental vector recalculation is performed on the knowledge affected by the update, and the corresponding level index is refreshed, including: Based on the updated feedback credibility and the deviation in application effect, adjust the knowledge weight and knowledge confidence of the relevant knowledge fragments. Update the cross-modal association relationships between the relevant knowledge fragments based on the knowledge verification results; Knowledge whose knowledge weight, knowledge confidence, or cross-modal association changes are identified as knowledge affected by the update; Identify adjacent knowledge that is associated with the knowledge affected by the update; The incremental vector is recalculated for the knowledge affected by the update and the adjacent knowledge. Based on the vector representation recalculated from the incremental vector, refresh the hierarchical indexes corresponding to the domain, crop type, seedling stage, and knowledge type of the knowledge affected by the update and the adjacent knowledge; Keep the hierarchical indexes that are not affected by the update unchanged.
10. The method for dynamic updating and precise push of seedling knowledge base based on adaptive RAG according to claim 1, characterized in that, Based on the updated seedling knowledge base, the dynamic user profile, and the current seedling stage, a personalized knowledge recommendation list is generated, and recommended content from the personalized knowledge recommendation list is pushed to the user, including: Match relevant knowledge, technical points, and precautions corresponding to the current seedling stage from the updated seedling knowledge base; The gating fusion similarity, scene relevance, knowledge freshness, source authority, feedback credibility, and conflict penalty factor corresponding to the matched relevant knowledge are determined, and a corresponding comprehensive recommendation score is formed. Based on the dynamic user profile and the comprehensive recommendation score, the relevant knowledge, technical points and precautions obtained by matching are sorted in a personalized manner to form the personalized knowledge recommendation list; The recommended content from the personalized knowledge recommendation list is pushed to the user.
