GEO mobile domain name AI priority reference optimization method and system
By constructing a GEO reputation weight matrix through multi-layer semantic parsing and compressed encoding of mobile top-level domains, and combining user query intent and terminal environment, the problem of domain name reference results deviation in existing technologies is solved, achieving more accurate and timely service adaptation, and improving the reference accuracy of the generative engine and the accessibility of mobile services.
Patent Information
- Application Number
- CN202611408578.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-11
- Publication Date
- 2026-11-10
AI Technical Summary
Existing technologies have shortcomings in handling the semantic representation of mobile top-level domains, dynamic reputation assessment, and generative engine reference adaptation, resulting in reference results that deviate from the actual service experience, making it difficult to accurately understand the deep connection between the domain name and the user's intent, and mobile adaptation compatibility and reputation correction cannot be updated in a timely manner.
By performing multi-layer semantic parsing on the original business data of the mobile top-level domain, a structured data packet is formed. Then, through metadata compression encoding oriented towards the large model context window, a GEO reputation weight matrix is constructed. User query requests from the generative AI engine are received and semantically encoded to calculate the geospatial matching degree and entity type fit. Combined with the user terminal environment fingerprint, adaptive routing and reputation correction are performed to achieve iterative updates of the reputation weight vector.
It improves the accuracy of the generative engine in reading and referencing domain name data, ensuring that geographically relevant, type-matched, and reputable results are presented first, reducing mismatches and irrelevant references, and improving the accuracy and timeliness of service acquisition.
Smart Images

Figure CN123075010A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of domain name optimization technology, and in particular to a method and system for optimizing GEO mobile domain name AI priority reference. Background Technology
[0002] Generative AI engines and search-enhanced generation technologies are widely used. As mobile top-level domains serve as entry points to the mobile internet, the priority access of their content significantly impacts user service efficiency. Current practices often treat mobile top-level domain pages as ordinary web documents, building indexes through web crawling, keyword extraction, structured annotation, and text vector embedding, and then retrieving and ranking them based on semantic similarity, link authority, historical click-through rates, or domain weight. A unified mobile entry point typically relies on user agent identification, redirection, or static adaptation tables for navigation.
[0003] The aforementioned processing method lacks specialized semantic processing for the context window of generative engines. Direct slicing or truncation can easily lead to the loss of key semantics such as entity attributes, geospatial information, and business types, making it difficult for generative engines to accurately understand the deep relationship between domain names and user intent. Reputation evaluation relies heavily on static links or historical statistics, which are difficult to change in real time with user interaction. High-quality emerging domain names are easily underestimated, and high-authority domain names may receive inappropriate priority references.
[0004] The search enhancement generation stage often relies primarily on general semantic similarity for filtering, neglecting geospatial matching and entity type fit. This can easily lead to the return of domains with location discrepancies or service type mismatches. Mobile adaptation depends on preset rules, and it's difficult to effectively address differences in terminal environments, compatibility with unified entry points, and actual interaction behaviors. As a result, reputation weights cannot be corrected in a timely manner, leading to a deviation between the cited results and the actual service experience.
[0005] Existing technologies have significant limitations in handling the semantic representation of mobile top-level domains, dynamic reputation assessment, and generative engine reference adaptation. More refined optimization mechanisms are needed to improve reference accuracy and mobile service accessibility. Summary of the Invention
[0006] This invention provides a method and system for optimizing GEO mobile domain name AI priority citation, which can solve the problems in the prior art.
[0007] A first aspect of this invention provides a method for optimizing the AI-priority citation of GEO mobile domain names, comprising:
[0008] Multi-layer semantic parsing is performed on the original business data of the mobile top-level domain. Structured data packets are determined through entity recognition and attribute extraction. Compressed encoded data of each mobile top-level domain is obtained through metadata compression encoding oriented towards a large model context window.
[0009] Based on compressed encoded data, a GEO reputation weight matrix is constructed through vectorized modeling, and the reputation weight vector corresponding to each mobile top-level domain is written into it.
[0010] The system receives user query requests from the generative AI engine and performs semantic encoding to obtain intent feature vectors. During the retrieval enhancement generation stage of the generative AI engine, it performs quality assessment on candidate mobile top-level domains based on the intent feature vectors to obtain a set of candidate domains.
[0011] The intent feature vector and the reputation weight vector of each domain in the candidate domain set are analyzed, the geospatial matching degree and entity type fit degree are calculated, and the GEO trust score is obtained. The mobile top-level domains that reach the preset score threshold and their unified application entry are pushed to the generative AI engine.
[0012] It receives access requests triggered by users through a generative AI engine, constructs a compatibility mapping relationship between the user's terminal environment fingerprint and the unified application entry, completes adaptive routing, captures the interaction process, constructs interaction behavior signals and converts them into reputation correction quantities, and writes back the reputation weight vector to drive iterative updates.
[0013] In one optional embodiment, multi-layer semantic parsing is performed on the original business data of the mobile top-level domain, structured data packets are determined through entity recognition and attribute extraction, and compressed encoding of metadata oriented towards a large model context window is used to obtain the compressed encoded data of each mobile top-level domain, including:
[0014] Multi-layer semantic parsing is performed on the original business data of the mobile top-level domain. The first semantic layer extracts the business entity type and industry affiliation pointed to by the domain. The second semantic layer extracts the service scope and geographical coverage dimension of the business entity. The third semantic layer extracts the mapping relationship between the business entity and the user demand category. The extraction results of each semantic layer are organized into a structured data packet containing entity identifiers, attribute key-value pairs and inter-layer reference indexes through entity recognition and attribute extraction.
[0015] The structured data packet is subjected to metadata compression encoding oriented towards the large model context window. By constructing a semantic dependency graph, entity identifiers and attribute key-value pairs with co-occurrence relationships in the structured data packet are grouped into semantic unit nodes. The importance weight of the nodes is calculated based on the reference frequency and inter-layer reference index between semantic unit nodes. The semantic unit nodes are prioritized according to the importance weight of the nodes. High-priority semantic unit nodes are extracted according to the capacity constraint of the large model context window and encoded into compressed data.
[0016] The compressed encoded data is bound and stored with the corresponding mobile top-level domain identifier to obtain the compressed encoded data of each mobile top-level domain.
[0017] In one optional embodiment, based on compressed encoded data, a GEO reputation weight matrix is constructed through vectorized modeling, and the reputation weight vector corresponding to each mobile top-level domain is written in, including:
[0018] Extract semantic unit nodes and node importance weights from the compressed encoded data of each mobile top-level domain. Transform the attribution relationship between entity identifiers and business entity types in semantic unit nodes into type vector components. Transform the service scope and geographic coverage dimension in attribute key-value pairs into spatial vector components. Transform the mapping relationship strength in inter-layer reference index into association vector components. Combine type vector components, spatial vector components and association vector components to form an initial feature vector.
[0019] The initial feature vector is subjected to weighted fusion based on node importance weights. By using the node importance weights as multiplicative factors on each vector component derived from the corresponding semantic unit node, a weighted feature vector is obtained. The user access frequency and single access duration of the mobile top-level domain within a preset time window are collected, and the joint distribution entropy of access frequency and duration is calculated to obtain the domain activity stability metric. The domain activity stability metric and the weighted feature vector are subjected to a component-wise exponential mapping transformation to obtain the reputation weight vector of each mobile top-level domain.
[0020] Construct a GEO reputation weight matrix with the mobile top-level domain identifier as the row dimension and the vector component index as the column dimension, and write each reputation weight vector into the GEO reputation weight matrix according to the row dimension.
[0021] In one optional embodiment, a user query request from a generative AI engine is received and semantically encoded to obtain an intent feature vector. During the retrieval enhancement generation stage of the generative AI engine, a quality assessment is performed on candidate mobile top-level domains based on the intent feature vector, resulting in a candidate domain set including:
[0022] The system receives user query requests from the generative AI engine, performs query intent decomposition on the user query requests to obtain core demand descriptions and auxiliary constraints, performs semantic hierarchical parsing on the core demand descriptions to identify target entity types and expected service categories in the core demand descriptions, performs geographic semantic extraction on the auxiliary constraints to identify geographic location limitations and coverage limitations in the auxiliary constraints, and performs semantic encoding on the target entity types, expected service categories, geographic location limitations, and coverage limitations to obtain intent feature vectors.
[0023] In the retrieval enhancement generation stage of the generative AI engine, the reputation weight vector of each mobile top-level domain is extracted from the GEO reputation weight matrix. The semantic space distance between the intent feature vector and each reputation weight vector is calculated. The semantic space distance and the spatial vector component in the reputation weight vector of the corresponding mobile top-level domain are used to perform a collaborative matching degree calculation to obtain the intent matching degree score of each mobile top-level domain. Based on the intent matching degree score, the candidate mobile top-level domains are evaluated for quality. Mobile top-level domains that meet the quality admission criteria are selected to obtain a set of candidate domains.
[0024] In one optional embodiment, parsing the intent feature vector and the reputation weight vector of each domain within the candidate domain set, calculating the geospatial matching degree and entity type fit degree, obtaining the GEO trust score, and pushing the mobile top-level domains that reach the preset score threshold and their unified application entry to the generative AI engine includes:
[0025] The reputation weight vector corresponding to each mobile top-level domain is extracted from the candidate domain set. The geographical location limitation and coverage limitation in the intent feature vector are analyzed. The geographical center point and geographical influence radius corresponding to the geographical location limitation are constructed. The spatial vector components in each reputation weight vector are analyzed. The service range and geographical coverage dimension encoded in the spatial vector components are extracted and transformed into the domain name geographical service boundary. The degree of spatial overlap between the geographical influence radius and the geographical service boundary of each domain name is determined to obtain the geospatial matching degree.
[0026] Extract the target entity type and expected service category from the intent feature vector, decode the business entity type attribution relationship in the type vector component of each reputation weight vector, determine the type hierarchy inclusion relationship between the target entity type and the business entity type attribution relationship, obtain the entity type fit, perform a two-dimensional collaborative fusion calculation on the geospatial matching degree and the entity type fit, and obtain the GEO trust score of each mobile top-level domain.
[0027] Select mobile top-level domains that have reached a preset score threshold in GEO trust score, obtain the unified application entry corresponding to the mobile top-level domain from the GEO mobile domain unified application entry database, and push the mobile top-level domain and its unified application entry to the generative AI engine.
[0028] In one optional embodiment, receiving an access request triggered by a user via a generative AI engine, constructing a compatibility mapping relationship between the user's terminal environment fingerprint and the unified application entry point, completing adaptive routing, capturing interaction process signals to construct interaction behavior signals and converting them into reputation correction quantities, and writing back the reputation weight vector to drive iterative updates include:
[0029] Receive access requests triggered by users via the generative AI engine, extract the target mobile phone top-level domain name and user terminal identifier carried in the access request, parse the hardware capability parameters and software environment parameters in the user terminal identifier, and construct the user terminal environment fingerprint.
[0030] Extract multiple candidate routing channels and their corresponding routing features from the unified application entry point, perform vectorized matching between the user terminal environment fingerprint and the routing features of each candidate routing channel, construct a compatibility mapping relationship, and complete adaptive routing based on vector similarity.
[0031] Monitor the user's interaction process at the target address, capture the session duration and function call integrity identifier, determine the user stickiness based on the session duration, determine the business loop completion based on the function call integrity identifier, construct the user stickiness and business loop completion into interaction behavior signals, analyze the signal polarity and signal amplitude of the interaction behavior signals, and convert them into reputation correction amount for the target mobile top-level domain.
[0032] Locate the reputation weight vector in the GEO reputation weight matrix that corresponds to the top-level domain of the target mobile phone, write the reputation correction amount back to the reputation weight vector, and drive the iterative update of the reputation weight vector.
[0033] In one optional embodiment, multiple candidate routing channels and their corresponding routing features are extracted from the unified application entry point. The user terminal environment fingerprint is then vectorized and matched with the routing features of each candidate routing channel to construct a compatibility mapping relationship. Adaptive routing is then performed based on vector similarity, including:
[0034] Obtain the unified application entry corresponding to the top-level domain of the target mobile phone, extract multiple candidate routing channels configured by the unified application entry, parse the resource consumption characteristics and response performance characteristics of each candidate routing channel, vectorize the resource consumption characteristics and response performance characteristics, and generate routing feature vectors for each candidate routing channel.
[0035] The terminal resource status and network connection status in the user terminal environment fingerprint are analyzed, and the terminal resource status and network connection status are vectorized to generate user terminal feature vectors.
[0036] Calculate the vector similarity between the user terminal feature vector and the routing feature vector of each candidate routing channel, and construct a compatibility mapping relationship between each candidate routing channel and its corresponding vector similarity.
[0037] The system selects the candidate routing channel with the highest routing priority weight from the compatibility mapping relationship and routes the access request to the routing target address corresponding to the candidate routing channel. When routing fails, the system selects the next best candidate routing channel in descending order of routing priority weight for rerouting until routing is successful, thus completing adaptive routing.
[0038] A second aspect of this invention provides a GEO mobile domain name AI priority citation optimization system, comprising:
[0039] The semantic parsing unit is used to perform multi-layer semantic parsing on the original business data of the mobile top-level domain. It determines the structured data packet through entity recognition and attribute extraction, and obtains the compressed encoded data of each mobile top-level domain through metadata compression encoding oriented to the large model context window.
[0040] The reputation modeling unit is used to construct the GEO reputation weight matrix based on compressed encoded data through vectorized modeling, and write the reputation weight vector corresponding to each mobile top-level domain.
[0041] The intent evaluation unit is used to receive user query requests from the generative AI engine and perform semantic encoding to obtain intent feature vectors. During the retrieval enhancement generation stage of the generative AI engine, the unit performs quality evaluation on candidate mobile top-level domains based on the intent feature vectors to obtain a set of candidate domains.
[0042] The trust scoring unit is used to analyze the intent feature vector and the reputation weight vector of each domain in the candidate domain set, calculate the geospatial matching degree and entity type fit degree, obtain the GEO trust score, and push the mobile top-level domains and their unified application entry that reach the preset score threshold to the generative AI engine.
[0043] The routing update unit is used to receive access requests triggered by users through the generative AI engine, build a compatibility mapping relationship between the user terminal environment fingerprint and the unified application entry, complete adaptive routing, capture the interaction process, build interaction behavior signals and convert them into reputation correction quantities, and write back the reputation weight vector to drive iterative updates.
[0044] A third aspect of the present invention provides an electronic device, comprising:
[0045] processor;
[0046] Memory used to store processor-executable instructions;
[0047] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.
[0048] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.
[0049] In this embodiment of the invention, multi-layer semantic parsing is performed on the original business data of mobile top-level domains. Structured data packets are formed through entity recognition and attribute extraction, and then compressed and encoded according to a large model context window. This reduces context occupancy, improves information density and semantic fidelity, makes it easier for the generative engine to read, understand, and reference domain data, reduces invalid interference, and improves the quality of data supply generated for retrieval enhancement. Vectorized modeling is performed based on the compressed encoded data to construct a GEO reputation weight matrix, and reputation weight vectors are written for each mobile top-level domain. Domain quality is thus transformed into a calculable, comparable, and ranking quantitative basis. The generative engine can identify high-quality sources, suppress low-quality content, increase the probability of priority citation, and enhance the credibility of the answer during candidate recall and ranking. User queries are received and semantically encoded to obtain intent feature vectors. The quality of candidate mobile top-level domains is evaluated during the retrieval enhancement generation stage to form a candidate domain set. The intent features and reputation weight vectors are parsed to calculate geospatial matching degree and entity type fit, generating a GEO trust score. Domains reaching the threshold and unified application entry points are pushed to the generative engine, prioritizing geographically relevant, type-matched, and reputable results to reduce mismatches and irrelevant references, thus improving service retrieval accuracy. Upon receiving access requests, a compatibility mapping is established between the terminal environment fingerprint and the unified application entry point, enabling adaptive routing and improving the success rate of opening requests across different devices and systems. Interaction behaviors are captured and converted into reputation correction values, written back to the reputation weight vector, forming a feedback loop that continuously updates reputation, ensuring high-quality entry points receive more references, suppressing abnormal behavior, and maintaining the timeliness of results. Attached Figure Description
[0050] Figure 1 A flowchart illustrating the optimization method for prioritizing GEO mobile domain name references using AI.
[0051] Figure 2 A comparison chart of compressed encoded token usage for large model context windows;
[0052] Figure 3 A graph showing the intent matching score and quality admission threshold of candidate mobile top-level domains;
[0053] Figure 4 Flowchart of adaptive routing and reputation update logic. Detailed Implementation
[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.
[0055] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.
[0056] Figure 1 This is a flowchart illustrating the GEO mobile domain name AI priority reference optimization method according to an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes:
[0057] Multi-layer semantic parsing is performed on the original business data of the mobile top-level domain. Structured data packets are determined through entity recognition and attribute extraction. Compressed encoded data of each mobile top-level domain is obtained through metadata compression encoding oriented towards a large model context window.
[0058] Based on compressed encoded data, a GEO reputation weight matrix is constructed through vectorized modeling, and the reputation weight vector corresponding to each mobile top-level domain is written into it.
[0059] The system receives user query requests from the generative AI engine and performs semantic encoding to obtain intent feature vectors. During the retrieval enhancement generation stage of the generative AI engine, it performs quality assessment on candidate mobile top-level domains based on the intent feature vectors to obtain a set of candidate domains.
[0060] The intent feature vector and the reputation weight vector of each domain in the candidate domain set are analyzed, the geospatial matching degree and entity type fit degree are calculated, and the GEO trust score is obtained. The mobile top-level domains that reach the preset score threshold and their unified application entry are pushed to the generative AI engine.
[0061] It receives access requests triggered by users through a generative AI engine, constructs a compatibility mapping relationship between the user's terminal environment fingerprint and the unified application entry, completes adaptive routing, captures the interaction process, constructs interaction behavior signals and converts them into reputation correction quantities, and writes back the reputation weight vector to drive iterative updates.
[0062] In one optional embodiment, multi-layer semantic parsing is performed on the original business data of the mobile top-level domain, structured data packets are determined through entity recognition and attribute extraction, and compressed encoding of metadata oriented towards a large model context window is used to obtain the compressed encoded data of each mobile top-level domain, including:
[0063] Multi-layer semantic parsing is performed on the original business data of the mobile top-level domain. The first semantic layer extracts the business entity type and industry affiliation pointed to by the domain. The second semantic layer extracts the service scope and geographical coverage dimension of the business entity. The third semantic layer extracts the mapping relationship between the business entity and the user demand category. The extraction results of each semantic layer are organized into a structured data packet containing entity identifiers, attribute key-value pairs and inter-layer reference indexes through entity recognition and attribute extraction.
[0064] The structured data packet is subjected to metadata compression encoding oriented towards the large model context window. By constructing a semantic dependency graph, entity identifiers and attribute key-value pairs with co-occurrence relationships in the structured data packet are grouped into semantic unit nodes. The importance weight of the nodes is calculated based on the reference frequency and inter-layer reference index between semantic unit nodes. The semantic unit nodes are prioritized according to the importance weight of the nodes. High-priority semantic unit nodes are extracted according to the capacity constraint of the large model context window and encoded into compressed data.
[0065] The compressed encoded data is bound and stored with the corresponding mobile top-level domain identifier to obtain the compressed encoded data of each mobile top-level domain.
[0066] In one specific implementation, when performing multi-layer semantic parsing on the original business data of the mobile top-level domain, the entire parsing process is divided into three independent semantic layers, each responsible for extracting business semantic information from different granularities and dimensions. The original business data typically originates from domain registration information, official website content, app store descriptions, user review summaries, and third-party industry databases. Before entering the semantic parsing process, this data must undergo format normalization processing, uniformly encoded into a UTF-8 format text sequence to ensure consistency across subsequent parsing layers.
[0067] In the first semantic layer, the focus is on extracting the business entity type and industry affiliation of the domain name. Business entity types cover major categories such as e-commerce retail, financial payment, local life services, content media, travel navigation, and healthcare. A pre-trained named entity recognition model scans the original text to identify the main entity name and its industry tag. Industry affiliation determination employs a multi-tag classification strategy, allowing a single domain to have multiple industry tags simultaneously. For example, a mobile top-level domain may cover both food delivery and fresh food delivery; in this case, both industry affiliations are labeled in the first semantic layer. The output of the first semantic layer is stored as key-value pairs, with the keys being "entity_type" and "industry_tag," and the values being a list of corresponding category strings.
[0068] In the second semantic layer, based on the entity type results from the first semantic layer, the service scope and geographic coverage dimension of the business entity are further extracted. The extraction of the service scope focuses on identifying the specific set of functions provided by the domain name, such as whether it supports online booking, instant delivery, intercity services, and international services. The geographic coverage dimension is constructed by parsing the city names, province labels, latitude and longitude coordinates, and administrative division codes in the original business data to establish the geographic service boundary of the domain name. For geographic coverage information, a hierarchical geocoding method is adopted, progressively refining the service area from the national, provincial, and city levels down to the district and county levels, forming a multi-granularity geographic labeling system. The extraction results from the second semantic layer include two types of attributes: "service_scope" and "geo_coverage." "geo_coverage" is recorded in the form of a geocoded list, supporting subsequent geospatial matching calculations.
[0069] In the third semantic layer, the mapping relationship between business entities and user demand categories is extracted. User demand categories are derived from cluster analysis of a large number of historical query logs, forming a standardized demand classification system, such as "nearby restaurant search," "mobile phone top-up," and "local coupon acquisition." The third semantic layer matches the business description of a domain name with each category in the demand classification system through semantic similarity calculation, establishing a demand mapping weight list for each domain name. A higher weight indicates a stronger association between the domain name and the corresponding demand category. The extraction results of this layer are stored as a "demand_mapping" attribute, which is an ordered list of demand category identifiers and matching weights.
[0070] After extracting from the three semantic layers, the results from each layer are organized into structured data packets through entity recognition and attribute extraction. The structured data packets contain three core fields: entity identifier (entity_id), attribute key-value pair set (attribute_pairs), and cross-layer reference index (cross_layer_index). The entity identifier uses a globally unique encoding, generated by concatenating the domain name hash value and a timestamp, ensuring that structured data packets from different domains do not generate identifier conflicts. The attribute key-value pair set summarizes all extraction results from the three semantic layers, while the cross-layer reference index records the dependencies between attributes in different semantic layers. For example, the "geo_coverage" attribute of the second semantic layer references the "industry_tag" attribute of the first semantic layer as a constraint. This reference relationship is explicitly recorded in the structured data packet in the form of index pointers, providing a foundation for the subsequent construction of the semantic dependency graph.
[0071] When performing metadata compression encoding on structured data packets with a large model context window, the core objective is to retain the information with the highest semantic value within the limited context window capacity. When constructing the semantic dependency graph, entity identifiers and attribute key-value pairs in the structured data packets are used as nodes in the graph. Edges between nodes are defined by two types of relationships: co-occurrence relationships, where co-occurrence edges are established between entity identifiers and attribute key-value pairs that appear simultaneously in the same original business text; and inter-layer reference relationships, which are cross-layer dependencies recorded through inter-layer reference indexes. The weight of co-occurrence relationships is obtained by statistically analyzing co-occurrence frequency. Let the node... With nodes The co-occurrence frequency is The weight of the co-occurring edge .
[0072] Based on the constructed semantic dependency graph, the importance weight of each semantic unit node is calculated. The node importance weight considers two factors: the node's reference frequency in the graph (i.e., the sum of its in-degrees) and the hierarchical weight of the semantic layer to which the node belongs. Let the node... The weighted sum of the in-degrees is The hierarchical weight of its semantic layer is Then the node importance weight Among them, the third semantic layer has the highest hierarchical weight because the mapping relationship of user needs is the most critical to the retrieval enhancement generation stage of the generative AI engine; the first semantic layer has the next highest hierarchical weight; and the second semantic layer's hierarchical weight is dynamically adjusted according to the specific geographic coverage granularity.
[0073] Based on node importance weights All semantic unit nodes are sorted in descending order to obtain a priority-sorted list. The capacity constraint of the large model context window is based on the maximum number of tokens. This means that semantic unit nodes are encoded into a token sequence in descending order of priority. When the cumulative number of tokens reaches a certain threshold... The encoding process involves truncation to ensure that the compressed data strictly meets the context window capacity limit. The encoding format uses a compact JSON-LD structure, which serializes entity identifiers, attribute key-value pairs, and their priority weights into a byte stream, and appends a checksum to support integrity verification during subsequent decoding.
[0074] When binding and storing compressed encoded data with the corresponding mobile top-level domain identifier, a key-value storage structure is adopted. The canonical string of the mobile top-level domain (such as ".mobile", ".phone", etc., combined with the complete representation of the second-level domain) is used as the primary key, and the byte stream of compressed encoded data and its metadata (including encoding timestamp, original data version number, and compression ratio statistics) are used as the values, written to the distributed storage engine. After binding and storage are completed, each mobile top-level domain has an independent record of compressed encoded data. Subsequent vectorized modeling and reputation weight matrix construction use this record as the input data source, ensuring the traceability of data origination and the atomicity of update operations.
[0075] like Figure 2 The diagram illustrates the changes in token usage for five mobile top-level domains before and after metadata compression encoding for a large model context window. Light gray bars represent the state before compression, blue bars represent the state after compression, and the red dashed line represents the context window capacity constraint. Before compression, the number of tokens was 3860, 3420, 3980, 2760, and 3240 respectively; after compression, they decreased to 1980, 1760, 2040, 1520, and 1680 respectively, with compression rates ranging from approximately 48.7% to 55.1%. This process utilizes a compression encoding mechanism... Figure 2 All compressed tokens are lower than This indicates that the encoding results meet the large model context window capacity limit, while retaining high-priority semantic units, reducing context usage, improving information density and semantic fidelity, and making it easier for generative AI engines to read and reference domain name data.
[0076] In one optional embodiment, based on compressed encoded data, a GEO reputation weight matrix is constructed through vectorized modeling, and the reputation weight vector corresponding to each mobile top-level domain is written in, including:
[0077] Extract semantic unit nodes and node importance weights from the compressed encoded data of each mobile top-level domain. Transform the attribution relationship between entity identifiers and business entity types in semantic unit nodes into type vector components. Transform the service scope and geographic coverage dimension in attribute key-value pairs into spatial vector components. Transform the mapping relationship strength in inter-layer reference index into association vector components. Combine type vector components, spatial vector components and association vector components to form an initial feature vector.
[0078] The initial feature vector is subjected to weighted fusion based on node importance weights. By using the node importance weights as multiplicative factors on each vector component derived from the corresponding semantic unit node, a weighted feature vector is obtained. The user access frequency and single access duration of the mobile top-level domain within a preset time window are collected, and the joint distribution entropy of access frequency and duration is calculated to obtain the domain activity stability metric. The domain activity stability metric and the weighted feature vector are subjected to a component-wise exponential mapping transformation to obtain the reputation weight vector of each mobile top-level domain.
[0079] Construct a GEO reputation weight matrix with the mobile top-level domain identifier as the row dimension and the vector component index as the column dimension, and write each reputation weight vector into the GEO reputation weight matrix according to the row dimension.
[0080] In one specific implementation, the compressed encoded data contains three core structures: semantic unit nodes and their node importance weights, attribute key-value pairs, and inter-layer reference indexes. For each of these three structures, corresponding vector components are extracted, and then combined and weighted to ultimately form the reputation weight vector for each mobile top-level domain.
[0081] The semantic unit node stores the attribution relationship between entity identifiers and business entity types. Taking the business associated with a mobile top-level domain as an example, its entity types may cover various categories such as communication service providers, application distribution platforms, and mobile payment portals. The attribution relationship between entity identifiers and business entity types is one-hot encoded or mapped by category embedding to obtain type vector components. Its dimensions are aligned with the total number of categories in the predefined business entity type system, enabling the semantic differences of each mobile top-level domain in the entity type dimension to be accurately quantified.
[0082] The attribute key-value pairs contain service scope and geographic coverage dimension, which directly determine the applicability of the mobile top-level domain at the geospatial level. Service scope describes the business boundaries supported by the domain, while geographic coverage dimension records the actual administrative division level or latitude and longitude range it covers. The service scope is mapped to a continuous service breadth component, and the geographic coverage dimension is transformed into discrete or continuous spatial coordinate components using a geocoding scheme (such as H3 hierarchical grid coding or administrative division numbering). The two are then concatenated to form a spatial vector component. The design of spatial vector components allows for direct similarity calculations within the vector space when calculating geospatial matching, avoiding the high overhead of parsing each line of the original geographic string.
[0083] The inter-layer reference index records the mapping relationships and their strength between nodes in different semantic layers. The strength of the mapping relationship reflects the degree of association between the current mobile top-level domain and other domains or business entities; a higher strength value means a deeper coupling between the two in business logic or user behavior paths. The mapping relationship strengths in the inter-layer reference index are normalized and arranged into association vector components. The dimensions of the association vector components correspond to the preset number of association types. , and By concatenating them in a fixed order, we obtain the initial feature vector. Its dimension is the sum of the dimensions of the three types of components.
[0084] The initial feature vectors do not yet distinguish the differences in the contribution of different semantic unit nodes to the overall domain name reputation. Node importance weights. The weighting factor has already been calculated during the compression encoding stage based on the hierarchical weights and in-degree weights of the nodes in the semantic graph; this weight is directly reused here as the multiplicative factor. For the _____, ... The vector components derived from each semantic unit node will Multiply by the type vector component, spatial vector component, and association vector component corresponding to the node respectively, and then sum the weighted components of all nodes in the same dimension to obtain the weighted feature vector. This weighted fusion process ensures that semantic nodes that are more representative of the business logic contribute more to the final vector, thus enabling the reputation weight vector to more accurately reflect the actual business value of the mobile top-level domain.
[0085] Weighted feature vectors constructed solely based on static semantic structures cannot capture the dynamic activity level of domain names in actual operation. Therefore, data is collected from each mobile top-level domain within a preset time window. User access frequency within Duration sequence of a single access ,in This represents the total number of visits within the time window. Treating visit frequency and duration as two-dimensional random variables, we estimate the joint probability distribution by discretizing and binning their joint distribution, and then calculate the joint distribution entropy. :
[0086] ;
[0087] in For the access frequency falling into the first The frequency is divided into bins and the duration falls into the first bin. Joint probability estimates for each time period bin. The lower the value, the more concentrated the access behavior of the domain name, and the more stable the user access pattern. The higher the value, the greater the dispersion of access behavior and the poorer the stability of domain activity. After linear normalization mapping to The range is used to obtain a measure of domain activity and stability. The lower the entropy, the better. The higher the value, the more positively it is associated with reputation.
[0088] Domain activity stability metric with weighted eigenvectors Perform a component-wise exponential mapping transformation, specifically, for Each component Perform the transformation:
[0089] ;
[0090] in For the transformed first Each reputation vector component As a modulation coefficient in the exponential mapping, domains with high activity and stability receive non-linear amplification across all components, while domains with low activity and stability have their component values suppressed after the exponential mapping. All components... The combined result is the reputation weight vector of the mobile phone's top-level domain. Its dimensions and Maintain consistency.
[0091] After calculating the reputation weight vectors for all mobile top-level domains, a GEO reputation weight matrix is constructed, with the mobile top-level domain identifier as the row dimension and the vector component index as the column dimension. Let there be a total of Each mobile top-level domain has a reputation weight vector dimension of [number]. ,but The scale is The reputation weight vector of each domain name The matrix is initialized and populated by writing the corresponding rows according to the row index of the domain name identifier. The GEO reputation weight matrix stores the complete reputation representation of each domain name in rows, which supports the rapid retrieval of the reputation weight vector of the target domain name through matrix row retrieval during the subsequent retrieval enhancement generation stage, and performs batch similarity calculation with the intent feature vector, significantly improving the evaluation efficiency of candidate domain names. The matrix structure also supports incremental row updates. When the reputation correction is written back, only the corresponding row needs to be located and the vector update operation needs to be performed, without rebuilding the entire matrix, ensuring low latency of iterative updates.
[0092] In one optional embodiment, a user query request from a generative AI engine is received and semantically encoded to obtain an intent feature vector. During the retrieval enhancement generation stage of the generative AI engine, a quality assessment is performed on candidate mobile top-level domains based on the intent feature vector, resulting in a candidate domain set including:
[0093] The system receives user query requests from the generative AI engine, performs query intent decomposition on the user query requests to obtain core demand descriptions and auxiliary constraints, performs semantic hierarchical parsing on the core demand descriptions to identify target entity types and expected service categories in the core demand descriptions, performs geographic semantic extraction on the auxiliary constraints to identify geographic location limitations and coverage limitations in the auxiliary constraints, and performs semantic encoding on the target entity types, expected service categories, geographic location limitations, and coverage limitations to obtain intent feature vectors.
[0094] In the retrieval enhancement generation stage of the generative AI engine, the reputation weight vector of each mobile top-level domain is extracted from the GEO reputation weight matrix. The semantic space distance between the intent feature vector and each reputation weight vector is calculated. The semantic space distance and the spatial vector component in the reputation weight vector of the corresponding mobile top-level domain are used to perform a collaborative matching degree calculation to obtain the intent matching degree score of each mobile top-level domain. Based on the intent matching degree score, the candidate mobile top-level domains are evaluated for quality. Mobile top-level domains that meet the quality admission criteria are selected to obtain a set of candidate domains.
[0095] In one specific implementation, after receiving a user query request from the generative AI engine, the primary task is to perform structured query intent decomposition on the request. User query requests are typically presented in the form of natural language phrases or sentences, containing both the user's core demand description and several restrictive conditions. Through intent decomposition, the user query request is broken down into two parts: a core demand description and auxiliary constraints. The core demand description carries the target information that the user truly wants to obtain; for example, in the query "find mobile operators that provide 4G service packages," "mobile operators that provide 4G service packages" is the core demand description. The auxiliary constraints carry the user's limiting requirements on the scope of the results; for example, the geographical limitation "only East China" is an auxiliary constraint. This decomposition method avoids confusing constraint information with core semantics, thereby improving the accuracy of subsequent semantic encoding.
[0096] When performing semantic hierarchical parsing on the core requirement description, a multi-layer semantic parsing framework is used to identify two key semantic elements: target entity type and expected service category. Target entity type refers to the entity classification to which the user expects the query results to belong, such as "mobile phone operator," "mobile phone brand," and "mobile application distribution platform," all of which belong to different granularity target entity types. Expected service category refers to the business capabilities or service attributes that the user hopes the entity will possess, such as "data package sales," "terminal device retail," and "application content distribution." The semantic hierarchical parsing process leverages the named entity recognition and intent classification capabilities of a pre-trained language model to map the core requirement description to a predefined entity type system and service category system, thereby obtaining a structured set of semantic tags.
[0097] When performing geographic semantic extraction on auxiliary constraints, the focus is on identifying two types of geographic semantic elements: geographic location constraints and coverage constraints. Geographic location constraints describe the geographic coordinates or regions where the user expects the target domain name service to be located; coverage constraints describe the user's expectations for the breadth of service coverage, such as "national coverage" or "provincial coverage." The geographic semantic extraction process uses joint reasoning between a geographic entity recognition model and a geographic ontology database to standardize geographic descriptions in natural language into quantifiable geospatial representations, including latitude and longitude coordinate ranges or hierarchical encodings, so as to accurately match them with the spatial vector components in the reputation weight vector.
[0098] By jointly semantically encoding four types of structured semantic elements—target entity type, desired service category, geographic location limitation, and coverage limitation—the intent feature vector is obtained. The encoding process maps the four types of elements to their corresponding semantic subspaces, and then fuses them into a unified dense vector representation through concatenation and linear transformation. Specifically, the sub-vector dimensions corresponding to the target entity type and the desired service category are related to the type vector component in the reputation weight vector. The dimensions remain consistent, with the sub-vector dimensions corresponding to geographical location constraints and coverage constraints matching the spatial vector components in the reputation weight vector. The dimensions are kept consistent, thus ensuring that the intent feature vector and the reputation weight vector are comparable in each subspace.
[0099] After entering the retrieval enhancement generation stage of the generative AI engine, from the GEO reputation weight matrix Extract the reputation weight vector corresponding to each mobile top-level domain one by one. Calculate the intention feature vector With each reputation weight vector The semantic space distance between them. The semantic space distance is calculated using cosine similarity as a metric. Defined as ,in Represents the dot product of two vectors. and Let L2 and L2 represent the intent feature vector and reputation weight vector, respectively. The closer the cosine similarity value is to 1, the closer the intent feature vector and the domain's reputation weight vector are in the overall semantic space, and the smaller the semantic space distance.
[0100] Relying solely on overall semantic space distance for filtering has limitations, as differences in geographic service coverage among different domains are often key factors determining user satisfaction. Therefore, after obtaining the semantic space distance, it is further compared with the spatial vector components of the corresponding mobile top-level domain's reputation weight vector. Perform collaborative matching degree calculation. The collaborative matching degree calculation process extracts the intent feature vector. Subvectors corresponding to geographic information ,calculate and Dot product similarity between ,Right now Ultimately, the intent matching score for each mobile top-level domain is calculated. It is obtained by weighted fusion of overall cosine similarity and geographical collaborative matching degree, and the calculation method is as follows: ,in and These are the overall semantic similarity weight coefficient and the geographical collaborative matching weight coefficient, respectively, and satisfy the following conditions: The values of both are pre-configured based on the importance of geographical factors in the actual business scenario.
[0101] Intent matching score based on each mobile top-level domain A quality assessment is performed on the candidate mobile top-level domains. Quality admission criteria are based on a preset intent matching threshold. Confirmed, only when the intent matching score of a certain mobile top-level domain is... satisfy Only then is the domain name included in the candidate domain name set. Threshold The settings take into account both the semantic clarity of the current query request and the overall distribution characteristics of domain names in the GEO reputation weight matrix. For query requests with highly explicit semantics, the setting can be appropriately increased. The value of is chosen to ensure the accuracy of the candidate domain name set; for query requests with ambiguous semantics or broad coverage, the value should be appropriately reduced. The value of is chosen to ensure that the candidate domain name set has sufficient recall breadth. After quality assessment and screening, the set of mobile top-level domain names that meet the quality admission criteria constitutes the final candidate domain name set. This set will serve as the input basis for subsequent GEO trust score calculation and push decision, supporting the generative AI engine to complete high-quality domain name recommendations in the retrieval enhancement generation stage.
[0102] like Figure 3 As shown, this diagram illustrates the distribution of intent matching scores after quality assessment of candidate mobile top-level domains during the retrieval enhancement generation stage of the generative AI engine. The horizontal axis represents the intent matching score. The red dashed line represents the quality admission threshold. The image The values of 0.88, 0.81, 0.76, 0.69, and 0.64 respectively were all above the threshold and were included in the candidate domain name set; while the others... The scores were 0.58, 0.47, and 0.39 respectively, all below the threshold and therefore rejected. This score is based on the overall cosine similarity. Geographical Coordination Matching Degree The weighted fusion yields, i.e. Take the following from the diagram: , By using threshold filtering, we can ensure that the candidate domain name set has both semantic relevance and geographical coverage matching, thereby improving the accuracy of subsequent GEO trust score calculation and push.
[0103] In one optional embodiment, parsing the intent feature vector and the reputation weight vector of each domain within the candidate domain set, calculating the geospatial matching degree and entity type fit degree, obtaining the GEO trust score, and pushing the mobile top-level domains that reach the preset score threshold and their unified application entry to the generative AI engine includes:
[0104] The reputation weight vector corresponding to each mobile top-level domain is extracted from the candidate domain set. The geographical location limitation and coverage limitation in the intent feature vector are analyzed. The geographical center point and geographical influence radius corresponding to the geographical location limitation are constructed. The spatial vector components in each reputation weight vector are analyzed. The service range and geographical coverage dimension encoded in the spatial vector components are extracted and transformed into the domain name geographical service boundary. The degree of spatial overlap between the geographical influence radius and the geographical service boundary of each domain name is determined to obtain the geospatial matching degree.
[0105] Extract the target entity type and expected service category from the intent feature vector, decode the business entity type attribution relationship in the type vector component of each reputation weight vector, determine the type hierarchy inclusion relationship between the target entity type and the business entity type attribution relationship, obtain the entity type fit, perform a two-dimensional collaborative fusion calculation on the geospatial matching degree and the entity type fit, and obtain the GEO trust score of each mobile top-level domain.
[0106] Select mobile top-level domains that have reached a preset score threshold in GEO trust score, obtain the unified application entry corresponding to the mobile top-level domain from the GEO mobile domain unified application entry database, and push the mobile top-level domain and its unified application entry to the generative AI engine.
[0107] In one specific implementation, the geographic location constraints and coverage constraints carried in the intent feature vector are subjected to structured parsing. Geographic location constraints are typically embedded in the intent feature vector in the form of city names, administrative division identifiers, or coordinate information from the user's query, while coverage constraints reflect the spatial scale to which the user expects the service to reach. After parsing, the geographic location constraints are mapped to the coordinates of the geographic center point. The coverage area is defined as the geographical radius of influence. Together, these constitute the framework for describing the geographic needs of user queries.
[0108] For each mobile top-level domain in the candidate domain set, extract the spatial vector component from its reputation weight vector. This component encodes the domain name's service scope and geographical coverage dimensions during the construction phase. Perform reverse decoding to reconstruct the domain name's geographical service boundaries, including the service center coordinates. With service coverage radius The domain name geographic service boundary describes the spatial area that the mobile top-level domain name can actually provide services for, and is the basic data structure for geospatial matching.
[0109] The calculation of geospatial matching degree is based on the geographical influence radius. The calculation is based on the degree of spatial overlap between the defined circular area of user needs and the circular area of services defined by the domain name geographic service boundary. The Euclidean distance between the centers of the two circles is then calculated. ,Right now .when When the two circles completely enclose each other, the degree of overlap is the highest; when At the specified interval, the two circles are completely separated, with zero overlap. When the overlap is between these two intervals, the overlap ratio is estimated based on the ratio of the sum of the areas of the intersecting arcs of the two circles to the area of the circle required by the user. This overlap ratio is then normalized to the interval... To obtain the geospatial matching degree The higher the value, the better the service scope of the mobile top-level domain matches the user's geographical query needs.
[0110] After calculating the geospatial matching degree, the entity type fit is further evaluated. The target entity type is extracted from the intent feature vector. With expected service category The former describes the type of business entity the user expects to interact with (such as telecom operators, mobile app stores, mobile phone brand service providers, etc.), while the latter describes the specific service function the user expects to obtain (such as package inquiry, device activation, bill payment, etc.). This is derived from the type vector components of the reputation weight vector of each mobile top-level domain. Decode the business entity type attribution relationship The attribution relationship is stored in the form of a hierarchical type tree, covering multi-level annotations from coarse-grained industry classifications to fine-grained business subclasses.
[0111] Determining type hierarchy containment relationships is achieved by performing node path matching on the hierarchical type tree. If and If they are in the same node in the type tree, it is considered an exact match and a fit coefficient is assigned. ;like yes If the parent class is a match, it is determined that the parent class contains the match, and a fit coefficient is assigned. Its value decreases with increasing hierarchical distance; the greater the hierarchical distance, the higher the value. The smaller; if yes If the subclass is a subclass specialized match, then it is determined to be a fit coefficient. The matching coefficient is typically lower than an exact match but higher than an irrelevant type; if the two types have no ancestor-descendant path relationship in the type tree, they are considered a type mismatch and assigned a matching coefficient of zero. Based on type-level matching, the expected service category is also considered. The service coverage is calculated by intersecting the set of service functions actually provided by the domain name. As a supplementary correction to the type hierarchy fit, the final entity type fit... It is obtained by weighting and combining the type hierarchy coefficient and service coverage, i.e. ,in Type hierarchy matching coefficient (take) , or (Applicable value) This represents the type hierarchy weight, with a value range of [value range missing]. This is used to balance the contribution ratio between type structure matching and service function coverage.
[0112] Obtain geospatial matching degree Fit with entity type Then, a two-dimensional collaborative fusion calculation is performed to obtain the GEO trust score of each mobile top-level domain. The fusion computing employs a nonlinear product enhancement mechanism, the formula of which is: ,in This is the geographic dimension weighting coefficient, used to control the contribution of geospatial matching degree to the fusion result. This is a synergistic gain coefficient used to amplify the overall score when both dimensions are highly matched, allowing domains that excel in both geographic and type dimensions to receive significantly higher GEO trust scores. Compared to simple linear weighting, this non-linear fusion method effectively avoids situations where a high score in one dimension masks a low score in another, more accurately selecting mobile top-level domains that truly meet the user's dual geographic and business needs.
[0113] After completing the GEO trust score calculation, set a preset score threshold. A threshold-based screening process is performed on the GEO trust scores of all mobile top-level domains in the candidate domain set, retaining those that meet the criteria. The domain names constitute the final push candidate set. The settings comprehensively consider the accuracy requirements of the current retrieval enhancement generation stage and the size and distribution of the candidate domain name set, and can be adaptively adjusted based on historical push performance. For mobile top-level domains that pass the threshold screening, their corresponding unified application entry is retrieved from the GEO mobile domain unified application entry database. The unified application entry is a standardized and encapsulated access endpoint description, including the domain's protocol type, deep link path, mobile adaptation parameters, and service function routing identifier. This enables the generative AI engine to directly provide accessible application access paths when presenting results to users, without requiring secondary searches or manual input. The screened mobile top-level domains and their unified application entry are packaged into a structured push payload and pushed to the generative AI engine through a standardized interface. The generative AI engine then integrates this payload into the generated response, achieving accurate and directly accessible mobile top-level domain service responses to user queries.
[0114] In one optional embodiment, receiving an access request triggered by a user via a generative AI engine, constructing a compatibility mapping relationship between the user's terminal environment fingerprint and the unified application entry point, completing adaptive routing, capturing interaction process signals to construct interaction behavior signals and converting them into reputation correction quantities, and writing back the reputation weight vector to drive iterative updates include:
[0115] Receive access requests triggered by users via the generative AI engine, extract the target mobile phone top-level domain name and user terminal identifier carried in the access request, parse the hardware capability parameters and software environment parameters in the user terminal identifier, and construct the user terminal environment fingerprint.
[0116] Extract multiple candidate routing channels and their corresponding routing features from the unified application entry point, perform vectorized matching between the user terminal environment fingerprint and the routing features of each candidate routing channel, construct a compatibility mapping relationship, and complete adaptive routing based on vector similarity.
[0117] Monitor the user's interaction process at the target address, capture the session duration and function call integrity identifier, determine the user stickiness based on the session duration, determine the business loop completion based on the function call integrity identifier, construct the user stickiness and business loop completion into interaction behavior signals, analyze the signal polarity and signal amplitude of the interaction behavior signals, and convert them into reputation correction amount for the target mobile top-level domain.
[0118] Locate the reputation weight vector in the GEO reputation weight matrix that corresponds to the top-level domain of the target mobile phone, write the reputation correction amount back to the reputation weight vector, and drive the iterative update of the reputation weight vector.
[0119] like Figure 4 The diagram shown illustrates the logic flowchart for adaptive routing and reputation update.
[0120] In one specific implementation, after receiving an access request triggered by a user via a generative AI engine, the target mobile top-level domain identifier and the user terminal identifier are extracted from the request message. The user terminal identifier is typically carried in the request header or body as a device fingerprint string, containing two types of information: hardware capability parameters and software environment parameters. Hardware capability parameters include fields describing the device's physical capabilities, such as screen resolution, processor architecture type, memory capacity, and network interface type (e.g., 4G / 5G / Wi-Fi). Software environment parameters include fields describing the device's software stack, such as operating system version, browser kernel identifier, installed runtime environment (e.g., WebView version, mini-program runtime version), and language and region settings. These hardware capability parameters and software environment parameters are then feature-encoded according to predefined field mapping rules and concatenated into a fixed-dimensional vector representation, which is the user terminal environment fingerprint. This vector fully describes the current operating environment state of the user terminal.
[0121] A unified application entry point typically corresponds to multiple candidate routing channels, each with a different technical implementation path, such as an H5 webpage channel, a deep-link channel to a native app, a mini-program channel, or a direct API connection channel. For each candidate routing channel, its routing features are extracted and encoded into a routing feature vector. ,in This is an index for candidate routing channels. Routing characteristics include the minimum operating system version required for the channel, the type of runtime environment it depends on, the supported network protocol versions, and rendering capability requirements. It also includes the user terminal environment fingerprint. Routing feature vectors of each candidate routing channel Cosine similarity is calculated to obtain a compatibility score. The calculation method is as follows The compatibility mapping relationship consists of a set of similarity scores between the user terminal environment fingerprint and the routing feature vectors of each candidate routing channel. The compatibility scores are selected from these sets. The highest-ranking candidate routing channel is selected as the adaptive routing target, forwarding user requests to the routing target address corresponding to that channel, thus completing adaptive routing. This mechanism ensures that user access requests are directed to the technical channel most compatible with their terminal environment, avoiding access failures or experience degradation caused by channel incompatibility with the terminal.
[0122] After routing is completed, the system continuously monitors the user's interaction process at the routed target address. The core interaction signals captured fall into two categories: session duration. With Function Call Integrity Identifier Session duration Records the total duration in seconds from when the user enters the target route address to when the session ends (either by actively logging out or disconnecting due to timeout). Function call integrity identifier. It is a Boolean or multi-level enumeration identifier used to mark whether the user has completed a business loop operation in the current session, such as completing a search and clicking on the result, completing a form submission, completing a service appointment, or completing content consumption, etc.
[0123] Based on session duration By comparing with the preset reference duration threshold Compare and determine the level of user stickiness. :when At that time, the user stickiness was considered high. Take a positive value; when Below At that time, user stickiness was deemed low. Take negative or zero values, the specific value is as follows: and The ratio is obtained through a linear mapping. Based on the function call integrity identifier. Determine the completion rate of the business loop. :like If marked as fully completed, then Take the largest positive value; if If marked as partially completed, then Take the middle positive value; if If marked as incomplete, then Take a zero or negative value. This indicates the user's level of engagement. Completion of business loop Combining and constructing interactive behavior signals This signal fully reflects the user's actual experience quality during this visit.
[0124] Analyzing interactive behavior signals The signal polarity and signal amplitude. Signal polarity is determined by... and The polarity is determined jointly by the signs: when both are positive, the polarity is positive, indicating that this interaction contributes positively to the domain name's reputation; when either is negative, the polarity is negative, indicating that this interaction has a negative need to correct the domain name's reputation; when both are mixed (positive and negative), the polarity is determined comprehensively according to weighted rules. Signal amplitude. Depend on and The weighted sum is obtained, that is ,in and The contribution weights for user stickiness and business loop completion are respectively, satisfying... The values of both can be pre-configured according to the business scenario. The signal polarity and signal amplitude are comprehensively encoded and converted into a reputation correction value for the target mobile phone's top-level domain. The reputation correction factor is the same as the reputation weight vector. For vectors of the same dimension, the direction and magnitude of the correction of each component are determined by the polarity and magnitude of the interaction behavior signal, ensuring that the reputation correction amount can accurately reflect the quality evaluation result of this interaction.
[0125] Positioning GEO Reputation Weight Matrix The reputation weight vector corresponding to the target mobile phone top-level domain , credit correction amount Write back the reputation weight vector and perform an iterative update operation. The update rule uses an additive update strategy with a decay coefficient: ,in The learning rate parameter controls the step size of the correction to the reputation weight vector for a single interaction, preventing excessive fluctuations in the reputation weight vector caused by extreme single interaction behaviors. After the update is complete, the new reputation weight vector... Write back to the GEO reputation weight matrix The corresponding row position in the matrix is used to complete the local update of the matrix. By continuously accumulating multiple interaction behavior signals and repeatedly executing the above write-back process, the reputation weight vector can be dynamically adjusted according to changes in the user's actual access experience. This ensures that the GEO reputation weight matrix always reflects the latest reputation status of each mobile top-level domain, thereby providing a more accurate reputation basis for domain name priority citation in the subsequent generative AI engine retrieval enhancement generation stage, forming a data-driven reputation closed-loop iteration mechanism.
[0126] In one optional embodiment, multiple candidate routing channels and their corresponding routing features are extracted from the unified application entry point. The user terminal environment fingerprint is then vectorized and matched with the routing features of each candidate routing channel to construct a compatibility mapping relationship. Adaptive routing is then performed based on vector similarity, including:
[0127] Obtain the unified application entry corresponding to the top-level domain of the target mobile phone, extract multiple candidate routing channels configured by the unified application entry, parse the resource consumption characteristics and response performance characteristics of each candidate routing channel, vectorize the resource consumption characteristics and response performance characteristics, and generate routing feature vectors for each candidate routing channel.
[0128] The terminal resource status and network connection status in the user terminal environment fingerprint are analyzed, and the terminal resource status and network connection status are vectorized to generate user terminal feature vectors.
[0129] Calculate the vector similarity between the user terminal feature vector and the routing feature vector of each candidate routing channel, and construct a compatibility mapping relationship between each candidate routing channel and its corresponding vector similarity.
[0130] The system selects the candidate routing channel with the highest routing priority weight from the compatibility mapping relationship and routes the access request to the routing target address corresponding to the candidate routing channel. When routing fails, the system selects the next best candidate routing channel in descending order of routing priority weight for rerouting until routing is successful, thus completing adaptive routing.
[0131] In one specific implementation, after obtaining the unified application entry point corresponding to the top-level domain of the target mobile phone, all candidate routing channels are enumerated from the configuration information of that entry point. Each candidate routing channel represents an access link from the user terminal to the actual business service, which may correspond to different protocol types (such as HTTP / 2, WebSocket, deep links in native apps, etc.), different service node addresses, or different content delivery network nodes. For each candidate routing channel, its resource consumption characteristics and response performance characteristics are analyzed. Resource consumption characteristics include historical statistical descriptions of the terminal-side memory usage, CPU call overhead, and network bandwidth consumption during the connection establishment and session maintenance process; response performance characteristics cover quantifiable network quality indicators such as the channel's average first-byte response latency, data throughput rate, and packet loss rate. The above two types of features are numerically encoded according to a preset feature dimension order and concatenated into a fixed-dimensional routing feature vector. ,in This serves as the index for candidate routing channels. Each channel corresponds to an independent feature vector, which together constitute the feature set of candidate routing channels.
[0132] While generating the routing feature vector, the terminal resource status and network connection status information carried in the current user terminal environment fingerprint are parsed. The parsed terminal resource status includes indicators reflecting the terminal hardware's availability, such as current available physical memory capacity, CPU idle rate, storage read / write speed, and battery level. The parsed network connection status includes the current network type (e.g., 5G, 4G LTE, Wi-Fi 6), real-time signal strength, current bandwidth estimate, and network latency measurement. The terminal resource status and network connection status are numerically encoded according to the same dimensional alignment as the routing feature vector and concatenated to generate the user terminal feature vector. The dimension design of this vector is related to the routing feature vector. To maintain consistency and ensure mathematical comparability in subsequent vector similarity calculations, the specific strategy for dimension alignment is as follows: for the dimension describing the lower limit of resource demand in the routing feature vector, the corresponding upper limit of the resource capacity currently available by the terminal is filled into the user terminal feature vector, thereby enabling the two to match within the same semantic space.
[0133] After constructing the two types of vectors, the user terminal feature vectors are calculated sequentially. With each candidate route channel routing feature vector The vector similarity between them is used to obtain the first... Compatibility score of candidate routing channels The calculation method uses cosine similarity, which is obtained by dividing the inner product of two vectors by the product of their respective magnitudes to obtain a normalized similarity value. The value range is [range missing]. The closer the value is to 1, the better the resource requirements and performance characteristics of the candidate routing channel match the current terminal environment. Before calculating the cosine similarity, L2 normalization preprocessing is performed on both types of vectors to eliminate the interference of differences in the dimensions of different feature dimensions on the similarity calculation results. After the calculation is completed, the candidate routing channels are indexed. and the corresponding compatibility score A key-value mapping structure is constructed to form a compatibility mapping relationship. This mapping relationship uses the channel index as the key, the compatibility score as the value, and also associates the corresponding route target address, providing a complete query basis for subsequent routing decisions.
[0134] After the compatibility mapping relationship is established, it is based on the compatibility score. All candidate routing channels are sorted in descending order, with higher-ranked channels assigned higher routing priority weights. The assignment strategy for routing priority weights can employ linear normalization, setting the weight of the highest-scoring channel to 1 and decreasing the weights of other channels proportionally to their scores. Alternatively, a softmax transformation can be used to convert the compatibility score into a probability distribution of priority weights, resulting in smoother weight differences. During actual routing execution, the candidate routing channel with the highest priority weight is selected first, and the current access request is forwarded to the routing destination address corresponding to that channel. The routing destination address can be a deep link URI in the native app, a URL for a mobile H5 page, or a page path in a mini-program, among other forms, determined by the unified application entry configuration information.
[0135] When a priority channel fails, a degraded rerouting mechanism is triggered. Routing failure criteria include: connection timeout, destination address unreachable, server returning a 5xx error code, and protocol handshake failure. Once any of these failure conditions is detected, the next candidate routing channel is immediately selected from the compatibility mapping relationship in descending order of routing priority weight, and the access request is rerouted to the routing destination address corresponding to that channel. This degraded process continues iteratively until a candidate routing channel successfully responds to the user's access request, completing adaptive routing. If all candidate routing channels fail, a unified service unavailability message is returned to the user terminal, and this full failure event is recorded in the routing log for use in subsequent reputation correction processes.
[0136] Throughout the adaptive routing process, the quality of the route feature vector construction directly impacts the accuracy of the compatibility mapping. To enhance the ability of the route feature vector to represent the actual channel state, both resource consumption features and response performance features are encoded using a dual-index approach: the statistical mean and standard deviation within a recent sliding window, rather than statically configured values. Specifically, a fixed-length historical performance observation queue is maintained for each candidate route channel. Each record in the queue corresponds to measured performance data from a single historical access. When the queue is full, a first-in, first-out (FIFO) strategy is used for updating. Before each routing decision, the latest statistical features are extracted from the queue to regenerate the route feature vector, ensuring that the compatibility mapping reflects the dynamic changes in the current network environment and service state, rather than relying on outdated static descriptions. This dynamic update mechanism is consistent with the iterative update mechanism of the reputation weight vector, jointly supporting the overall adaptive capability of the method.
[0137] The generation of user terminal feature vectors also needs to consider the real-time nature of the terminal status. The CPU idle rate and available memory capacity in the terminal resource status may fluctuate significantly in a short period. Therefore, the current terminal status data is re-collected each time an access request is received, rather than reusing the terminal feature vector generated in the previous request. The collection of network connection status relies on the network quality detection interface provided by the terminal operating system to obtain the real-time bandwidth estimate and latency measurement of the current connection. By performing vector similarity matching between the real-time collected terminal status and the dynamically updated routing feature vectors, routing decisions can be automatically adjusted when the user terminal environment changes (such as switching from Wi-Fi to 5G, or when terminal memory is low), guiding access requests to the most suitable candidate routing channel, thus ensuring the continuity and stability of the access experience under different terminal environment conditions.
[0138] A second aspect of this invention provides a GEO mobile domain name AI priority citation optimization system, comprising:
[0139] The semantic parsing unit is used to perform multi-layer semantic parsing on the original business data of the mobile top-level domain. It determines the structured data packet through entity recognition and attribute extraction, and obtains the compressed encoded data of each mobile top-level domain through metadata compression encoding oriented to the large model context window.
[0140] The reputation modeling unit is used to construct the GEO reputation weight matrix based on compressed encoded data through vectorized modeling, and write the reputation weight vector corresponding to each mobile top-level domain.
[0141] The intent evaluation unit is used to receive user query requests from the generative AI engine and perform semantic encoding to obtain intent feature vectors. During the retrieval enhancement generation stage of the generative AI engine, the unit performs quality evaluation on candidate mobile top-level domains based on the intent feature vectors to obtain a set of candidate domains.
[0142] The trust scoring unit is used to analyze the intent feature vector and the reputation weight vector of each domain in the candidate domain set, calculate the geospatial matching degree and entity type fit degree, obtain the GEO trust score, and push the mobile top-level domains and their unified application entry that reach the preset score threshold to the generative AI engine.
[0143] The routing update unit is used to receive access requests triggered by users through the generative AI engine, build a compatibility mapping relationship between the user terminal environment fingerprint and the unified application entry, complete adaptive routing, capture the interaction process, build interaction behavior signals and convert them into reputation correction quantities, and write back the reputation weight vector to drive iterative updates.
[0144] A third aspect of the present invention provides an electronic device, comprising:
[0145] processor;
[0146] Memory used to store processor-executable instructions;
[0147] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.
[0148] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.
[0149] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.
[0150] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. The GEO mobile domain name AI priority reference optimization method is characterized by, include: Multi-layer semantic parsing is performed on the original business data of the mobile top-level domain. Structured data packets are determined through entity recognition and attribute extraction. Compressed encoded data of each mobile top-level domain is obtained through metadata compression encoding oriented towards a large model context window. Based on compressed encoded data, a GEO reputation weight matrix is constructed through vectorized modeling, and the reputation weight vector corresponding to each mobile top-level domain is written into it. The system receives user query requests from the generative AI engine and performs semantic encoding to obtain intent feature vectors. During the retrieval enhancement generation stage of the generative AI engine, it performs quality assessment on candidate mobile top-level domains based on the intent feature vectors to obtain a set of candidate domains. The intent feature vector and the reputation weight vector of each domain in the candidate domain set are analyzed, the geospatial matching degree and entity type fit degree are calculated, and the GEO trust score is obtained. The mobile top-level domains that reach the preset score threshold and their unified application entry are pushed to the generative AI engine. It receives access requests triggered by users through a generative AI engine, constructs a compatibility mapping relationship between the user's terminal environment fingerprint and the unified application entry, completes adaptive routing, captures the interaction process, constructs interaction behavior signals and converts them into reputation correction quantities, and writes back the reputation weight vector to drive iterative updates.
2. The method according to claim 1, characterized in that, Multi-layer semantic parsing is performed on the raw business data of the mobile top-level domains. Structured data packets are determined through entity recognition and attribute extraction. Metadata compression encoding targeting a large model context window is then used to obtain the compressed encoded data for each mobile top-level domain, including: Multi-layer semantic parsing is performed on the original business data of the mobile top-level domain. The first semantic layer extracts the business entity type and industry affiliation pointed to by the domain. The second semantic layer extracts the service scope and geographical coverage dimension of the business entity. The third semantic layer extracts the mapping relationship between the business entity and the user demand category. The extraction results of each semantic layer are organized into a structured data packet containing entity identifiers, attribute key-value pairs and inter-layer reference indexes through entity recognition and attribute extraction. The structured data packet is subjected to metadata compression encoding oriented towards the large model context window. By constructing a semantic dependency graph, entity identifiers and attribute key-value pairs with co-occurrence relationships in the structured data packet are grouped into semantic unit nodes. The importance weight of the nodes is calculated based on the reference frequency and inter-layer reference index between semantic unit nodes. The semantic unit nodes are prioritized according to the importance weight of the nodes. High-priority semantic unit nodes are extracted according to the capacity constraint of the large model context window and encoded into compressed data. The compressed encoded data is bound and stored with the corresponding mobile top-level domain identifier to obtain the compressed encoded data of each mobile top-level domain.
3. The method according to claim 1, characterized in that, Based on compressed encoded data, a GEO reputation weight matrix is constructed through vectorized modeling, and the reputation weight vectors corresponding to each mobile top-level domain are written in: Extract semantic unit nodes and node importance weights from the compressed encoded data of each mobile top-level domain. Transform the attribution relationship between entity identifiers and business entity types in semantic unit nodes into type vector components. Transform the service scope and geographic coverage dimension in attribute key-value pairs into spatial vector components. Transform the mapping relationship strength in inter-layer reference index into association vector components. Combine type vector components, spatial vector components and association vector components to form an initial feature vector. The initial feature vector is subjected to weighted fusion based on node importance weights. By using the node importance weights as multiplicative factors on each vector component derived from the corresponding semantic unit node, a weighted feature vector is obtained. The user access frequency and single access duration of the mobile top-level domain within a preset time window are collected, and the joint distribution entropy of access frequency and duration is calculated to obtain the domain activity stability metric. The domain activity stability metric and the weighted feature vector are subjected to a component-wise exponential mapping transformation to obtain the reputation weight vector of each mobile top-level domain. Construct a GEO reputation weight matrix with the mobile top-level domain identifier as the row dimension and the vector component index as the column dimension, and write each reputation weight vector into the GEO reputation weight matrix according to the row dimension.
4. The method according to claim 1, characterized in that, The system receives user query requests from the generative AI engine and performs semantic encoding to obtain intent feature vectors. During the retrieval enhancement generation phase of the generative AI engine, quality assessments are performed on candidate mobile top-level domains based on these intent feature vectors, resulting in a candidate domain set including: The system receives user query requests from the generative AI engine, performs query intent decomposition on the user query requests to obtain core demand descriptions and auxiliary constraints, performs semantic hierarchical parsing on the core demand descriptions to identify target entity types and expected service categories in the core demand descriptions, performs geographic semantic extraction on the auxiliary constraints to identify geographic location limitations and coverage limitations in the auxiliary constraints, and performs semantic encoding on the target entity types, expected service categories, geographic location limitations, and coverage limitations to obtain intent feature vectors. In the retrieval enhancement generation stage of the generative AI engine, the reputation weight vector of each mobile top-level domain is extracted from the GEO reputation weight matrix. The semantic space distance between the intent feature vector and each reputation weight vector is calculated. The semantic space distance and the spatial vector component in the reputation weight vector of the corresponding mobile top-level domain are used to perform a collaborative matching degree calculation to obtain the intent matching degree score of each mobile top-level domain. Based on the intent matching degree score, the candidate mobile top-level domains are evaluated for quality. Mobile top-level domains that meet the quality admission criteria are selected to obtain a set of candidate domains.
5. The method according to claim 1, characterized in that, The intent feature vector and the reputation weight vector of each domain in the candidate domain set are analyzed to calculate the geospatial matching degree and entity type fit, thereby obtaining the GEO trust score. Mobile top-level domains that reach the preset score threshold and their unified application entry points are then pushed to the generative AI engine, including: The reputation weight vector corresponding to each mobile top-level domain is extracted from the candidate domain set. The geographical location limitation and coverage limitation in the intent feature vector are analyzed. The geographical center point and geographical influence radius corresponding to the geographical location limitation are constructed. The spatial vector components in each reputation weight vector are analyzed. The service range and geographical coverage dimension encoded in the spatial vector components are extracted and transformed into the domain name geographical service boundary. The degree of spatial overlap between the geographical influence radius and the geographical service boundary of each domain name is determined to obtain the geospatial matching degree. Extract the target entity type and expected service category from the intent feature vector, decode the business entity type attribution relationship in the type vector component of each reputation weight vector, determine the type hierarchy inclusion relationship between the target entity type and the business entity type attribution relationship, obtain the entity type fit, perform a two-dimensional collaborative fusion calculation on the geospatial matching degree and the entity type fit, and obtain the GEO trust score of each mobile top-level domain. Select mobile top-level domains that have reached a preset score threshold in GEO trust score, obtain the unified application entry corresponding to the mobile top-level domain from the GEO mobile domain unified application entry database, and push the mobile top-level domain and its unified application entry to the generative AI engine.
6. The method according to claim 1, characterized in that, It receives access requests triggered by users via a generative AI engine, constructs a compatibility mapping relationship between the user's terminal environment fingerprint and the unified application entry point, completes adaptive routing, captures interaction process signals to construct interaction behavior signals and converts them into reputation correction quantities, and writes back the reputation weight vector to drive iterative updates, including: Receive access requests triggered by users via the generative AI engine, extract the target mobile phone top-level domain name and user terminal identifier carried in the access request, parse the hardware capability parameters and software environment parameters in the user terminal identifier, and construct the user terminal environment fingerprint. Extract multiple candidate routing channels and their corresponding routing features from the unified application entry point, perform vectorized matching between the user terminal environment fingerprint and the routing features of each candidate routing channel, construct a compatibility mapping relationship, and complete adaptive routing based on vector similarity. Monitor the user's interaction process at the target address, capture the session duration and function call integrity identifier, determine the user stickiness based on the session duration, determine the business loop completion based on the function call integrity identifier, construct the user stickiness and business loop completion into interaction behavior signals, analyze the signal polarity and signal amplitude of the interaction behavior signals, and convert them into reputation correction amount for the target mobile top-level domain. Locate the reputation weight vector in the GEO reputation weight matrix that corresponds to the top-level domain of the target mobile phone, write the reputation correction amount back to the reputation weight vector, and drive the iterative update of the reputation weight vector.
7. The method according to claim 6, characterized in that, Extract multiple candidate routing channels and their corresponding routing features from the unified application entry point. Perform vectorized matching between the user terminal environment fingerprint and the routing features of each candidate routing channel to construct a compatibility mapping relationship. Adaptive routing is then performed based on vector similarity, including: Obtain the unified application entry corresponding to the top-level domain of the target mobile phone, extract multiple candidate routing channels configured by the unified application entry, parse the resource consumption characteristics and response performance characteristics of each candidate routing channel, vectorize the resource consumption characteristics and response performance characteristics, and generate routing feature vectors for each candidate routing channel. The terminal resource status and network connection status in the user terminal environment fingerprint are analyzed, and the terminal resource status and network connection status are vectorized to generate user terminal feature vectors. Calculate the vector similarity between the user terminal feature vector and the routing feature vector of each candidate routing channel, and construct a compatibility mapping relationship between each candidate routing channel and its corresponding vector similarity. The system selects the candidate routing channel with the highest routing priority weight from the compatibility mapping relationship and routes the access request to the routing target address corresponding to the candidate routing channel. When routing fails, the system selects the next best candidate routing channel in descending order of routing priority weight for rerouting until routing is successful, thus completing adaptive routing.
8. A GEO mobile domain name AI priority reference optimization system, used to implement the method as described in any one of claims 1-7, characterized in that, include: The semantic parsing unit is used to perform multi-layer semantic parsing on the original business data of the mobile top-level domain. It determines the structured data packet through entity recognition and attribute extraction, and obtains the compressed encoded data of each mobile top-level domain through metadata compression encoding oriented to the large model context window. The reputation modeling unit is used to construct the GEO reputation weight matrix based on compressed encoded data through vectorized modeling, and write the reputation weight vector corresponding to each mobile top-level domain. The intent evaluation unit is used to receive user query requests from the generative AI engine and perform semantic encoding to obtain intent feature vectors. During the retrieval enhancement generation stage of the generative AI engine, the unit performs quality evaluation on candidate mobile top-level domains based on the intent feature vectors to obtain a set of candidate domains. The trust scoring unit is used to analyze the intent feature vector and the reputation weight vector of each domain in the candidate domain set, calculate the geospatial matching degree and entity type fit degree, obtain the GEO trust score, and push the mobile top-level domains and their unified application entry that reach the preset score threshold to the generative AI engine. The routing update unit is used to receive access requests triggered by users through the generative AI engine, build a compatibility mapping relationship between the user terminal environment fingerprint and the unified application entry, complete adaptive routing, capture the interaction process, build interaction behavior signals and convert them into reputation correction quantities, and write back the reputation weight vector to drive iterative updates.
9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.
10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.
