Sales auxiliary system and method based on AI technology

Abstract

The invention relates to the technical field of sales assistance, in particular to an AI technology-based sales assistance system and method, and the system comprises a function modeling module, an interaction analysis module, an inventory decision module, a strategy compiling module and a terminal adaptation module. According to the method, function nodes are extracted through a natural language processing technology, a topological relation is constructed, dynamic hierarchical management of a sales function system is realized, an attention intensity model is established through interaction duration data, a priority strategy is generated in combination with an inventory pressure coefficient, and recommendation weight distribution is optimized when resources are insufficient. A decision tree algorithm is used for carrying out multi-dimensional parameter conflict detection and dynamic sorting, a terminal adaptive personalized verbal skill is generated, the resource allocation real-time response capability is improved, an inventory and strategy linkage mechanism is enhanced, accurate matching of verbal skill and inventory is realized, artificial decision delay is eliminated, customer demands and inventory pressure are balanced, and sales conversion resource allocation targeting is enhanced.

Description

Technical Field

[0001] The present invention relates to the field of sales assistance technology, and in particular to a sales assistance system and method based on AI technology. Background Art

[0002] The field of sales support technology involves systems and methods that support and optimize the sales process through information technology. Core areas include customer information management, sales data analysis, sales process tracking, customer relationship maintenance, and sales strategy development. This technology encompasses the information management of the entire sales process, from lead acquisition and opportunity conversion to order management and after-sales service. It primarily leverages computer systems and online platforms to collect, analyze, and provide feedback on sales data, thereby improving sales efficiency and decision-making accuracy. It is a crucial technical support for companies to enhance their sales performance in the competitive market.

[0003] Traditional sales support systems are information management tools used to assist salespeople in completing sales tasks. They address technical issues such as the collection, organization, and querying of sales data and customer interactions. Traditional sales support systems record customer information in spreadsheets, manually organize sales records, retrieve raw sales data through a simple query interface, and manually generate sales reports using static templates to support salespeople's daily work.

[0004] Traditional sales support systems use spreadsheets for recording and manual organization. Data updates rely on manual operations, resulting in information synchronization delays. Customer interaction records lack behavioral analysis dimensions, reducing the depth of demand mining. Static report generation mechanisms cannot reflect real-time business changes. The inventory management module and sales strategy module run independently, resulting in data gaps. Fixed script templates are difficult to adapt to the dynamic market environment. The sales process lacks an intelligent linkage mechanism, and priority settings are rigidly set, ignoring the impact of inventory fluctuations. This ultimately leads to structural defects such as one-sided decision-making basis, delayed resource allocation, and deviations in customer demand response. Summary of the Invention

[0005] In order to solve the technical problems that traditional sales support systems use spreadsheet records and manual sorting mode, data updates rely on manual operations resulting in information synchronization delays, customer interaction records lack behavioral analysis dimensions that reduce the depth of demand mining, static report generation mechanisms cannot reflect real-time business changes, inventory management modules and sales strategy modules run independently to produce data gaps, fixed speech templates are difficult to adapt to dynamic market environments, sales process links lack intelligent linkage mechanisms, priority settings are fixed and ignore the impact of inventory fluctuations, and ultimately lead to structural defects such as one-sided decision-making basis, delayed resource allocation, and deviation in customer demand response, the embodiment of the present invention provides a sales support system and method based on AI technology. The technical solution is as follows: In one aspect, a sales assistance system based on AI technology is provided, the system comprising: The function modeling module is used to extract the keywords of the manual through natural language processing technology, divide them into core function nodes, auxiliary function nodes, and extended function nodes using the hierarchical analysis method, establish the function node topology map, and pass it to the interactive analysis module; An interactive parsing module is used to match the functional node topology, calculate the label matching time, calculate the node attention intensity through the entropy weight model, generate a dynamic priority list, and pass it to the inventory decision module; An inventory decision module is configured to receive the dynamic priority list, calculate the inventory pressure coefficient of a single product, add inventory weight tags to the first three nodes in the list when the pressure coefficient exceeds a pressure threshold, generate a weighted priority list, and pass it to the strategy compilation module; A policy compilation module is used to parse the inventory weight tags in the weighted priority list, perform conflict detection and priority sorting on the tag states through a decision tree classifier, generate an execution parameter package, and pass it to the terminal adaptation module; The terminal adaptation module is used to parse the inventory adaptation speech parameters in the execution parameter package, locate the terminal device according to the session ID code, update the inventory guidance speech field in the recommendation column, and generate an interaction log record.

[0006] As a further solution of the present invention, the functional node topology diagram includes the core functional node hierarchy, the auxiliary functional node association path, and the extended functional node topology relationship; the dynamic priority list includes the node attention strength value, the entropy weight method weight coefficient, and the time sequence priority index; the weighted priority list includes the inventory weight tag identifier, the inventory turnover rate threshold parameter, and the priority node identifier; the execution parameter package includes the decision tree splitting threshold, the execution priority queue, and the conflict detection status code; the interaction log record includes the speech version number, the device UUID mapping table, and the recommendation field update log; The entropy weight coefficient is the objective weight value of the indicator obtained by calculating the information entropy.

[0007] As a further solution of the present invention, the functional modeling module includes: The keyword extraction submodule obtains the instruction manual text content, identifies paragraph sentences and word structures, filters out low-frequency words, calculates word dependency strength based on word syntactic collocation and co-occurrence frequency, constructs a word dependency graph, applies a clustering algorithm to calculate the coupling index between words, and generates a keyword relevance coefficient value; The keyword association coefficient value quantifies the semantic association strength between the keywords in the specification and describes the coupling relationship between the keywords; The functional hierarchy division submodule constructs a word weight list based on the keyword relevance coefficient value, uses the hierarchical analysis method to construct a word group comparison relationship, calculates the consistency ratio, and obtains the functional hierarchy weight coefficient distribution value; The function level weight coefficient distribution value is used to represent the corresponding critical weight of the function module in the overall structure, and is calculated based on the correlation between keywords; The node topology construction submodule calls the functional hierarchy weight coefficient distribution value, sets the core function threshold and auxiliary function interval, constructs a node type classification table, calculates the association strength between nodes based on the weight value, integrates the node pair relationships with connection probabilities higher than the set ratio in the probability table, and obtains a functional node topology map.

[0008] As a further solution of the present invention, the interactive parsing module includes: The node matching submodule obtains the functional node topology map, detects the duration of node label matching, calculates the matching duration of each node based on the node label and network topology, records the corresponding data, and generates a matching duration value; The attention intensity calculation submodule calculates the attention intensity of the node based on the matching duration value through the entropy weight model, compares the label matching degree of the node, calculates the intensity coefficient of the node attention, and generates the node attention intensity value; A priority generation submodule establishes a node sorting sequence according to the node attention strength value, sets a strength fluctuation threshold and an inventory response coefficient, dynamically weights nodes with strength values higher than the threshold according to the response coefficient, and generates a dynamic priority list; The intensity fluctuation threshold determines whether the node attention intensity has changed significantly, and the change exceeds ±20%, which is set based on the original mean fluctuation range; The inventory response factor reflects the weight of the impact of inventory on priority adjustment, and its value range is 0.3–0.8, which is set according to the system's sensitivity to inventory availability.

[0009] As a further solution of the present invention, the attention intensity value is calculated using the formula: ; in, represents the attention intensity value of node j at time t, Represents the matching duration between node i and node j at time t, which is normalized by Z-score and unified into dimensionless parameters. Represents the topological structure weight coefficient of node i and node j at time t, Represents the number of interactions between node i and node j at time t, which is normalized by Z-score and unified into a dimensionless parameter. Represents the degree of difference between the labels of node i and node j at time t, represents the feature synergy value of node i and node j at time t, represents the average value of the feature coordination value between node j and its adjacent nodes at time t, and n represents the number of nodes involved in the matching calculation.

[0010] As a further solution of the present invention, the inventory decision module includes: The pressure coefficient calculation submodule obtains the inventory turnover rate, inventory age distribution, and demand volatility parameters of the single product in the dynamic priority list, establishes a three-dimensional coordinate system to calculate the vector modulus, performs Z-score standardization conversion, and generates the single product inventory pressure coefficient; The weight mark trigger submodule calls the single product inventory pressure coefficient, calculates the exponential moving average of the pressure value in the past seven days, and constructs a pressure change channel. When the real-time value breaks through the upper track of the channel and exceeds the pressure threshold, a weight trigger signal is generated; The pressure threshold is a critical standard for measuring whether inventory pressure is abnormally high. If it exceeds this value, intervention will be triggered. It is set differently based on category attributes, season, and life cycle factors. The list weight injection submodule parses the weight trigger signal, locates the first three nodes in the list, calculates the ratio of the pressure coefficient to the reference value as the amplification factor, establishes a linear mapping relationship and writes the node attributes to generate a weighted priority list.

[0011] As a further solution of the present invention, the policy compilation module includes: A weight parsing submodule parses the inventory weight tags in the weighted priority list, converts the weight percentages into decimal floating point numbers, matches the type code to the capacity upper limit, and generates a weight feature vector; The conflict detection submodule calls the weight feature vector, inputs it into the decision tree classifier, calculates the standard deviation of the priority level parameter, sets the node splitting threshold, analyzes the degree of dispersion through the Gini coefficient, calculates the conflict index deviation value, detects the over-limit node, and outputs the conflict index matrix; The decision tree classifier uses the input weight feature vector to classify the conflict situation into differentiated priority levels through node splitting and Gini coefficient calculation; The parameter generation submodule calls the exception identifier in the conflict index matrix, establishes a linear relationship between priority and conflict index, sets a weighted formula for weight and conflict, calculates and sorts, and encapsulates it into an execution parameter package.

[0012] As a further solution of the present invention, the conflict index deviation value is calculated using the formula: ; in, Represents the conflict index deviation value of the i-th node, Represents the number of weight items in the i-th node, Represents the j-th value of the j-th weight feature vector in the i-th node, Represents the upper limit of the capacity corresponding to the type code matched by the jth item in the i-th node, Represents the value of the j-th priority parameter in the i-th node, represents the arithmetic mean of the priority level parameters in the i-th node, represents the overall average value of the conflict index deviation value of the node, is the maximum capacity of the system, is a small constant to avoid zero, set to 0.0001.

[0013] As a further solution of the present invention, the terminal adaptation module includes: The session parameter parsing submodule parses the inventory adaptation speech parameters in the execution parameter package, extracts the session ID feature sequence, performs a bit-by-bit comparison operation with the device code library, selects entries with matching fault tolerance values, and generates a device positioning code; The speech field update submodule parses the interaction frequency matrix of the device location code, calculates the covariance coefficient between the field click rate and the inventory turnover rate, establishes a priority sorting vector, updates the field version hash value, and generates a speech version identifier; The interaction log generation submodule extracts the speech field change record corresponding to the speech version identifier, captures the device response delay difference, counts the field click trigger frequency, encodes the operation type code, response status code and timestamp into triples, and generates an interaction log record.

[0014] In another aspect, an AI-based sales assistance method is provided, wherein the AI-based sales assistance method is performed based on the above-mentioned AI-based sales assistance system and comprises the following steps: S1: Extract the keywords of the instruction manual through natural language processing technology, divide them into core function nodes, auxiliary function nodes, and extended function nodes using the hierarchical analysis method, and establish a function node topology map; S2: Count the residence time of the labels in the functional node topology graph, combine the node weights and input the entropy weight model to calculate the attention intensity value and generate a dynamic priority list; S3: Receives the dynamic priority list, calculates the inventory pressure coefficient of each item, and adds inventory weight tags to the first three nodes in the list when the pressure coefficient exceeds a preset threshold, thereby generating a weighted priority list. S4: parsing the inventory weight tags in the weighted priority list, performing conflict detection and priority sorting on the tag states through a decision tree classifier, and generating an execution parameter package; S5: Parse the inventory adaptation speech parameters in the execution parameter package, locate the terminal device according to the session ID code, update the inventory guidance speech field in the recommendation column, and generate an interaction log record.

[0015] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least: Based on natural language processing technology, a functional node topology structure is constructed to achieve intelligent stratification and dynamic mapping of the product function system. The entropy weight method is used to quantify the node attention intensity in user interaction behavior. Combined with the inventory pressure coefficient, the weight marking mechanism is triggered. The decision tree classifier is used for multi-dimensional strategy conflict detection and self-matching sorting, and finally the real-time parameterized update of terminal recommendation scripts is achieved. This technical path establishes a dynamic feedback mechanism between sales strategy and inventory status, enhances the data correlation between demand identification and resource scheduling, optimizes the collaborative efficiency of sales decision-making and inventory turnover, and forms a closed-loop business response system. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0017] Figure 1 It is a system flow chart of the present invention; Figure 2 Schematic diagram of the system framework of the present invention; Figure 3 It is a diagram of the method steps of the present invention. DETAILED DESCRIPTION

[0018] The technical solution of the present invention is described below in conjunction with the accompanying drawings.

[0019] In the embodiments of the present invention, words such as "exemplarily" and "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as an "exemplary" in the present invention should not be interpreted as being preferred or advantageous over other embodiments or designs. Rather, the use of the word "exemplary" is intended to present concepts in a concrete manner. Furthermore, in the embodiments of the present invention, "and / or" can mean both or either of the two.

[0020] In the embodiments of the present invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, when the distinction is not emphasized, the meanings they convey are the same. The terms "of," "corresponding," and "corresponding" may sometimes be used interchangeably. It should be noted that, when the distinction is not emphasized, the meanings they convey are the same.

[0021] In the embodiments of the present invention, sometimes a subscript such as W1 may be written as a non-subscript such as W1. When the difference is not emphasized, the meanings to be expressed are the same.

[0022] In order to make the technical problems, technical solutions and advantages to be solved by the present invention clearer, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0023] The embodiment of the present invention provides a sales support system based on AI technology, such as Figure 1 The flowchart of the sales assistance system based on AI technology is shown, which includes: The function modeling module is used to extract the keywords of the manual through natural language processing technology, divide them into core function nodes, auxiliary function nodes, and extended function nodes using the hierarchical analysis method, establish the function node topology map, and pass it to the interactive analysis module; The interactive parsing module is used to match the functional node topology, calculate the label matching time, calculate the node attention intensity through the entropy weight model, generate a dynamic priority list, and pass it to the inventory decision module; The inventory decision module receives the dynamic priority list, calculates the inventory pressure coefficient of each product, adds inventory weight tags to the first three nodes in the list when the pressure coefficient exceeds the pressure threshold, generates a weighted priority list, and passes it to the strategy compilation module; The strategy compilation module is used to parse the inventory weight tags in the weighted priority list, perform conflict detection and priority sorting on the tag status through the decision tree classifier, generate an execution parameter package, and pass it to the terminal adaptation module; The terminal adaptation module is used to parse the inventory adaptation speech parameters in the execution parameter package, locate the terminal device according to the session ID code, update the inventory guidance speech field in the recommendation column, and generate interaction log records.

[0024] The function node topology diagram includes the core function node hierarchy, auxiliary function node association paths, and extended function node topology relationships. The dynamic priority list includes the node attention strength value, entropy weight coefficient, and time sequence priority index. The weighted priority list includes the inventory weight tag identifier, inventory turnover rate threshold parameter, and priority node identifier. The execution parameter package includes the decision tree splitting threshold, execution priority queue, and conflict detection status code. The interaction log records include the script version number, device UUID mapping table, and recommendation field update log. The weight coefficient of the entropy weight method is the objective weight value of the indicator obtained by calculating the information entropy.

[0025] Specifically, if Figure 2 As shown, the functional modeling module includes: The keyword extraction submodule obtains the instruction manual text content, identifies paragraph sentences and word structures, filters out low-frequency words, calculates word dependency strength based on word syntactic collocation and co-occurrence frequency, constructs a word dependency graph, applies a clustering algorithm to calculate the coupling index between words, and generates a keyword relevance coefficient value; The keyword extraction submodule processes the instruction manual text content. First, the full text of the instruction manual is extracted, the language encoding format is set to UTF-8, and the regular matching method is used to disassemble the text structures such as multiple paragraphs, titles, and bullet points. The short sentence structure in the text is identified and divided into sentence groups. Then, the word segmentation tool is used to disassemble each sentence group into a set of independent terms. In this process, the minimum length of the term is set to two characters, and proper nouns such as names of people and places are eliminated. The low-frequency terms are filtered out by setting the term statistical threshold to terms with an appearance frequency lower than 0.3% of the total number of terms. For example, an instruction manual contains about 18,000 terms. The low-frequency word threshold is set to 54 times. Words with a statistical frequency lower than this value will be eliminated from the subsequent processing flow. Then, lexical analysis is performed on the remaining words. By analyzing the position of words in the sentence group and the collocation of the words before and after, a dependency matrix between the words is constructed. The measure of dependency strength is based on the frequency of co-occurrence of the words in the same semantic unit. The calculation method is: if word A and word B co-occur 25 times in 100 sentence groups, the co-occurrence frequency is 0.25. The dependency strength is set to the geometric mean of the frequency multiplied by the total number of sentence groups divided by the frequency of A or B in the sentence group. Assuming that A and B appear 40 and 50 times respectively, the dependency strength is calculated as In this way, a word dependency strength graph is constructed, and then similarity clustering is performed on the graph. The minimum coupling distance principle is used to divide multiple categories of terms into several cluster groups. The average dependency strength between terms in each cluster is higher than the set clustering threshold (set to 0.6). If the dependency strength is less than 0.6, the edge will be regarded as a weak connection and disconnected in the graph. Finally, several keyword coupling clusters are obtained. The average strength of the co-occurrence of terms between clusters is further calculated and normalized to the range of 0 to 1 to obtain the keyword correlation coefficient value. For example, the correlation between a group of keywords {sales, customer, quotation} is 0.78, while the correlation between another group {delivery, document, interface} is 0.43. The former group is closer and is used to describe the more strongly correlated functional clusters in the manual.

[0026] Table 1: Keyword co-occurrence statistics

[0027] As shown in Table 1, the dependency strength is calculated by counting the co-occurrence frequency of keywords in the text and combining it with their individual occurrence frequency, providing a basis for the subsequent construction of the word dependency graph and the keyword correlation coefficient value. The keyword correlation coefficient value quantifies the semantic correlation strength between keywords in the manual and characterizes the coupling relationship between keywords.

[0028] The functional hierarchy division submodule builds a word weight list based on the keyword relevance coefficient value, uses the hierarchical analysis method to build the word group comparison relationship, calculates the consistency ratio, and obtains the functional hierarchy weight coefficient distribution value; The functional hierarchy division submodule is processed based on the keyword correlation coefficient value. First, a word weight list is constructed, and the correlation coefficient value of each group of keywords is used as the correlation reference between the keywords in the group. The keyword pairs and their corresponding correlation coefficients are listed in sequence to form a two-dimensional array structure. For example, there are five groups of keyword pairs, and their correlation coefficients are 0.78, 0.43, 0.66, 0.52, and 0.81, respectively. These five values are used as the initial weights of the corresponding groups and input into the next processing process. Subsequently, the weight differences between the terms are analyzed, and groups with similar weights are extracted and word groups are constructed. If the correlation coefficient difference between any keyword in two adjacent word groups and any word in the other group is less than 0.15, it is determined that the two groups can be merged into the same functional set. Then, the word group comparison relationship construction operation is performed on the merged word groups, and the number of terms, average weight, and internal consistency between each pair of word groups are compared. The consistency ratio is defined as the ratio of the standard deviation of the internal correlation of the word group to the average value. If the ratio is less than 0.25, the word group is judged to have good consistency. When constructing the comparative relationship, the differences between any two word groups are compared one by one, and the average weight difference and the difference in the number of terms are calculated and judged with the preset threshold. The threshold is set to 0.2 and 5 items. If the average weight difference between the two word groups is less than 0.2 and the difference in the number of terms is less than 5, they are regarded as similar functional levels and merged into the same upper node. Then the functional groups after the word groups are merged are counted, and the association between their terms is summarized. The average value of the strength is normalized to the interval [0, 1] as the final weight coefficient of the functional level. For example, if a functional group contains 7 pairs of keywords, and their correlations are 0.76, 0.81, 0.68, 0.79, 0.73, 0.77, and 0.72, the average value is 0.76+0.81+0.68+0.79+0.73+0.77+0.72 / 7≈0.7529, and the standard deviation is about 0.0433. The consistency ratio is , because the ratio is lower than 0.25, it meets the consistency requirement, and finally the normalized value of 0.7529 of this group is used to construct the functional hierarchy, and the normalization coefficient of the functional hierarchy is set to the highest average value as the benchmark (for example, if the maximum value is 0.83, the current functional group weight is ), repeat the process for the functional groups, and finally form a complete functional hierarchy weight coefficient distribution value.

[0029] The node topology construction submodule calls the function level weight coefficient distribution value, sets the core function threshold and auxiliary function interval, builds a node type classification table, calculates the association strength between nodes based on the weight value, integrates the node pairs whose connection probability is higher than the set ratio in the probability table, and obtains the function node topology map; After calling the above-mentioned function level weight coefficient distribution value, the node topology construction submodule first sets a core function threshold and auxiliary function interval. Assuming that the weight coefficient of the current function group falls in the interval [0.65, 0.95], those with weight values greater than 0.85 are classified as core functions, those with weight values between 0.70 and 0.85 are classified as auxiliary function intervals, and those with weight values lower than 0.70 are regarded as edge functions and are not included in the core part of the topological structure. Then, the function nodes are classified according to the weight values, and a node classification table is set to divide the types into core nodes, auxiliary nodes, and edge nodes. Then, the correlation strength calculation between nodes is performed, and the edge weights of the node corresponding terms in the original dependency graph are called. The average strength value of the direct connection between nodes is calculated pairwise. For example, there are four edges between nodes A and B, with weights of 0.71, 0.85, and 0.91, respectively. .82, 0.76, 0.79, then the association strength is 0.71+0.82+0.76+0.79 / 4=0.77. If this value is higher than the connection strength threshold set by the system (set to 0.75), the edge connection relationship is retained, and then the original co-occurrence probability table of the node pair is consulted. If the proportion of the node pair appearing simultaneously in multiple instruction manual samples exceeds the set proportion (such as set to 0.4), it is regarded as a stable connection pair and added to the node connection matrix. Finally, based on the node connection matrix and combined with the above-mentioned weight classification, a functional node topology diagram is drawn, and the type of each node and its connection relationship with other nodes are marked. For example, the thickness of the marked edge represents the connection strength, and the node color represents its functional classification. Finally, a structured functional node network diagram is generated as an important structural basis for functional analysis.

[0030] Specifically, if Figure 2 As shown, the interactive parsing module includes: The node matching submodule obtains the functional node topology, detects the duration of node label matching, calculates the matching duration of each node based on the node label and network topology, records its corresponding data, and generates a matching duration value; The node matching submodule is used to obtain the topological structure diagram of the functional nodes. During execution, the functional nodes involved in the sales support system are first identified and numbered. For example, nodes A, B, C, D represent interactive touchpoints of differentiated customers, such as website visits, email interactions, product clicks, etc. Then, the connection information of the nodes is extracted from the system database, and its adjacency matrix is constructed and a topological structure diagram based on the sales channel logic is established. This process can generate a two-dimensional adjacency matrix such as the matrix by calling the node interaction record. , define whether the nodes are directly connected and the size of the weight. On this basis, by retrieving the label data of the nodes in differentiated time periods (such as customer interest keywords: price, brand, technical support, etc.), label matching analysis is performed on the connected node pairs (such as nodes A and B). During the execution process, each pair of node labels needs to be compared item by item, and a character comparison method (such as Jaccard similarity) is used to count the number of overlapping labels and calculate the matching score. Then, the detection time window is set in combination with the relative position of the nodes in the topological path (such as node distance, connection path weight). For example, it is set to the interaction data in the last 7 days. Then, based on the above matching score and the continuous interaction behavior in the time window, the label matching time between the nodes in each unit time period is gradually counted, and finally a three-dimensional array is formed. , where t is the time period, (i, j) is the node index, and the matching duration unit is mainly minutes. In the implementation process, if nodes A and B have 4 interactions between June 10 and June 16, 2025, each lasting 10, 12, 8, and 15 minutes and the label matching degree is higher than 70%, then the cumulative matching duration is 45 minutes. The matching duration is further normalized by Z-score, that is, for each group The data is deducted from the mean of the overall matching duration of the node and divided by the standard deviation to obtain dimensionless data. The standardized data is as follows For example, if the average matching time between node A and multiple nodes is 40 minutes and the standard deviation is 3 minutes, then its normalized value is , record the standardized matching time and save it to the system intermediate data pool to prepare for subsequent module calls.

[0031] The attention intensity calculation submodule calculates the attention intensity of the node based on the matching duration value through the entropy weight model, compares the label matching degree of the node, calculates the node attention intensity coefficient, and generates the node attention intensity value; The attention strength calculation submodule is based on the matching duration, and uses parameters such as structural weight, interaction frequency, label difference and feature collaboration to build a multi-dimensional indicator system to calculate the attention strength value of each node. In the specific execution process, the standardized matching duration value matrix is first read. , each element of which represents the matching strength between node i and node j at time t, and then the structural weight coefficient is extracted , which is set based on the product of the node's connectivity and connection weight in the topological structure. For example, the weight between nodes i and j is set to 0.8. If the average connectivity of node i with other nodes in the topological structure is 0.6, then =0.48, then read the number of interactions between nodes , assuming that the total number of interactions between nodes A and B in 7 days is 5 times, the system performs Z-score standardization on the number of node interactions and the value is 1.2. The next step is to read the degree of label difference , which is calculated as 1 minus the label overlap. For example, nodes A and B have 8 labels, of which 6 overlap. , characteristic synergy value The cosine similarity is calculated by constructing a vector based on the quantifiable behavioral features (such as click frequency and dwell time) in the customer behavior data. For example, the behavior vector of node A is (0.6, 0.8, 0.7), and that of node B is (0.5, 0.85, 0.6). The cosine similarity between the two is 0.998. =0.998, and By averaging the synergy values of node j and its adjacent nodes, if the synergy values of node B and its three adjacent nodes are 0.998, 0.932, and 0.887 respectively, then =0.998+0.932+0.887 / 3=0.939, calculate the attention intensity value, using the formula: ; in, represents the attention intensity value of node j at time t, Represents the matching duration between node i and node j at time t, which is normalized by Z-score and unified into dimensionless parameters. Represents the topological structure weight coefficient of node i and node j at time t, Represents the number of interactions between node i and node j at time t, normalized by Z-score, and unified into a dimensionless parameter Represents the degree of difference between the labels of node i and node j at time t, represents the feature synergy value of node i and node j at time t, represents the average value of the feature coordination value between node j and its adjacent nodes at time t, and n represents the number of nodes involved in the matching calculation.

[0032] Taking node B as an example, assume that there are three nodes connected to it, and the associated parameters are as follows: Table 2: Parameters for attention intensity calculation

[0033] As shown in Table 2, the calculation process is as follows: Numerator part = + + ≈ + + ≈0.3825+0.3468+0.2441=0.9734; ,The result shows that the attention intensity of node B in this time period is 1.027, ,which can be used as the basis for subsequent priority judgment.

[0034] The priority generation submodule establishes a node sorting sequence based on the node attention intensity value, sets the intensity fluctuation threshold and inventory response coefficient, and dynamically weights the nodes with intensity values higher than the threshold according to the response coefficient to generate a dynamic priority list; The priority generation submodule is based on the node attention strength value. First, the nodes are preliminarily sorted from large to small according to their strength values to form a node sequence list. For example, in a system, the strength values of nodes B, A, C, and D are 1.027, 0.985, 0.932, and 0.880, respectively. The initial sorting is [B, A, C, D]. Then, the strength fluctuation threshold setting content is read. This value is based on the average fluctuation range of the node in the past cycle, and ±20% is set as the upper and lower limits. If the average strength of node A in the past 5 days is 0.9, then its fluctuation upper limit is 0.9×1.2=1.08, and the lower limit is 0.9×0.8=0.72. When the attention strength of a node at the latest time point is greater than 1.08 or less than 0.72, it is judged that the strength has changed significantly, and then the inventory response coefficient is read. This coefficient is set between 0.3 and 0.8 based on the sensitivity of the company's inventory to sales strategy adjustments. For example, if the inventory warning level for a certain type of product is medium to high, the response coefficient is set to 0.6. Combined with the aforementioned sorting and strength value changes, the system will further weight the nodes whose attention intensity fluctuations exceed the threshold. The weighting method is: original attention intensity value × inventory response coefficient × fluctuation adjustment factor (default is 1.1). If the current intensity of node B is 1.027, the fluctuation is significant, and the inventory response coefficient is 0.6, then the weighted priority value is 1.027×0.6×1.1≈0.678. After updating the priority value of node B, the new priority list is obtained by re-sorting, such as adjusting [B, A, C, D] to [A, B, C, D], forming a dynamic priority sequence as the basis for the system's recommended processing order.

[0035] Specifically, if Figure 2 As shown, the inventory decision module includes: The pressure coefficient calculation submodule obtains the inventory turnover rate, inventory age distribution, and demand volatility parameters of the single product in the dynamic priority list, establishes a three-dimensional coordinate system to calculate the vector modulus, performs Z-score standardization conversion, and generates the single product inventory pressure coefficient; The pressure coefficient calculation submodule obtains the inventory turnover rate, inventory age distribution, and demand volatility parameters of each single product in the dynamic priority list. First, the inventory turnover rate is disassembled and specifically counted in the form of "sales quantity during the period / average inventory quantity". In the data of the past 30 days, the average daily sales of item A is 120 pieces and the average inventory is 300 pieces. Its inventory turnover rate is 120×30 / 300=12. Then the inventory age distribution is obtained, the inventory entry time is counted and classified according to the differentiated number of days. Assume that 40% of the inventory of item A in the current inventory has an entry time of less than 7 days and 30% has an entry time of 7 to 14 days. 20% are between 14 and 30 days, 10% are more than 30 days, then the inventory age vector can be represented as an array [0.4, 0.3, 0.2, 0.1]. Then calculate the demand volatility, which is the ratio of the standard deviation of the daily sales of the item in the past 30 days to the average daily sales. If the standard deviation is 35, then the demand volatility is 35 / 120≈0.29. Use the above three parameters as coordinate axes to construct a three-dimensional coordinate system, locate each item in the three-dimensional space and calculate the vector modulus. Let the three-dimensional vector be (x, y, z), that is, the inventory turnover rate is x=12, the inventory age standard deviation is converted to y=0.8, and the demand volatility is z=0.29, then the modulus is On this basis, Z-score standardization is performed. Assuming that the mean of the original commodity inventory pressure module is 10.2 and the standard deviation is 1.5, the standardized value of item A is Z=12.03-10.2 / 1.5≈1.22. In order to distinguish the degree of pressure, the Z-score value range is defined as: low pressure area [-1, 0), normal area [0, 1), high pressure area [1, 2), and overpressure area [2, ∞). Here, item A is in the high pressure area. When forming the inventory pressure coefficient, the Z-score value is output in floating point form and recorded as 1.22 for subsequent weight adjustment calculations.

[0036] The weight mark trigger submodule calls the single product inventory pressure coefficient, calculates the exponential moving average of the pressure value in the past seven days, and constructs a pressure change channel. When the real-time value breaks through the upper track of the channel and exceeds the pressure threshold, a weight trigger signal is generated; The pressure threshold is a critical standard for measuring whether inventory pressure is abnormally high. If it exceeds this value, intervention actions will be triggered. It is set differently based on category attributes, season, and life cycle factors. The weight tag triggers the submodule to call the inventory pressure coefficient, calculate the daily pressure value of the past seven days, and perform exponential moving average processing. Suppose the Z-score value of a certain product in the past 7 days is [1.22, 1.28, 1.35, 1.45, 1.48, 1.50, 1.55], and the exponential smoothing coefficient is set to =0.5, taking the first day's 1.22 as the initial value, the calculation formula is , perform the following steps: Day 2: ; Day 3: ; Day 4: ; Calculate sequentially until the 7th day, and we get It is about 1.488, and a pressure change channel is constructed. The minimum and maximum values of the EMA of the previous 6 days are set as the lower track 1.25 and the upper track 1.48. The current real-time pressure Z-score is 1.55. It is judged whether it breaks through the upper track and exceeds the pressure threshold. The threshold is set to 1.5. The current value is 1.55>1.48 and>1.5. If the conditions are met, a weight trigger signal is generated. The threshold is set according to differentiated categories. If it is a seasonal high-volatility category (such as down jackets), the threshold is raised to 1.8. If it is a daily consumer category (such as bottled water), the threshold is lowered to 1.2 to ensure that the intervention judgment is reasonable.

[0037] The list weight injection submodule parses the weight trigger signal, locates the first three nodes in the list, calculates the ratio of the pressure coefficient to the baseline value as the amplification factor, establishes a linear mapping relationship and writes the node attributes to generate a weighted priority list; After parsing the aforementioned weight trigger signal, the list weight injection submodule locates the first three nodes in the list and calls their current inventory pressure coefficient to compare with the preset benchmark value of the category. The benchmark value is the average Z-score of the category in the past year, which is set to 0.9. The Z-score values of the existing nodes A, B, and C are 1.55, 1.48, and 1.40 respectively. The corresponding ratios are A: 1.55 / 0.9≈1.72, B: 1.48 / 0.9≈1.64, and C: 1.40 / 0.9≈1.56 , taking the ratio as the amplification factor, define the linear mapping interval [1.0, 2.0] → weight adjustment range [0, 3], that is, set the linear mapping function y=3(x-1.0), and the weight increase value of A is 3×0.72=2.16. The final weight is set to the initial weight + 2.16. If the initial weight is 1.0, it is updated to 3.16, the weight of B is 2.92, and the weight of C is 2.68. They are written into the node attribute table respectively to form a weighted priority list, which is convenient for subsequent sorting and dynamic scheduling.

[0038] Table 3: Weighted priority node data table

[0039] As shown in Table 3, the node weights are differentially improved because the pressure values are significantly higher than the baseline values, which is ultimately reflected in the weighted priority ranking structure.

[0040] Specifically, if Figure 2 As shown, the policy compilation module includes: The weight parsing submodule parses the inventory weight tags in the weighted priority list, converts the weight percentage to a decimal floating point number, matches the type code to the capacity upper limit, and generates a weight feature vector. When the weight parsing submodule parses the inventory weight tag in the weighted priority list, it first extracts the original weight value and type coding information of each resource item in the list. For example, if the weight of an item is "60%" and the type code is "T05", the system converts "60%" into a decimal floating point number 0.6, and determines that the capacity upper limit of the corresponding resource type "T05" in the database is 120 units. Then, a weight feature vector is constructed. The vector is encoded as a triplet consisting of a floating point weight value, a corresponding capacity upper limit value, and a resource type index value. In the specific operation, if the list contains five resource items, the item conversion results are merged to form a five-dimensional feature. Vector group. For example, a priority list of {(60%, T05), (20%, T01), (10%, T03), (5%, T07), (5%, T02)} is converted to {(0.6, 120, 5), (0.2, 80, 1), (0.1, 100, 3), (0.05, 90, 70), (0.05, 60, 2)}. Based on this, the fields are called item by item, and the maximum allowable inventory capacity is found according to the resource type code. The weight field is converted to a floating point type according to the percentage system, and then numbered in order by resource type to construct the feature vector. Finally, a complete vector group is formed for subsequent conflict analysis.

[0041] The conflict detection submodule calls the weight feature vector, inputs it into the decision tree classifier, calculates the standard deviation of the priority level parameter, sets the node splitting threshold, analyzes the degree of dispersion through the Gini coefficient, calculates the conflict index deviation value, detects the over-limit node, and outputs the conflict index matrix; The decision tree classifier uses the input weight feature vector to classify the conflict situation into differentiated priority levels through node splitting and Gini coefficient calculation; After receiving the weight feature vector group, the conflict detection submodule first inputs it into the constructed decision tree classification structure. The structure makes layer-by-layer judgments in a node-splitting manner. First, the standard deviation of the priority level parameters in each group of weight vectors is calculated. For example, the level parameter set in a certain node is {4, 2, 3, 5, 1}, and its arithmetic mean is 3. The standard deviation calculation formula is: ; Determine whether the value exceeds the set threshold (such as 1.2). If so, the node needs to continue splitting. Then the system sets the node's splitting threshold to the current standard deviation and analyzes its Gini coefficient according to the following formula: ; in is the proportion of the kth priority level. For example, if the priority level frequency in a node is {1:1, 2:1, 3:1, 4:1, 5:1}, then the proportion of each category is 0.2. Substituting it into the Gini coefficient is: ; This value indicates that the priority distribution within the node is highly discrete. The system then calculates the conflict index deviation value according to the following formula: ; in, Represents the conflict index deviation value of the i-th node, Represents the number of weight items in the i-th node, Represents the j-th value of the j-th weight feature vector in the i-th node, Represents the upper limit of the capacity corresponding to the type code matched by the jth item in the i-th node, Represents the value of the j-th priority parameter in the i-th node, represents the arithmetic mean of the priority level parameters in the i-th node, represents the overall average value of the conflict index deviation value of the node, is the maximum capacity of the system, is a small constant to avoid zero, set to 0.0001.

[0042] Set to 200, Set to 0.0001, 1~5, is 3, Temporarily set it to 0.35.

[0043] Taking the five data items in a node as an example, the following table is constructed: Table 4: Conflict Index Calculation Parameters

[0044] As shown in Table 4, substitute into the formula for calculation.

[0045] Item 1: ; Item 2: ; Item 3: ; Item 4: ; Item 5: ; ; The final node conflict index deviation value is 0.2864. If the conflict deviation warning benchmark set by the system is 0.25, the node is marked as an abnormal node, and its number and deviation value are output to the conflict index matrix.

[0046] The parameter generation submodule calls the exception identifier in the conflict index matrix, establishes a linear relationship between priority and conflict index, sets the weighted formula of weight and conflict, calculates and sorts, and encapsulates it into an execution parameter package; The parameter generation submodule extracts the node entries marked as abnormal based on the conflict index matrix content, such as the node numbered " "The node whose conflict index deviation value is 0.2864 is greater than the warning baseline value of 0.25, will be included in the subsequent processing set, and the system will sort it according to the priority level. Corresponding conflict index value Establish a corresponding relationship and assign parameters in a linear weighted manner. For example, if the weight setting function is: ; in =0.6, =0.4, then for the first data: ; The other entries are calculated in the same way, and finally five weighted ranking values are obtained. The system arranges them from high to low according to these ranking values, readjusts the order of each node entry, and encapsulates the parameter combination of the node entry into an execution parameter package after numbering. The parameter package is transmitted to the subsequent scheduling module in the form of an array.

[0047] Specifically, if Figure 2 As shown, the terminal adaptation module includes: The session parameter parsing submodule parses the inventory adaptation speech parameters in the execution parameter package, extracts the session ID feature sequence, performs a bit-by-bit comparison with the device code library, selects entries with matching fault tolerance values, and generates a device location code. In the session parameter parsing submodule, the inventory adaptation speech parameters contained in the execution parameter package are first extracted. The extraction process traverses the key-value pair structure, reads the values associated with fields such as "speech type", "request timestamp", "device number" and "semantic target number", and performs character grouping processing on the extracted semantic target number and constructs a standardized vector structure to facilitate the subsequent identification of the directional context parameters of the speech style. Then, the session ID feature sequence attached to the speech parameter is obtained. The feature sequence is represented by 32-bit hexadecimal encoding. In order to enhance the accuracy of the parsing, the feature sequence needs to be encoded in 32 bits. The feature sequence is divided into four segments, each 8 bits long. Each segment is converted to ASCII and then modulo 37 to form four basic comparison sub-segments. Bit-by-bit comparison is performed with the corresponding fields in the device code library. This operation is based on the byte-level XOR logic operation. Each comparison value in the result is XORed and aggregated to generate the difference vector of the current entry. The difference is less than or equal to "3" as the fault tolerance value standard during the screening process. Only when at least three of the four comparison sub-segments meet the fault tolerance constraint, the entry is considered to meet the screening conditions. For example, assuming the original session ID is: "ABCD1234EFGH5678IJKL9012MNOP3456" is split into four parts: "ABCD1234", "EFGH5678", "IJKL9012", and "MNOP3456". Character-level ASCII mapping is performed on the first part to obtain [65, 66, 67, 68, 49, 50, 51, 52]. After modulo 37, it is [28, 29, 30, 31, 12, 13, 14, 15], perform bit-by-bit XOR calculation on the array and the transcoded array of the device code "ADCF1832" to generate a comparison difference vector [0, 2, 2, 1, 1, 3, 2, 2]. The average difference is calculated to be 1.625, which is lower than the set fault tolerance threshold of 3, so the entry is selected into the candidate pool; after completing the four-segment comparison, the one with 3 or more qualified difference segments is judged as a matching entry, and finally sorted in ascending order by the average difference value, and the first candidate is selected as the device positioning code.

[0048] The submodule for updating the speech field parses the interaction frequency matrix of the device location code, calculates the covariance coefficient between the field click rate and inventory turnover rate, establishes a priority sorting vector, updates the field version hash value, and generates a speech version identifier. In the speech field update submodule, first read the corresponding interaction frequency matrix from the device semantic interaction database based on the device positioning code generated in the previous step. The matrix rows are device field identifiers, the columns are standard time granularity (such as hours), and the matrix values are click frequency records. For example, the click frequency of field "F001" is 15 times in the "2025-06-1708:00" period and 12 times in the "2025-06-1709:00" period. Construct a two-dimensional array with the column frequency and its inventory turnover record. In this array, each column represents the frequency-turnover pair value in a time period. Calculate the covariance between the click frequency vector and the turnover rate vector by performing the following steps: Based on the total sample number n=24 (assuming 24 time points in a day are calculated by hour), first calculate the frequency mean , turnover mean , and then calculate the covariance value: ; Taking field F001 as an example, its click frequency within a day is [15, 12, 9, 13, 10, ...] (a total of 24 items), and its turnover rate is [1.2, 1.0, 0.9, 1.1, 1.0, ...]. Calculated using the above formula, the covariance is 0.38. Fields are sorted by covariance value. A higher covariance indicates a stronger positive correlation between click frequency and turnover rate. Thus, a priority sorting vector such as ["F003", "F001", "F005", ...] is constructed. The current field hash value is then compared with the hash value sequence of the previous version record. If the field content changes due to a field sort update or the addition of a new field, a new version hash identifier is regenerated. The SHA256 algorithm is used to concatenate and digest the field set, and finally the new speech version identifier is output.

[0049] The interaction log generation submodule extracts the change records of the speech field corresponding to the speech version identifier, captures the device response delay difference, counts the field click trigger frequency, and encodes the operation type code, response status code, and timestamp into a triplet to generate the interaction log record. In the interaction log generation submodule, the script version identifier generated in step S502 is first structurally parsed to identify the corresponding field change records. By comparing the differences in the position and existence status of the field in the current version and the previous version, operation type codes such as "add", "delete", and "position change" are recorded. The response timestamp of the device after the field click operation is extracted from the system response link, and the response delay is calculated as: ; For example, if the click time is 10:03:12.504 and the response time is 10:03:12.917, the response delay is 0.413 seconds. The average response delay is calculated by accumulating the response difference of this field within 24 hours. , count the total number of clicks on the field and record it, obtaining a triplet format (operation type code, response status code, timestamp). The response status code is encoded according to the HTTP or TCP result, and the values are "200", "404", "500", etc. The combination form is such as (INS, 200, 10:03:12.504). A separate log entry is generated for each click in the record to track the change process of the device response status. If a field is clicked 37 times in a day and the average response delay is 0.342 seconds, there will be 37 triplet entries in the record, each corresponding to a specific click behavior and response situation. Finally, the log collection is packaged into a JSON structure and written to the database.

[0050] Table 5: Sample table of device interaction records

[0051] As shown in Table 5, some interaction records of a device on a certain day are listed. The record content includes field operation type, click and response time and its delay. The response status is used to judge the integrity and stability of device interaction.

[0052] See also Figure 3 The sales assistance method based on AI technology is executed based on the above-mentioned sales assistance system based on AI technology, and includes the following steps: S1: Extract the keywords of the instruction manual through natural language processing technology, divide them into core function nodes, auxiliary function nodes, and extended function nodes using the hierarchical analysis method, and establish a function node topology map; S2: Count the dwell time of labels in the topology graph of functional nodes, combine the node weights and input the entropy weight model to calculate the attention intensity value and generate a dynamic priority list; S3: Receives a dynamic priority list, calculates the inventory pressure coefficient of a single product, and adds inventory weight tags to the first three nodes in the list when the pressure coefficient exceeds a preset threshold to generate a weighted priority list. S4: Parse the inventory weight tags in the weighted priority list, perform conflict detection and priority sorting on the tag status through a decision tree classifier, and generate an execution parameter package; S5: Parse the inventory adaptation script parameters in the execution parameter package, locate the terminal device according to the session ID code, update the inventory guidance script field in the recommendation column, and generate an interaction log record.

[0053] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A sales support system based on AI technology, characterized in that: The system comprises: The function modeling module is used to extract the keywords of the manual through natural language processing technology, divide them into core function nodes, auxiliary function nodes, and extended function nodes using the hierarchical analysis method, establish the function node topology map, and pass it to the interactive analysis module; An interactive parsing module is used to match the functional node topology, calculate the label matching time, calculate the node attention intensity through the entropy weight model, generate a dynamic priority list, and pass it to the inventory decision module; An inventory decision module is configured to receive the dynamic priority list, calculate the inventory pressure coefficient of a single product, add inventory weight tags to the first three nodes in the list when the pressure coefficient exceeds a pressure threshold, generate a weighted priority list, and pass it to the strategy compilation module; A policy compilation module is used to parse the inventory weight tags in the weighted priority list, perform conflict detection and priority sorting on the tag states through a decision tree classifier, generate an execution parameter package, and pass it to the terminal adaptation module; The terminal adaptation module is used to parse the inventory adaptation speech parameters in the execution parameter package, locate the terminal device according to the session ID code, update the inventory guidance speech field in the recommendation column, and generate an interaction log record.

2. The AI-based sales support system according to claim 1, characterized in that: The functional node topology diagram includes the core functional node hierarchy, auxiliary functional node association paths, and extended functional node topology relationships; the dynamic priority list includes the node attention strength value, entropy weight coefficient, and time sequence priority index; the weighted priority list includes the inventory weight tag identifier, inventory turnover rate threshold parameter, and priority node identifier; the execution parameter package includes the decision tree splitting threshold, execution priority queue, and conflict detection status code; the interaction log record includes the speech version number, device UUID mapping table, and recommendation field update log; The entropy weight coefficient is the objective weight value of the indicator obtained by calculating the information entropy.

3. The AI-based sales support system according to claim 1, characterized in that: The functional modeling module includes: The keyword extraction submodule obtains the instruction manual text content, identifies paragraph sentences and word structures, filters out low-frequency words, calculates word dependency strength based on word syntactic collocation and co-occurrence frequency, constructs a word dependency graph, applies a clustering algorithm to calculate the coupling index between words, and generates a keyword relevance coefficient value; The functional hierarchy division submodule constructs a word weight list based on the keyword relevance coefficient value, uses the hierarchical analysis method to construct a word group comparison relationship, calculates the consistency ratio, and obtains the functional hierarchy weight coefficient distribution value; The node topology construction submodule calls the functional hierarchy weight coefficient distribution value, sets the core function threshold and auxiliary function interval, constructs a node type classification table, calculates the association strength between nodes based on the weight value, integrates the node pair relationships with connection probabilities higher than the set ratio in the probability table, and obtains a functional node topology map.

4. The AI-based sales support system according to claim 3, characterized in that: The interactive parsing module includes: The node matching submodule obtains the functional node topology map, detects the duration of node label matching, calculates the matching duration of each node based on the node label and network topology, records the corresponding data, and generates a matching duration value; The attention intensity calculation submodule calculates the attention intensity of the node based on the matching duration value through the entropy weight model, compares the label matching degree of the node, calculates the intensity coefficient of the node attention, and generates the node attention intensity value; The priority generation submodule establishes a node sorting sequence according to the node attention strength value, sets a strength fluctuation threshold and an inventory response coefficient, dynamically weights nodes with strength values higher than the threshold according to the response coefficient, and generates a dynamic priority list.

5. The AI-based sales support system according to claim 4, characterized in that: The attention intensity value is calculated using the formula: ； in, represents the attention intensity value of node j at time t, Represents the matching duration between node i and node j at time t, which is normalized by Z-score and unified into dimensionless parameters. Represents the topological structure weight coefficient of node i and node j at time t, Represents the number of interactions between node i and node j at time t, which is normalized by Z-score and unified into a dimensionless parameter. Represents the degree of difference between the labels of node i and node j at time t, represents the feature synergy value of node i and node j at time t, It represents the average value of the feature synergy between node j and its adjacent nodes at time t, n represents the number of nodes involved in the matching calculation, and the variable is normalized by Z-score and unified into a dimensionless parameter.

6. The AI-based sales support system according to claim 4, characterized in that: The inventory decision module includes: The pressure coefficient calculation submodule obtains the inventory turnover rate, inventory age distribution, and demand volatility parameters of the single product in the dynamic priority list, establishes a three-dimensional coordinate system to calculate the vector modulus, performs Z-score standardization conversion, and generates the single product inventory pressure coefficient; The weight mark trigger submodule calls the single product inventory pressure coefficient, calculates the exponential moving average of the pressure value in the past seven days, and constructs a pressure change channel. When the real-time value breaks through the upper track of the channel and exceeds the pressure threshold, a weight trigger signal is generated; The pressure threshold is a critical standard for measuring whether inventory pressure is abnormally high. If it exceeds this value, intervention will be triggered. It is set differently based on category attributes, season, and life cycle factors. The list weight injection submodule parses the weight trigger signal, locates the first three nodes in the list, calculates the ratio of the pressure coefficient to the reference value as the amplification factor, establishes a linear mapping relationship and writes the node attributes to generate a weighted priority list.

7. The AI-based sales support system according to claim 6, characterized in that: The policy compilation module includes: A weight parsing submodule parses the inventory weight tags in the weighted priority list, converts the weight percentages into decimal floating point numbers, matches the type code to the capacity upper limit, and generates a weight feature vector; The conflict detection submodule calls the weight feature vector, inputs it into the decision tree classifier, calculates the standard deviation of the priority level parameter, sets the node splitting threshold, analyzes the degree of dispersion through the Gini coefficient, calculates the conflict index deviation value, detects the over-limit node, and outputs the conflict index matrix; The decision tree classifier uses the input weight feature vector to classify the conflict situation into differentiated priority levels through node splitting and Gini coefficient calculation; The parameter generation submodule calls the exception identifier in the conflict index matrix, establishes a linear relationship between priority and conflict index, sets a weighted formula for weight and conflict, calculates and sorts, and encapsulates it into an execution parameter package.

8. The AI-based sales support system according to claim 7, characterized in that: The conflict index deviation value is calculated using the formula: ； in, Represents the conflict index deviation value of the i-th node, Represents the number of weight items in the i-th node, Represents the j-th value of the j-th weight feature vector in the i-th node, Represents the upper limit of the capacity corresponding to the type code matched by the jth item in the i-th node, Represents the value of the j-th priority parameter in the i-th node, represents the arithmetic mean of the priority level parameters in the i-th node, represents the overall average value of the conflict index deviation value of the node, is the maximum capacity of the system, is a small constant to avoid zero, set to 0.0001.

9. The AI-based sales support system according to claim 7, characterized in that: The terminal adaptation module includes: The session parameter parsing submodule parses the inventory adaptation speech parameters in the execution parameter package, extracts the session ID feature sequence, performs a bit-by-bit comparison operation with the device code library, selects entries with matching fault tolerance values, and generates a device positioning code; The speech field update submodule parses the interaction frequency matrix of the device location code, calculates the covariance coefficient between the field click rate and the inventory turnover rate, establishes a priority sorting vector, updates the field version hash value, and generates a speech version identifier; The interaction log generation submodule extracts the speech field change record corresponding to the speech version identifier, captures the device response delay difference, counts the field click trigger frequency, encodes the operation type code, response status code and timestamp into triples, and generates an interaction log record.

10. A sales assistance method based on AI technology, characterized in that: The method is used to implement the AI-based sales assistance system according to any one of claims 1 to 9, comprising the following steps: S1: Extract the keywords of the instruction manual through natural language processing technology, divide them into core function nodes, auxiliary function nodes, and extended function nodes using the hierarchical analysis method, and establish a function node topology map; S2: Count the residence time of the labels in the functional node topology graph, combine the node weights and input the entropy weight model to calculate the attention intensity value and generate a dynamic priority list; S3: Receives the dynamic priority list, calculates the inventory pressure coefficient of each item, and adds inventory weight tags to the first three nodes in the list when the pressure coefficient exceeds a preset threshold, thereby generating a weighted priority list. S4: parsing the inventory weight tags in the weighted priority list, performing conflict detection and priority sorting on the tag states through a decision tree classifier, and generating an execution parameter package; S5: Parse the inventory adaptation speech parameters in the execution parameter package, locate the terminal device according to the session ID code, update the inventory guidance speech field in the recommendation column, and generate an interaction log record.

Images