Man-machine collaborative generative engine optimization system and recommendation method

Abstract

The invention discloses a man-machine collaborative generative engine optimization system and recommendation method, and relates to the technical field of generative engine optimizing.The method comprises the steps that dominant intentions and implicit intentions in user input content are analyzed through an intention management module, intention network nodes are constructed, and each node records the generation time sequence, intensity evolution and association dimension of the intentions; identifying intention activeness of the nodes through multi-dimensional dynamic scores, executing hotspot memory layer storage on high-activeness data, and executing temperature point memory layer degradation on low-activeness data; generating a cross-session index tag according to the association dimension between the intention network nodes; comparing the difference between the current decision and the historical optimal scheme, marking a key deviation point, and generating a backtracking report; and generating an optimization decision according to the key deviation point of the backtracking report, storing the optimization decision as a new node in a memory layer, and updating the multi-dimensional dynamic score according to a decision effect.

Description

Technical Field

[0001] This invention relates to the field of generative engine optimization technology, and more particularly to a human-machine collaborative generative engine optimization system and recommendation method. Background Technology

[0002] In recent years, generative artificial intelligence technology has made significant progress in natural language processing, content generation, and intelligent recommendation, and has been widely applied in scenarios such as dialogue systems, writing assistance, and personalized services. However, existing generative engines still have many limitations: most systems can only recognize the explicit intent of user input, and have a weak understanding of implicit intent, emotional context, and contextual relationships, resulting in a discrepancy between the output and the user's actual needs; the memory mechanism lacks dynamic management capabilities and cannot achieve hierarchical storage and efficient retrieval based on intent activity, resulting in slow response to high-frequency requests and easy forgetting of low-frequency intents; the system often struggles to maintain cross-conversation intent. Figure 1 The generation process suffers from insufficient modeling of long-term user preferences and pattern evolution, hindering consistency and behavioral continuity. The "black box" nature of the generation process results in poor interpretability of decisions and a lack of mechanisms for retrospective analysis and optimization of historical deviations. Furthermore, limited ability to integrate external real-time data leads to a lack of timeliness and contextual relevance in generated content. Therefore, there is an urgent need to develop a generative engine optimization system and method that supports human-machine collaboration, dynamic intent management, multi-dimensional memory optimization, and cross-session decision retrospective capabilities to comprehensively improve the accuracy, personalization, and user experience of generated content. Summary of the Invention

[0003] This application addresses the problems of one-sided intent cognition, rigid memory mechanisms, fragmented cross-session decision-making, black-box decision-making process, and weak real-time data fusion in existing technologies by providing a human-machine collaborative generative engine optimization system and recommendation method. It achieves the technical effects of three-dimensional intent parsing, intelligent scheduling of memory resources, continuous cross-session intent association, traceable and optimizable decision-making, and efficient fusion of multi-source data.

[0004] This application provides a human-machine collaborative generative engine optimization recommendation method, including: S1: The intent management module parses the explicit and implicit intents in the user's input content and constructs intent network nodes. Each node records the generation sequence, intensity evolution, and association dimensions of the intent. S2: Identify the intent activity of nodes through multi-dimensional dynamic scoring, perform hot spot memory layer storage for high-activity data, and perform warm spot memory layer degradation for low-activity data; S3: Generate cross-session index labels based on the correlation dimension between intent network nodes. When the decision deviation detection module detects that the current decision deviates from the historical pattern, it loads the associated historical data based on the index labels; compares the difference between the current decision and the historical optimal solution and marks the key deviation points, generating a backtracking report; S4: Generate optimization decisions based on key deviations in the backtracking report, store the optimization decisions as new nodes in the memory layer, and inherit the generation sequence, intensity evolution baseline value and correlation dimension of the original node. Update the multi-dimensional dynamic score based on the decision effect.

[0005] Furthermore, the intent management module includes: a multimodal parsing engine and a node generator; The multimodal parsing engine processes text, speech, and image input streams, and outputs a fused feature vector containing semantic features, emotional intensity, and visual focus. , in, To fuse feature vectors, For semantic features, For emotional intensity, As the visual focus; The node generator is used to construct intent network nodes, which record the generation time sequence, intensity evolution and association dimensions, and generate unique node identifiers.

[0006] Furthermore, the multi-dimensional dynamic scoring includes: time decay coefficient, intensity evolution gradient, association density, and comprehensive activity score; The time decay coefficient refers to the degree to which the activity of the intended node decays over time. , in, The time decay coefficient, For the domain attenuation coefficient, The current system time. Create a timestamp for the node; The intensity evolution gradient is used to capture the intensity change trend of the intended node: , in, The gradient value, For strength standard deviation, At the current intensity, The intensity is the historical average. The association density is used to measure the connection strength of nodes in the intent network: , in, For correlation density, This represents the number of associated nodes. For the weight of the associated edge, Cosine similarity of feature vectors; The overall activity score is a global activity metric for intent nodes: , in, To score overall activity level, , , For the corresponding weights, and .

[0007] Furthermore, the cross-session index label includes: extracting the association dimension of intent network nodes and generating SHA-256 digest values; dividing time-segment blocks according to the node generation time sequence; and extracting peak and valley intensities from the intensity evolution gradient tracking results to generate an intensity evolution fingerprint. , in, For strength evolution fingerprint, Peak intensity For valley strength, 0.1 ensures that the XOR operation object is an integer; finally, a composite index label containing the association dimension hash, time period encoding and strength evolution fingerprint is obtained.

[0008] Furthermore, the backtracking report includes: identifying and labeling three types of deviations: parameter deviation, missing associations, and temporal conflicts; a structured comparison of key dimensions between the current decision and the historical best solution, and dimensions where the deviation exceeds the threshold; the key dimensions include explicit dimensions, implicit dimensions, and feature vector dimensions.

[0009] Furthermore, the memory layer storage includes: acquiring real-time data from at least one external data source, matching and fusing the real-time data with the association dimensions of intent network nodes to generate enhanced data nodes; when an enhanced data node is associated with a high-activity intent, it is stored in the hotspot memory layer and marked as real-time enhanced data; when an enhanced data node is associated with a low-activity intent, it is stored in the lukewarm memory layer and a dynamic verification tag is added; the external data source includes API interfaces, IoT device streams, and user behavior logs.

[0010] Furthermore, the generation of enhanced data nodes includes: performing multi-source real-time data consistency verification before generating enhanced data nodes; uniformly calibrating real-time data from different external data sources to the system's reference time axis; performing feature vector consistency detection based on fused feature vectors and real-time data; and outputting data source confidence.

[0011] Furthermore, the optimization decision includes: when the annotation parameters of the backtracking report deviate, calling the historical best solution parameters stored in the hotspot memory layer to perform parameter calibration; when the annotation association is missing, activating the association dimension graph of the intent network node to perform association completion; when the annotation time sequence conflicts, calling the intensity evolution fingerprint to perform time sequence alignment verification; and fusing the calibration parameters, the completed association dimension and the alignment verification results to generate the optimization decision.

[0012] Furthermore, storing the optimized decision as a new node in the memory layer includes evaluating the impact of the decision: , in, The distance between feature vectors. In order to optimize the decision feature vector, The feature vector of the historical optimal solution; when When the percentage is ≤10%, it is stored in the hotspot memory layer and marked as a fine-tuning decision; when the percentage is 10% < 10% ≤ 30%, it is stored in the warm point memory layer and marked as an innovation decision; when the percentage is ≤10%, it is stored in the warm point memory layer and marked as an innovation decision. When the percentage is greater than 30%, create a separate memory partition for storage and mark it as a migration decision.

[0013] A human-machine collaborative generative engine optimization system includes: The intent management module is used to parse explicit and implicit intents in user input and build intent network nodes. A multi-dimensional dynamic scoring module is used to calculate the time decay coefficient, intensity evolution gradient, correlation density, and overall activity score. The memory layer storage module is used to perform hierarchical storage operations for the hot spot memory layer and the temperature point memory layer; The cross-session index generation module is used to generate composite index labels that include association dimension hashes, time period codes, and strength evolution fingerprints; The decision deviation detection module is used to identify the deviation between the current decision and the historical pattern and trigger the backtracking process; The backtracking report generation module is used to annotate three types of deviations: parameter deviation, missing correlations, and temporal conflicts. The decision generation module has been optimized to perform parameter calibration, correlation completion, and time-series alignment verification. The decision-making influence assessment module is used to calculate the distance between the feature vectors of optimized decisions and to label the decision type; The closed-loop update module is used to store new nodes and update multi-dimensional dynamic scores.

[0014] One or more technical solutions provided in this application have at least the following technical effects or advantages: By employing multimodal intent collaborative parsing technology, we can simultaneously capture explicit needs and latent intents, constructing intent network nodes with temporal evolution characteristics, thereby achieving a comprehensive understanding of complex user intents. Through a dynamic hierarchical memory mechanism, we can intelligently schedule storage resources based on intent activity, ensuring millisecond-level response for high-value information and automatic degradation of inefficient data. Through cross-session indexing technology, we can establish a long-term intent association network to continuously track user preferences and behavioral patterns. Through multi-dimensional backtracking analysis, we can generate interpretable decision deviation reports, driving precise parameter calibration, completion of association dimensions, and resolution of temporal conflicts. Finally, through quantitative evaluation of decision influence and closed-loop updates, we can achieve the effects of optimized generation quality, improved decision-making efficiency, and reduced operation and maintenance costs. Attached Figure Description

[0015] Figure 1 This is a flowchart of the human-machine collaborative generative engine optimization recommendation method in an embodiment of the present invention; Figure 2 This is a diagram of the human-machine collaborative generative engine optimization system architecture in an embodiment of the present invention. Detailed Implementation

[0016] To facilitate understanding of the present invention, a more complete description of this application will be given below with reference to the accompanying drawings, which illustrate preferred embodiments of the invention. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to enable a more thorough and complete understanding of the disclosure of the present invention.

[0017] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention; the term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0018] Example 1: As Figure 1 As shown, a generative engine optimization recommendation method based on human-machine collaboration is presented.

[0019] S1: The intent management module parses the explicit and implicit intents in the user's input content and constructs intent network nodes. Each node records the generation sequence, intensity evolution, and association dimensions of the intent. The intent management module includes: a multimodal parsing engine and a node generator; Specifically, the system synchronously receives raw data streams of text, speech, and images, establishes a cross-modal spatiotemporal alignment channel to ensure temporal consistency of the input streams; it parses the semantic structure and sentiment of the text stream to generate semantic feature vectors; it separates phonemes and suprasegmental features from the speech stream and quantifies sentiment intensity parameters; and it identifies visual focal regions from the image stream, extracts spatial attention distribution, and constructs a three-dimensional feature matrix. , in, To fuse feature vectors, For semantic features, For emotional intensity, As the visual focus; The system uses millisecond-level timestamps synchronized with the International Atomic Time Standard to record time sequence and capture the trend of intent activity changes, including real-time interaction activity and the average intensity of similar intents in historical benchmark data. It associates historical similar nodes and binds device type, geographical location, and network environment metadata. Based on the time accurate to milliseconds, the first 4 bits of the fusion feature vector hash, the first 2 bits of the device type code, and the first 2 bits of the geographical location grid code, a unique identifier is generated.

[0020] S2: Identify the intent activity of nodes through multi-dimensional dynamic scoring, perform hot spot memory layer storage for high-activity data, and perform warm spot memory layer degradation for low-activity data; The time decay coefficient refers to the degree to which the activity of the intended node decays over time. , in, The time decay coefficient, For the domain attenuation coefficient, The system time is the current time, calibrated every 5 minutes via the NTP protocol. Create a timestamp for the node; The intensity evolution gradient is used to capture the intensity change trend of the intended node: , in, The gradient value, For strength standard deviation, At the current intensity, The intensity is the historical average. The association density is used to measure the connection strength of nodes in the intent network: , in, For correlation density, This represents the number of associated nodes. Cosine similarity of feature vectors As the weight of the associated edge, when explicitly specified by the user, =1, when inferring implicit associations in behavioral patterns in a system. =0.6, when continuing associations from historical sessions, =0.3; The overall activity score is a global activity metric for intent nodes: , in, To score overall activity level, , , For the corresponding weights, and ,default =0.4, =0.3, =0.3, focusing on time decay.

[0021] when A value >0.6 is marked as high activity. A value <0.4 is marked as low activity; according to Based on the assessment results, highly active data is stored in a high-speed storage area, supporting millisecond-level response times. This is prioritized for real-time generation tasks, and hotspot data is refreshed periodically to ensure that frequently accessed data resides in the high-speed area. Low-activity data is migrated to a low-cost storage area, where access latency is higher.

[0022] S3: Generate cross-session index labels based on the correlation dimension between intent network nodes. When the decision deviation detection module detects that the current decision deviates from the historical pattern, it loads the associated historical data based on the index labels; compares the difference between the current decision and the historical optimal solution and marks the key deviation points, generating a backtracking report; The cross-session index label includes: encoding the associated dimensions into a JSON string in a fixed order of explicit dimensions, implicit dimensions, and feature vectors to ensure reproducibility; calculating the hash value of the JSON string and extracting the first 16 characters as a unique identifier; dividing time-segmented blocks according to the node generation time sequence and using the International Atomic Time (TAI) to synchronize the system clock to avoid time zone deviations; and extracting peak and valley intensities from the intensity evolution gradient tracking results to generate an intensity evolution fingerprint. , in, For strength evolution fingerprint, Peak intensity For valley strength, 0.1 ensures that the XOR operation object is an integer; finally, a composite index label containing the association dimension hash, time period encoding and strength evolution fingerprint is obtained.

[0023] Specifically, calculate the feature vector distance threshold: , in, The distance between feature vectors. For the current feature vector, For the feature vector of the historical optimal solution, when When the value is greater than 0.2, deviation detection is triggered.

[0024] Load associated historical data and match it using index tags. Retrieve cross-session related nodes using the association dimension hash in the composite index tags. Filter historical data from the same period based on the time period code. Verify the consistency of intensity changes using intensity evolution fingerprints. Generate a backtracking report.

[0025] The backtracking report includes: identifying and labeling three types of deviations: parameter deviation, missing associations, and temporal conflicts; Specifically, the relative deviation between the current decision and the historical best solution in the explicit dimension is calculated. When the deviation exceeds the neighborhood threshold ±15%, it is marked as parameter deviation and verified based on the feature vector distance. The coverage rate of associated nodes in the implicit dimension is calculated using the association density formula; if the coverage rate is <80%, it is determined to be a missing association. , in, For coverage, The current number of associated nodes, and through Detect feature vector similarity, and add labels when the similarity is <0.7; Extract the strength evolution fingerprint of the current decision and calculate the Hamming distance with the historical best solution; when the distance is ≥3, it is determined to be a temporal conflict, the trend direction is verified, and the sign change of the gradient value is analyzed, such as historical... >0 is on the rise, while currently <0 indicates a downward trend.

[0026] The retrospective report uses a parameter-side table to directly compare the numerical indicators of the current decision with the historical best solution, highlighting numerical deviations through differential calculations. It dynamically generates a network graph to visualize the connections between intent nodes, identifying nodes and edges with missing associations based on association density coverage. An association missing alarm is triggered when the coverage is below 80%. A three-dimensional vector space mapping model is constructed to project fused feature vectors onto a coordinate system, displaying the spatial offset of semantic features, emotional intensity, and visual focus. When a relative deviation of explicit parameters > ±15%, an association coverage < 80%, or a Hamming distance ≥ 3 in the intensity evolution fingerprint is detected, a red highlight is triggered to indicate severe deviation; a yellow highlight is used for critical deviations (within the threshold ±5%).

[0027] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages: This application utilizes multimodal intent collaborative parsing technology to simultaneously capture explicit needs and latent intents, constructing intent network nodes with temporal evolution characteristics to achieve an understanding of complex user intents. Through a dynamic hierarchical memory mechanism, storage resources are intelligently scheduled based on intent activity, ensuring millisecond-level response for high-value information and automatic degradation of inefficient data. A long-term intent association network is established through cross-session indexing technology, enabling continuous tracking of user preferences and behavioral patterns. Interpretable decision deviation reports are generated through multi-dimensional backtracking analysis, driving precise parameter calibration, completion of association dimensions, and resolution of temporal conflicts. Finally, through quantitative evaluation of decision influence and closed-loop updates, the application achieves optimized generation quality, improved decision-making efficiency, and reduced operational costs. This addresses core issues in existing technologies such as one-sided intent cognition, rigid memory mechanisms, and fragmented cross-session decision-making.

[0028] Example 2: In Example 1, only the construction and hierarchical storage of intent network nodes were implemented, but the node data source was singular and lacked external real-time data injection, resulting in insufficient timeliness of the generated content. This example further supplements the content of Example 1.

[0029] The memory layer storage includes: acquiring real-time data from at least one external data source, matching and fusing the real-time data with the association dimensions of intent network nodes to generate enhanced data nodes; when an enhanced data node is associated with a high-activity intent, it is stored in the hotspot memory layer and marked as real-time enhanced data; when an enhanced data node is associated with a low-activity intent, it is stored in the warm spot memory layer and a dynamic verification tag is added; the external data sources include API interfaces, IoT device streams, and user behavior logs.

[0030] Specifically, dynamic information flows are obtained through API interface data, IoT device streams, and user behavior logs. During the data access phase, spatiotemporal benchmark calibration is performed to unify heterogeneous timestamps to the TAI atomic time base and convert geographic coordinates to the WGS84 standard, ensuring spatiotemporal consistency of multi-source data. When the cosine similarity between API keywords and semantic feature vectors is greater than 0.7, new semantic features are generated using weighted averages. The emotional intensity Qreal of behavior logs and historical Q values ​​are dynamically weighted to maximize the capture of emotional context. The geometric convex hull center is calculated from IoT device coordinates and visual focus to generate enhanced spatial features. Through explicit dimension fusion, implicit dimension enhancement, and spatial dimension integration, enhanced data nodes are output, inheriting the correlation dimension map and intensity evolution benchmark of the original nodes, forming fused nodes that combine historical continuity and real-time context.

[0031] When the overall activity score of the association intent is >0.6 and the data confidence is ≥0.9, it is stored in a non-volatile memory array and retrieved in milliseconds using a B+ tree index; it is marked as real-time augmented data, and a data source fingerprint and time stamp are added; it is given priority in participating in optimization decision generation. For association nodes with an overall activity score <0.4, feature compression is performed, reducing the 128-dimensional feature vector to 16 dimensions to retain core information; the association graph is pruned, and weights are removed. Edge connectivity with a resolution of <0.3; dynamic verification tags are added, including node ID, verification timestamp, and data traceability fingerprint; stored in columnar compressed cloud storage, verification is performed periodically, and a backtracking process is triggered when failure occurs.

[0032] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages: This application addresses the core shortcomings of single-source node data and insufficient timeliness of generated content by introducing a multi-source real-time data fusion mechanism. It solves the problem of weak contextual relevance by collaboratively accessing three real-time data channels: API interface, IoT device stream, and user behavior logs, combined with TAI atomic time base spatiotemporal calibration and WGS84 spatial coordinate normalization. A three-dimensional intelligent matching rule is adopted to generate enhanced data nodes that combine historical continuity with real-time context. High-activity nodes are stored in non-volatile memory and retrieved via B+ tree millisecond retrieval, marked as real-time enhanced data to directly participate in decision optimization. Low-activity nodes undergo feature dimensionality reduction, association pruning, and the addition of dynamic verification labels, reducing storage resource utilization while ensuring verification accuracy.

[0033] Example 3: Although Example 2 achieves the fusion of multi-source real-time data and intent nodes, there may be contradictions in the data from different sources, and the external data feature distribution offset is not detected, which reduces the reliability of the enhanced node. This example further supplements the content of Example 2.

[0034] The generation of enhanced data nodes includes: performing multi-source real-time data consistency verification before generating enhanced data nodes; uniformly calibrating real-time data from different external data sources to the system's reference time axis; and performing feature vector consistency detection based on fused feature vectors and real-time data to output data source confidence.

[0035] Specifically, a three-level time calibration channel is established to address the spatiotemporal asynchrony issue of multi-source data. This involves millisecond-level synchronization with International Atomic Time (TAI) via the NTP protocol, calibrating the system clock every 100 milliseconds to eliminate fundamental clock drift between devices; adding timestamp offset tags to delayed data to compensate for transmission latency; and setting a 200-millisecond dynamic window alignment between API interfaces and user behavior log data to absorb timing jitter caused by network fluctuations. Heterogeneous geographic coordinates are uniformly converted to the WGS84 global standard coordinate system, and visual focus is mapped to a 0.0-1.0 normalized screen space to ensure spatial consistency across devices.

[0036] Perform feature vector consistency testing, comparing the contextual relevance of keywords returned by the API with the semantic vectors of intent nodes. Semantic consistency is considered satisfactory when the text topic matching degree is ≥70% and the keyword overlap rate is >60%. Analyze the fluctuation tolerance of behavioral log sentiment parameters and historical sentiment parameters. Then, the real-time emotional intensity is accepted, among which, For log sentiment parameters, For historical sentiment parameters, The historical standard deviation is used. The spatial intersection-union ratio (IUU) between the coordinates of IoT devices and the visual focus of nodes is calculated. When the overlap area is ≥60%, the spatial association is considered valid.

[0037] A weighted scoring model is used to output the confidence level of the data source. The combined score of topic relevance and keyword coverage accounts for 40% of the score, the deviation between the real-time value and the historical benchmark accounts for 30%, and the score based on the coordinate overlap ratio accounts for 30%. When the data source confidence level is ≥90, it is considered high-confidence data, and enhanced nodes are generated directly. When the data source confidence level is 70≤data source confidence level<90, it is considered medium-confidence data, and historical data should be the primary source for fusion. When the data source confidence level is <70, it is considered low-confidence data, triggering a backtracking report, marking the data source conflict and deviation type, and storing the conflicting data in an independent audit partition for manual review and retrieval.

[0038] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages: This application addresses the conflicts and feature drift issues in the fusion of heterogeneous multi-source data by introducing a multi-source real-time data consistency verification and feature vector detection mechanism. A unified spatiotemporal benchmark calibration method is employed to synchronize real-time data from different sources to the system's benchmark timeline, eliminating data inconsistencies caused by asynchronous timing. By constructing a feature vector consistency detection model based on multi-dimensional weighted scoring, accurate quantitative assessment of data source confidence is achieved. This improves the generation quality of enhanced data nodes and system reliability, ensuring consistency and stability during multi-source data fusion, while optimizing system decision-making efficiency and resource utilization, providing reliable technical support for collaborative data processing in complex environments.

[0039] Example 4: Although Example 3 established a multi-source data verification and confidence leveling mechanism, it did not define how to generate executable optimization decisions based on the verification results, nor did it clarify the inheritance and evolution relationship between new decisions and historical experience. This example further supplements the content of Example 3.

[0040] S4: Generate optimization decisions based on key deviations in the backtracking report, store the optimization decisions as new nodes in the memory layer, and inherit the generation sequence, intensity evolution baseline value and correlation dimension of the original node. Update the multi-dimensional dynamic score based on the decision effect.

[0041] The optimization decision includes: when the annotation parameters in the backtracking report deviate, calling the historical best solution parameters stored in the hotspot memory layer to perform parameter calibration; when the annotation association is missing, activating the association dimension graph of the intent network node to perform association completion; when the annotation time sequence conflicts, calling the intensity evolution fingerprint to perform time sequence alignment verification; and fusing the calibration parameters, completed association dimensions and alignment verification results to generate the optimization decision.

[0042] Specifically, the optimization decision is triggered by the annotation type of the backtracking report. When parameter deviations are found in the backtracking report annotations, the system immediately calls the hotspot memory layer. This memory layer dynamically stores and maintains a set of parameters for solutions that have been verified as historically optimal in different scenarios. Based on the severity of the current deviation, the historically optimal parameters are weighted and averaged with the current parameters to generate a new parameter value that incorporates both historical success and current context. When the annotation indicates a missing correlation, the system activates the correlation dimension graph of the corresponding node in the intent network. This graph is a dynamic network that reveals deep, explicit, or implicit correlations between different decision dimensions. The system traverses other dimension nodes strongly correlated with the current decision node in the graph to retrieve the missing correlation dimensions. Subsequently, it initiates a real-time data request or calls a backup data source to complete the data and information under these missing dimensions, forming a more comprehensive and three-dimensional decision context environment. When a temporal conflict is identified by annotation, the system invokes the intensity evolution fingerprint database. This database records the correct patterns of intensity evolution over time in historical decisions in an abstract manner. The intensity evolution trajectory of the current decision is aligned and verified with the baseline pattern in the fingerprint database. By calculating the intensity difference or pattern similarity between the two at key time points, the specific stage at which the conflict occurred is located. The verification results are used to correct the temporal logic of decision execution, such as adjusting the start timing or duration of different sub-tasks, to ensure that the rhythm of the entire decision-making process is consistent with historical successful patterns. The calibrated parameters, the completed correlation dimension information, and the conclusions of the temporal alignment verification are integrated to generate a new, comprehensive optimized decision-making scheme. This scheme is an enhanced decision-making system that inherits historical experience and specifically corrects current shortcomings.

[0043] The generated optimization decision will be stored as a new node in the system's memory layer. This new node does not exist independently; it inherits the key attributes of the original problem node to maintain the continuity and traceability of the knowledge system. These attributes include: retaining the timestamp and contextual order information of the original decision; inheriting the baseline of the original node's intensity change pattern as a reference for future evaluation of its own or other nodes' intensity evolution; and inheriting all the related dimensions of the original node, which together with the supplemented new dimensions form a richer network of connections. Through this inheritance mechanism, new and old knowledge form a clear genealogical relationship, facilitating historical and relational analysis by the system.

[0044] After the optimized decision is applied to the actual environment, its effect will be continuously monitored. Based on the monitored effect data, the multi-dimensional dynamic score of the new node will be dynamically updated. This score is a comprehensive evaluation system, which typically includes, but is not limited to: timeliness score, accuracy score, robustness score, and relevance value score.

[0045] The optimization decision, stored as a new node in the memory layer, includes an evaluation of the decision's impact: , in, The distance between feature vectors. In order to optimize the decision feature vector, The feature vector of the historical optimal solution; when When the percentage is ≤10%, it is stored in the hotspot memory layer and marked as a fine-tuning decision; when the percentage is 10% < 10% ≤ 30%, it is stored in the warm point memory layer and marked as an innovation decision; when the percentage is ≤10%, it is stored in the warm point memory layer and marked as an innovation decision. When the percentage is greater than 30%, create a separate memory partition for storage and mark it as a migration decision.

[0046] Specifically, the scientific classification of decision innovation is achieved by calculating the relative Euclidean distance between the optimized decision feature vector and the historical best solution feature vector. This assessment is automatically triggered before decision storage, forming a three-level classification storage strategy: when the difference is ≤10%, it is classified as a fine-tuning decision, stored in the hot spot memory layer to achieve millisecond-level response, inheriting the complete association dimension map of the original node, suitable for scenarios such as parameter calibration and time-series fine-tuning; when the difference is in the 10%-30% range, it is defined as an innovative decision, stored in the warm spot layer after feature dimensionality reduction and association pruning, supporting cross-dimensional recombination and mesoscale pattern updates; when the difference is >30%, it is marked as a migration decision, built in an isolated storage environment in an independent partition, fully recording the breakthrough decision paradigm and cross-domain association network, avoiding interference with the existing knowledge system. Fine-tuning decisions are directly embedded in the real-time generation link, and high-frequency call efficiency is ensured through B+ tree indexing; innovative decisions are added with spatiotemporal verification tags to periodically verify data timeliness; migration decisions generate cross-session index snapshots and trigger a new decision tree growth engine. All new nodes strictly inherit the core attributes of the original nodes, including their generation sequence, intensity evolution baseline value, and correlation dimensions, maintaining the traceability of the decision-making genealogy. Simultaneously, the application effect of decisions is fed back to the multi-dimensional dynamic scoring module in real time, driving dynamic adjustments to the storage hierarchy. When the overall activity score continues to rise, innovative decisions can be upgraded to the hotspot layer; if they are not used for a long period, they are automatically downgraded and their resources reclaimed.

[0047] Based on the aforementioned patented method, this application also provides a human-machine collaborative generative engine optimization system, such as... Figure 2 As shown, the system includes: The intent management module is used to parse explicit and implicit intents in user input and build intent network nodes. A multi-dimensional dynamic scoring module is used to calculate the time decay coefficient, intensity evolution gradient, correlation density, and overall activity score. The memory layer storage module is used to perform hierarchical storage operations for the hot spot memory layer and the temperature point memory layer; The cross-session index generation module is used to generate composite index labels that include association dimension hashes, time period codes, and strength evolution fingerprints; The decision deviation detection module is used to identify the deviation between the current decision and the historical pattern and trigger the backtracking process; The backtracking report generation module is used to annotate three types of deviations: parameter deviation, missing correlations, and temporal conflicts. The decision generation module has been optimized to perform parameter calibration, correlation completion, and time-series alignment verification. The decision-making influence assessment module is used to calculate the distance between the feature vectors of optimized decisions and to label the decision type; The closed-loop update module is used to store new nodes and update multi-dimensional dynamic scores.

[0048] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages: This application achieves advancements in decision knowledge accumulation for generative engines by constructing an optimized decision influence assessment and dynamic storage mechanism. It addresses the core issue of the difficulty in objectively measuring the degree of decision innovation using a feature vector distance quantification model, precisely classifying optimized decisions into three categories: fine-tuning, innovation, and migration. This enables the system to identify the full spectrum of decision value, from parameter fine-tuning to paradigm breakthroughs. A hierarchical storage strategy enables intelligent scheduling of knowledge resources. Fine-tuning decisions reside in the hotspot memory layer to ensure millisecond-level response efficiency; innovation decisions are stored in the warm-point layer after feature dimensionality reduction to balance storage costs; and migration decisions are stored in an isolated, pollution-resistant environment in an independent partition, ensuring the real-time availability of frequently accessed core knowledge while providing a secure evolution sandbox for breakthrough solutions. A closed-loop knowledge evolution system is established, inheriting the generation sequence, intensity evolution benchmark value, and core attributes of the original nodes' association dimensions to ensure a traceable phylogenetic relationship between new and old decisions. Combined with a multi-dimensional dynamic scoring and update mechanism driven by decision effect feedback, the storage strategy can dynamically adjust with changes in decision value, achieving self-optimization of the knowledge system. It solves the bottleneck problems of rigid knowledge accumulation in traditional system decision-making, delayed identification of breakthrough solutions, and improper allocation of storage resources.

[0049] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for human-in-the-loop generative engine optimization recommendation, the method comprising: Comprise: S1: resolve explicit and implicit intent in user input content through intent management module, build intent network node, each node records the generation time sequence, intensity evolution and correlation dimension; S2: identify the intent activity of the node by multi-dimensional dynamic scoring, store the hot memory layer for high activity data, and degrade the warm memory layer for low activity data; S3: generate cross-session index label according to the correlation dimension between the intent network nodes, load the associated historical data based on the index label when the decision deviation detection module detects that the current decision deviates from the historical mode, compare the differences between the current decision and the historical optimal scheme and mark the key deviation points to generate a backtracking report; S4: generate an optimized decision according to the key deviation points of the backtracking report, store the optimized decision as a new node in the memory layer, and inherit the generation time sequence, intensity evolution benchmark value and correlation dimension of the original node, and update the multi-dimensional dynamic score according to the decision effect.

2. The human-in-the-loop, generative engine optimization recommendation method of claim 1, wherein, The intent management module comprises a multi-modal analysis engine and a node generator; The multi-modal analysis engine is used to process text, voice and image input streams, and output a fusion feature vector containing semantic features, emotional intensity and visual focus: ， wherein, is a fusion feature vector, is a semantic feature, is an emotional intensity, is a visual focus; The node generator is used to build an intent network node, which records the generation time sequence, intensity evolution and correlation dimension, and generates a unique node identifier.

3. The human-in-the-loop, generative engine optimization recommendation method of claim 1, wherein, The multi-dimensional dynamic score includes a time decay coefficient, an intensity evolution gradient, a correlation density and a comprehensive activity score; The time decay coefficient is the degree of activity decay of the intent node over time: ， wherein, is a time decay coefficient, is a domain decay coefficient, is the current time of the system, is a node creation timestamp; The intensity evolution gradient is used to capture the intensity change trend of the intent node: ， wherein, is a gradient value, is a standard deviation of intensity, is a current intensity, is a historical average intensity; The correlation density is used to measure the connection strength of the node in the intent network: ， wherein, is the correlation density, is the number of correlation nodes, is the correlation edge weight, is the cosine similarity of the feature vectors; The comprehensive activity score is a global activity indicator of the intent node: , wherein, is a comprehensive activity score, , , is a corresponding weight, and .

4. The human-in-the-loop, generative engine optimization recommendation method of claim 1, wherein, The cross-session index label includes extracting the correlation dimension of the intent network node, generating a SHA-256 digest value, dividing the time period block according to the node generation time sequence, extracting the peak and valley intensity, and generating an intensity evolution fingerprint: ， wherein, is the intensity evolution fingerprint, is the peak intensity, is the valley intensity, 0.1 ensures that the exclusive OR operation object is an integer; finally, the composite index tag containing the association dimension hash, time period encoding, and intensity evolution fingerprint is obtained.

5. The human-in-the-loop, generative engine optimization recommendation method of claim 1, wherein, The backtracking report includes identifying and marking three types of deviation: parameter deviation, correlation missing and time sequence conflict; Structured comparison of key dimensions between current decision and historical optimal scheme, highlight the relative deviation threshold value of the dimension; The key dimension includes explicit dimension, implicit dimension and feature vector dimension.

6. The human-in-the-loop, generative engine optimization recommendation method of claim 1, wherein, The memory layer storage includes obtaining real-time data from at least one external data source, matching and fusing the real-time data with the correlation dimension of the intent network node to generate an enhanced data node; When the enhanced data node is associated with high activity intent, store it in the hot memory layer and mark it as real-time enhanced data; When the enhanced data node is associated with low activity intent, store it in the warm memory layer and add a dynamic check tag; The external data source includes API interface, Internet of Things device stream and user behavior log.

7. The human-in-the-loop, generative engine optimization recommendation method of claim 6, wherein, The generating the enhanced data node comprises: before generating the enhanced data node, performing multi-source real-time data consistency verification, uniformly calibrating real-time data from different external data sources to a system reference time axis; performing feature vector consistency detection based on the fusion feature vector and the real-time data, and outputting data source confidence.

8. The human-in-the-loop, generative engine optimization recommendation method of claim 1, wherein, The optimization decision comprises: when the backtracking report label parameter deviates, calling the historical optimal scheme parameter stored in the hot spot memory layer to perform parameter calibration; when the label association is missing, activating the association dimension graph of the intent network node to perform association completion; when the label time sequence conflicts, calling the intensity evolution fingerprint to perform time sequence alignment verification; and fusing the calibration parameters, the completed association dimensions and the alignment verification results to generate the optimization decision.

9. The human-in-the-loop, generative engine optimization recommendation method of claim 1, wherein, The optimization decision is stored in the memory layer as a new node, which comprises evaluating the decision influence: ， wherein, is the feature vector distance, is the optimized decision feature vector, is the historical optimal solution feature vector; when ≤ 10% is stored in the hot memory layer, marked as fine-tuning decision; when 10% < 10 ≤ 30% is stored in the warm memory layer, marked as innovative decision; when > 30% is stored in the independent memory partition, marked as migration decision.

10. A human-in-the-loop generative engine optimization system applied to the human-in-the-loop generative engine optimization recommendation method of any one of claims 1 to 9, characterized in that, The system comprises: An intent management module for parsing explicit and implicit intents in user input content and constructing an intent network node; A multi-dimensional dynamic scoring module for calculating time decay coefficients, intensity evolution gradients, association densities and comprehensive activity scores; A memory layer storage module for performing hierarchical storage operations of hot spot memory layers and warm spot memory layers; A cross-session index generation module for generating a composite index label comprising association dimension hash, time period encoding and intensity evolution fingerprint; A decision deviation detection module for identifying deviations of the current decision from historical patterns and triggering a backtracking process; A backtracking report generation module for labeling three types of deviation: parameter deviation, association missing and time sequence conflict; An optimization decision generation module for performing parameter calibration, association completion and time sequence alignment verification; A decision influence evaluation module for calculating the feature vector distance of the optimization decision and marking the decision type; A closed-loop update module for storing new nodes and updating multi-dimensional dynamic scores.

Images