An enterprise AI product intelligent selection system based on a large model

Abstract

The application discloses an enterprise AI product intelligent selection system based on a large model, and belongs to the cross field of artificial intelligence application and enterprise digital management. The application maps the unstructured demand description of an enterprise user into a structured selection constraint condition by constructing a multi-dimensional product knowledge graph and a demand semantic understanding model, and combining the semantic understanding and reasoning capability of a large language model, and realizes intelligent matching and ordering recommendation of AI products by combining a multi-objective optimization algorithm. The system comprises a demand semantic analysis module, a product knowledge graph construction module, a constraint condition mapping module, a multi-dimensional evaluation scoring module and an explainable recommendation output module. The application realizes an automatic selection process from a fuzzy demand to accurate recommendation, reduces the time cost and decision risk of enterprise AI product selection, and can be widely applied to enterprise software procurement, SaaS product selection, cloud service scheme comparison and the like.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology application, specifically involving the application of large language model technology in enterprise digital product selection scenarios, and more specifically, involving an enterprise AI product intelligent selection system based on large models. Background Technology

[0002] In recent years, artificial intelligence (AI) technology has moved from the laboratory to large-scale industrial applications, and the AI ​​product market is characterized by its diverse categories, rapid iteration, and significant architectural differences. According to industry statistics, there are currently over a thousand manufacturers in China providing AI-related products and services, covering dozens of sub-sectors including machine learning development platforms, large-model API services, intelligent customer service systems, computer vision solutions, automatic speech recognition systems, and AI data annotation and management platforms. However, the technical architectures of AI products differ significantly. From underlying computing chips to deep learning frameworks, from model deployment methods to data privacy protection mechanisms, different products exhibit fundamental differences in technical routes, performance indicators, cost structures, and ecosystem compatibility. When selecting AI products, companies typically form decision-making teams composed of business personnel, technical architects, and the procurement department. However, business personnel often lack professional understanding of technical parameters, technical architects lack comprehensive knowledge of the fine-grained characteristics of various vendors' products, and the procurement department struggles to accurately estimate hidden costs, resulting in a significant knowledge asymmetry among these three groups. This information asymmetry and cognitive differences directly lead to long AI product selection cycles, low decision-making quality, and high procurement risks for enterprises. Industry research indicates that medium-sized enterprises typically spend three to six months completing a full AI product selection process, with approximately 60% of that time consumed in the requirements analysis and product information gathering phases, rather than actual solution comparison and decision-making. Therefore, leveraging artificial intelligence technology itself to address information asymmetry and decision-making efficiency issues during the AI ​​product selection process has become a critical technological bottleneck that enterprises urgently need to overcome in their digital transformation.

[0003] To address the aforementioned selection difficulties, several technical solutions have been proposed and attempted, but each solution has shortcomings in its core technical approach. The first type of solution is based on keyword retrieval and form filtering. Its technical principle essentially relies on inverted indexes and Boolean matching models. Users need to input keywords or fill out structured forms to obtain a list of candidate products. The matching effect heavily depends on literal overlap, resulting in the omission of many semantically matching but literally mismatched candidate products. Furthermore, it is completely unable to handle qualitative descriptions or implicit constraints. The second type of solution is based on rule engines, where domain experts predefine a selection rule base. Users fill in parameters item by item to trigger rule matching. Its core technology is a generative rule system. The drawback of this type of solution is that the rule base maintenance cost increases exponentially with product iteration. Hard threshold logic cannot handle flexible constraints with fuzzy quantifiers and conditional preferences. Moreover, when user needs involve new product characteristics, the rule base lacks corresponding attribute definitions, causing the system to fail directly. The third type of solution uses large language models for general recommendations. Its advancement lies in the introduction of semantic understanding capabilities. However, its recommendations are based on algorithmic model components, and the evaluation is primarily based on performance metrics on benchmark datasets. In contrast, enterprise AI product selection involves the entire commercial product solution, with evaluation dimensions extending far beyond algorithmic performance. These dimensions include enterprise-level non-functional factors such as business terms, deployment and maintenance, data compliance, and ecosystem integration. The decision-making logic for the two is fundamentally different. Furthermore, the evaluation methods used in existing solutions are based on objective performance assessments of benchmark data, rather than semantic matching and multi-objective optimization based on the personalized needs of enterprise users. This fails to answer the core decision-making question of which product is most suitable for a specific business scenario.

[0004] In summary, the common, unresolved technical shortcomings of the three existing technical solutions can be categorized into three levels. On the input side, existing solutions generally require users to input selection criteria in a structured manner, failing to handle the most natural, unstructured expressions of enterprise users' needs. This forces the selection process to expend significant manpower on requirement engineering transformation during the requirements analysis phase. On the processing side, existing solutions either employ flat keyword indexes, sacrificing the structured semantics of multi-dimensional product attributes, or rely on a limited set of manually predefined attributes, lacking scalability and failing to flexibly cover emerging product types as the AI ​​product market rapidly evolves. Regarding matching mechanisms, existing solutions generally lack logical modeling for parallel reasoning of hard constraints and soft preferences. On the output side, enterprise AI product selection is a high-decision-cost scenario, but existing solutions often output results in a black-box manner, failing to provide a traceable path and structured explanation of the decision-making basis, severely limiting the credibility and adoption rate of selection results in actual procurement decisions. Summary of the Invention

[0005] To address the aforementioned shortcomings of existing technologies, the technical problem this invention aims to solve is: how to leverage the semantic understanding and reasoning capabilities of large language models to automate the entire process from unstructured natural language demand descriptions by enterprise users to intelligent AI product recommendations, and to introduce an enterprise AI product intelligent selection system based on a large model.

[0006] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: A large-scale model-based intelligent selection system for enterprise AI products includes: The requirement semantic parsing module is used to receive natural language requirement description text input by enterprise users, perform semantic parsing on the requirement text through a large language model, and extract structured selection information. The structured selection information includes at least product category tags, explicit constraint set, implicit constraint set obtained by reasoning, and priority weight of each constraint. The constraint mapping module, connected to the requirement semantic parsing module, is used to classify each constraint item in the structured selection information into hard constraints and soft constraints, and generate executable product screening conditions and scoring function parameters. The product knowledge graph retrieval module is connected to the constraint mapping module. It is used to perform retrieval and matching in the pre-built AI product knowledge graph based on the hard constraints, filter out the candidate product set that meets all hard constraints, and trigger a relaxation strategy when the candidate product set is empty. The multi-dimensional evaluation and scoring module is connected to the product knowledge graph retrieval module. It is used to calculate the scores of four dimensions for each product in the candidate product set: semantic matching degree, technical performance score, adaptability score and cost-effectiveness score, and calculate the comprehensive weighted score according to the weight parameters of the soft constraints. The interpretable recommendation output module is connected to the multi-dimensional evaluation and scoring module. It is used to sort the candidate products in descending order according to the comprehensive weighted score to generate a recommendation list and output the recommendation results including recommendation ranking, sub-dimensional score details and natural language explanation. Specifically, the semantic parsing module, when performing semantic parsing using a large language model, employs a chain-like prompting strategy to decompose the semantic parsing task into three sub-tasks: intent recognition, entity extraction, and implicit constraint reasoning. The implicit constraint reasoning sub-task, based on the common-sense reasoning capabilities of the large language model, completes the selection constraints not explicitly mentioned by the user. The constraint mapping module calculates the weight parameters for each scoring dimension for soft constraints containing fuzzy quantifiers or preference descriptions through a semantic similarity-guided weight allocation mechanism. The multi-dimensional evaluation and scoring module uses a weighted TOPSIS multi-objective optimization algorithm to calculate a comprehensive weighted score, thus forming a fully automated selection architecture from unstructured requirement input to interpretable recommendation output.

[0007] This invention, through the organic integration of a demand semantic parsing module, a constraint mapping module, a product knowledge graph retrieval module, a multi-dimensional evaluation and scoring module, and an interpretable recommendation output module, for the first time constructs a fully automated selection architecture that extends from the input of unstructured natural language requirements from enterprise users to the output of executable procurement decisions. The modules collaborate in a pipeline manner through standardized data interfaces. The demand parsing results directly drive constraint mapping, the constraint mapping output directly guides knowledge graph retrieval, the retrieval results are directly input into multi-dimensional evaluation, and the evaluation results are directly sent to the interpretable output module. The entire process requires no manual intervention. Specifically, the limited chain-hint semantic parsing strategy solves the technical challenge of automatically converting unstructured requirements into structured constraints; the semantic similarity-guided soft constraint weight allocation mechanism solves the technical challenge of fuzzy preference quantification; and the weighted TOPSIS multi-objective optimization algorithm solves the technical challenge of multi-dimensional comprehensive optimal ranking. The synergistic effect of these three key technologies enables the system to achieve the desired technical performance throughout the selection cycle.

[0008] In the optimized semantic parsing module, the intent recognition subtask determines the AI ​​product category to which the user's requirement belongs, the entity extraction subtask extracts quantitative parameters and conditional descriptions from the requirement text, and the implicit constraint reasoning subtask automatically infers implicit selection conditions such as data scale, concurrent requirements, and security compliance levels based on the user's requirement context. This solution decomposes the complex semantic parsing task into three relatively independent subtasks—intent recognition, entity extraction, and implicit constraint reasoning—for separate execution. This reduces the cognitive load of a large language model in a single inference, improving the accuracy and stability of semantic parsing. Furthermore, it enables the system to complete implicit selection conditions not explicitly mentioned by the user, enhancing the completeness and intelligence of requirement parsing.

[0009] The optimized AI product knowledge graph includes at least product nodes, manufacturer nodes, function nodes, technical indicator nodes, deployment method nodes, and industry case nodes as its node types; and at least membership relationships, possession relationships, indicator value relationships, support relationships, and applicable relationships as its relation types. The knowledge graph is stored in a graph database and dynamically updated and maintained through a combination of automated data collection and manual review. This solution uses a multi-dimensional product knowledge graph stored in a graph database to represent product functional attributes, technical indicators, deployment methods, industry cases, and other information in a fine-grained structure using nodes and relational edges. Compared to traditional flat keyword indexes, it supports more complex graph traversal queries and semantic association reasoning. Through a dynamic update mechanism combining automated data collection and manual review, the system can maintain the timeliness and completeness of knowledge in line with the rapid iteration of the AI ​​product market, overcoming the bottleneck of rule engines requiring manual maintenance of the rule base, and enabling the system to flexibly expand to emerging product types.

[0010] In the optimized constraint mapping module, constraints containing precise numerical thresholds or uncompromising condition descriptions are classified as hard constraints, which are used to perform Boolean logic filtering in the knowledge graph; constraints containing fuzzy quantifiers, preference descriptions, or conditional modifications are classified as soft constraints. The semantic similarity-guided weight allocation mechanism is expressed as follows: w i = Where C is the set of structured constraints, desc i Let sim( be the semantic description embedding vector for the i-th rating dimension) , ) is the cosine similarity function, and λ is the temperature coefficient.

[0011] The above scheme precisely classifies constraints into hard and soft constraints, enabling the system to rigorously filter uncompromising conditions using Boolean logic during the knowledge graph retrieval stage, and flexibly express preference requirements using continuous weight parameters during the scoring stage. This achieves the separation and synergy between rigid screening and flexible optimization. The defined semantic similarity-guided weight allocation formula can automatically calculate weight parameters based on the similarity between the user's requirement text and the semantic descriptions of each scoring dimension. This allows fuzzy preferences to be precisely quantified into weight coefficients for different dimensions in the scoring function, thus solving the technical challenge of existing rule engines being unable to handle flexible constraints with fuzzy quantifiers and conditional modifications.

[0012] In the optimized multi-dimensional evaluation and scoring module, the semantic matching degree is calculated using dense vector embedding generated by a large language model, and the calculation formula is as follows: Match sem (p,T)= Among them, E product (p) is the vector obtained by embedding the complete description text of product p through the embedding model mapping, E query (T) is the query vector obtained by mapping the user's request text T through the same embedding model.

[0013] The above solution uses dense vector embedding generated by a large language model to replace the traditional TF-IDF or BM25 sparse vector representation. By mapping the complete product description text and the user requirement text to the same continuous vector space and calculating the cosine similarity, the system can identify semantically equivalent but literally completely different relationships. This deep semantic matching capability significantly improves the selection recall rate compared to traditional keyword retrieval solutions, avoiding the problem of missing relevant products due to differences in terminology.

[0014] The optimized multi-dimensional evaluation and scoring module uses the weighted TOPSIS multi-objective optimization algorithm to calculate the comprehensive weighted score. Specifically, this includes: vector normalization of the original scores of each product in the candidate product set across four dimensions: semantic matching degree, technical performance score, suitability score, and cost-effectiveness score; weighting the scores of each dimension using the weight parameters of the soft constraints; determining the positive and negative ideal solutions for each dimension; calculating the distance of each candidate product to the positive and negative ideal solutions; and using the relative proximity as the final comprehensive score. This scheme uses the TOPSIS method to determine the relative proximity as the comprehensive score by calculating the distance of each candidate product to the positive and negative ideal solutions. This enables a scientific compromise ranking among multiple evaluation dimensions that may have mutually restrictive relationships. For example, product A may have the best technical performance but a higher price, while product B may have the best cost-effectiveness but slightly lower performance. The TOPSIS method can objectively determine which product is closer to the user's ideal solution overall based on the user's set weight preferences. Compared to the simple weighted summation method, the TOPSIS method considers the relative advantages and disadvantages of indicators in each dimension, avoiding the irrationality caused by directly adding indicators of different dimensions, making the recommendation ranking results more scientific, objective and stable.

[0015] In the optimized product knowledge graph retrieval module, the relaxation strategy includes: when the candidate product set satisfying all hard constraints is empty, the system identifies the constraints that lead to zero results, relaxes or removes them one by one according to priority from low to high, until the candidate product set is not empty, and marks the relaxed or removed constraints and returns them to the user along with a prompt message. In this invention, when strict hard constraints cause the candidate product set to be empty, the system does not simply return the invalid message "no results," but automatically identifies the constraints that lead to zero results and relaxes them one by one according to priority from low to high, until a valid candidate set is generated. This relaxation mechanism simulates the thought process in human selection decisions and clearly marks the relaxed constraints to inform the user. This mechanism avoids the zero-result dilemma caused by overly strict initial user requirements, improving the system's usability and user experience.

[0016] In the optimized, interpretable recommendation output module, the natural language explanation is automatically generated by a large language model based on the matching relationship between the recommendation results and user needs. The detailed sub-dimension scores are visualized in the form of radar charts or bar charts. The recommendation results support the generation of a horizontal comparative analysis report of two to three candidate products. This solution provides a complete decision-making basis chain for enterprise selection teams by simultaneously outputting recommendation rankings, detailed sub-dimension scores, radar chart or bar chart visualizations, product comparison analysis tables, and natural language explanations. Unlike the black-box output method of existing solutions, the output results of this invention have complete traceability. The traceable recommendation basis makes the system output results feasible for formal adoption in the enterprise procurement process, thereby enhancing the practical value of the recommendation results.

[0017] The optimized AI product knowledge graph supports an active learning mechanism based on user selection decision feedback. After the system records the user's final selection decision, it automatically adjusts the temperature coefficient and product rating parameters in the semantic similarity-guided weight allocation mechanism based on the decision feedback to achieve continuous optimization of the rating model. This system can automatically optimize the rating model parameters based on the feedback data of the user's final selection decision, forming a closed-loop learning mechanism of "recommendation → decision → feedback → optimization". As the number of times the system is used increases, the rating model can gradually learn the selection preference patterns of enterprises of different industries and sizes, so that the accuracy of subsequent recommendations continues to improve. This avoids the performance solidification problem caused by the lack of adaptive ability in traditional systems. This active learning mechanism is one of the important features that distinguishes this invention from static rule engine systems, enabling the system to continuously evolve with the accumulation of data.

[0018] The optimized large language model adopts a general-purpose large model based on the Transformer architecture, integrated into the requirement semantic parsing module and the interpretable recommendation output module via API. The requirement semantic parsing module is configured with system prompts containing selection domain knowledge, output format constraints, and inference step guidance instructions. This invention integrates a general-purpose large model based on the Transformer architecture via API, enabling the system to directly utilize the semantic understanding and text generation capabilities of state-of-the-art large language models without needing to train domain-specific models from scratch, thus reducing system development costs and deployment difficulty. The system prompts configured in the requirement semantic parsing module, containing selection domain knowledge, output format constraints, and inference step guidance instructions, guide the large model to output more standardized, accurate, and complete results in the vertical domain of selection, reducing the illusions and deviations common in open-domain applications of large models, and ensuring the output quality and stability of the system in large-scale commercial deployments.

[0019] Compared with the prior art, the beneficial effects of the present invention are: 1. Compared with existing technologies, traditional selection methods are limited by structured input paradigms. Enterprise users must manually convert vague business requirements into keywords or form options before they can use the system. This process not only consumes a lot of manpower and time, but also inevitably introduces information distortion due to the difference in knowledge background between business personnel and technical personnel. This invention, by constructing a semantic parsing engine based on a large language model, combined with a chain-based prompt decomposition strategy and implicit constraint reasoning mechanism, can directly use the unstructured requirement description of enterprise users as system input, automatically completing intent recognition, entity extraction, and implicit condition completion. This allows vague expressions to be accurately parsed into a set of structured selection constraints. This technological breakthrough at the input end compresses the selection cycle to the minute level, while eliminating the obstacle that business personnel cannot accurately express their technical requirements due to insufficient professional knowledge.

[0020] 2. This invention constructs a dynamically expandable AI product knowledge graph, using a graph structure to represent product functional attributes, technical indicators, deployment methods, industry cases, and other information in a fine-grained manner. It also achieves deep semantic matching between requirement text and product descriptions through semantic vector embedding, enabling the system to identify semantically equivalent but literally different relationships, thus improving the recall and accuracy of product selection. More importantly, this invention designs a dual-track recommendation algorithm that combines hard-constraint logic filtering with soft-constraint weight optimization, coupled with weighted TOPSIS multi-objective optimization. This achieves multi-dimensional comprehensive optimal recommendation decisions while ensuring that key conditions are not violated, overcoming the binary dilemma of traditional solutions either missing the optimal solution or making overly broad recommendations.

[0021] 3. This invention constructs a complete interpretable recommendation output mechanism, providing quantitative scoring radar charts for each evaluation dimension and product comparison analysis tables while outputting recommendation rankings. Furthermore, a large language model automatically generates customized natural language decision-making basis explanations, enabling decision-making teams to clearly trace the scoring source and logical basis of each recommendation, thus improving the credibility and adoption rate of recommendation results in multi-party joint reviews. Simultaneously, the "semantic understanding-knowledge reasoning-multi-objective optimization-interpretable output" full-link technology paradigm pioneered by this invention also has significant reference value and promotional value for a wider range of digital decision-making scenarios, such as enterprise software procurement and cloud service solution comparison. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall architecture of the enterprise AI product intelligent selection system based on a large model according to the present invention; Figure 2 This is a schematic diagram of the workflow of the requirement semantic parsing module in this invention; Figure 3 This is a schematic diagram of the AI ​​product knowledge graph structure in this invention; Figure 4This is a logic diagram of the constraint mapping module in this invention. Figure 5 This is a flowchart of the knowledge graph retrieval and relaxation strategy in this invention; Figure 6 This is a diagram of the scoring system for the multi-dimensional evaluation and scoring module in this invention; Figure 7 This is a schematic diagram illustrating the semantic matching degree calculation principle in this invention; Figure 8 This is an example diagram of the interpretability recommendation output module in this invention; Figure 9 This is a flowchart illustrating the overall selection method in this invention. Detailed Implementation

[0023] 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.

[0024] like Figures 1-9 As shown, this invention provides an enterprise AI product intelligent selection system based on a large model. This system adopts a modular, layered architecture design, taking unstructured natural language descriptions of enterprise users' needs as input and interpretable recommendation results containing complete decision-making criteria as output, constructing a fully automated selection solution from needs understanding to recommendation decision-making. The system consists of five core modules: a needs semantic parsing module, a constraint mapping module, a product knowledge graph retrieval module, a multi-dimensional evaluation and scoring module, and an interpretable recommendation output module. These modules collaborate in a pipeline-like manner through standardized data interfaces.

[0025] The requirement semantic parsing module, serving as the system's input, is responsible for transforming enterprise users' free text-based selection requirements into machine-processable structured data. Taking natural language requirement descriptions as input, the module performs multi-level semantic parsing using a large language model, outputting structured selection information including product category tags, explicit constraint sets, inferred implicit constraint sets, and the priority weights of each constraint. The constraint mapping module, connected to the requirement semantic parsing module, receives the structured selection information and categorizes the constraints into hard and soft constraints according to a pre-defined classification logic. For hard constraints, it generates filtering conditions for Boolean retrieval within the knowledge graph; for soft constraints, it generates weight parameters required for subsequent scoring calculations. The product knowledge graph retrieval module, connected to the constraint mapping module, takes the hard constraint filtering conditions as input and performs retrieval matching within a pre-built AI product knowledge graph, outputting a set of candidate products that satisfy all hard constraint conditions, with a built-in relaxation strategy to handle cases where the candidate set is empty. The multi-dimensional evaluation and scoring module is connected to the product knowledge graph retrieval module. Taking the candidate product set and soft constraint weight parameters as input, it quantitatively evaluates each candidate product from four dimensions: semantic matching degree, technical performance score, suitability score, and cost-effectiveness score. A comprehensive weighted score is calculated using the weighted TOPSIS multi-objective optimization algorithm, and the system outputs a ranked result of the candidate products with scores. The interpretable recommendation output module is connected to the multi-dimensional evaluation and scoring module. Taking the comprehensive score ranking result as input, it generates a complete recommendation report containing recommendation ranking, detailed scores for each dimension, and natural language explanations, serving as the system's final output.

[0026] The five modules are connected in series, with the output data format of the previous module strictly matching the input data format of the next module. This allows data to flow seamlessly between modules without human intervention, and together they constitute the end-to-end automated selection architecture of this invention, from unstructured demand input to interpretable recommendation output.

[0027] The requirement semantic parsing module is responsible for transforming the raw natural language requirement description text input by enterprise users into structured selection information. It serves as the cognitive entry point and intelligent starting point of the entire system, and its parsing quality directly determines the processing accuracy of all subsequent modules. The core technology of the requirement semantic parsing module is based on semantic parsing using a large language model. During the parsing process, a chain-based prompting strategy is used to decompose the complex overall parsing task into three sub-tasks: intent recognition, entity extraction, and implicit constraint reasoning, which are executed separately.

[0028] The specific implementation of the chain-like prompt strategy is as follows: The module pre-constructs system prompt words for the selection scenario. These prompt words include a summary description of the knowledge graph of the selection domain, JSON schema constraints for the output format of each sub-task, and thought chain instructions guiding the large model to delve deeper layer by layer according to specific reasoning steps. When the user inputs the requirement text, the requirement semantic parsing module first uses the requirement text and the intent recognition prompt word template as input to the large language model, and calls the large model to execute the intent recognition sub-task to determine the AI ​​product category to which the user's requirement belongs. The output result of the intent recognition sub-task is not a simple category label, but a category probability distribution with confidence scores, such as outputting "Machine Learning Platform: 0.85, Intelligent Customer Service System: 0.10, OCR Recognition System: 0.05", to indicate the probability distribution that the requirement may correspond to multiple product categories. Subsequently, the requirement semantic parsing module uses the requirement text and the identified product category labels together as context, and, together with the entity extraction prompt word template, calls the large language model again to execute the entity extraction sub-task. The goal of the entity extraction subtask is to extract all explicitly stated quantitative parameters and condition descriptions from the requirements text, including but not limited to budget amount, data scale, number of users, deployment environment requirements, compliance requirement level, functional requirement list, and existing technology stack information. Entity extraction employs a combined approach of named entity recognition and relation extraction. For multiple constraints appearing simultaneously in the requirements text, the logical relationships between them also need to be identified. Subsequently, the requirements semantic parsing module uses the requirements text, the identified product category tags, and the extracted set of explicit constraints as context, and, in conjunction with the implicit reasoning prompt template, calls the large language model for the third time to execute the implicit constraint reasoning subtask. The implicit constraint reasoning subtask fully utilizes the common sense reasoning ability and domain knowledge association ability learned by the large language model through pre-training on a large amount of text corpus to automatically complete implicit constraints that are not explicitly mentioned by the user in the requirements text but are highly relevant to their selection decision. For example, when a user describes "We need to process a large amount of customer voice data," the requirement semantic parsing module automatically infers implicit constraints such as "requiring ASR speech recognition capabilities," "requiring a large-scale audio data storage solution," "potentially requiring GPU computing power support," and "potentially requiring audio data anonymization processing to meet privacy compliance requirements." When a user describes "We have three departments with a total of 200 people who need to use this platform," the requirement semantic parsing module automatically infers implicit constraints such as "requiring support for multi-tenant isolation," "requiring a fine-grained permission management system," and "requiring at least 200 concurrent user licenses." When a user describes "The company has high data security requirements," the requirement semantic parsing module automatically infers implicit constraints such as "requiring private deployment or dedicated cloud deployment," "requiring Level 3 or higher information security compliance certification," "requiring encrypted data transmission and storage," and "potentially requiring audit log functionality."

[0029] After completing the three sub-tasks, the requirement semantic parsing module merges and deduplicates the intent recognition results, entity extraction results, and implicit constraint inference results. It also performs conflict detection on constraints from different sub-tasks: for example, if an explicit constraint requires private deployment while the implicit inference result indicates a need for public cloud elastic scaling capabilities, the requirement semantic parsing module will mark this conflict as "awaiting user confirmation." Finally, the requirement semantic parsing module outputs complete structured selection information, which includes at least the following components: one or more product category tags and their confidence scores, a set of explicit constraints containing all explicitly appearing constraints, a set of implicit constraints containing all inference completion constraints, and a priority weight identifier for each constraint. This structured selection information is passed to the constraint mapping module in a standardized JSON format as the data foundation for subsequent processing.

[0030] The AI ​​product knowledge graph is the core knowledge foundation of this invention, providing structured data support for product retrieval and semantic matching. This knowledge graph is stored in a graph database, such as Neo4j, and represents various entities and their complex relationships in the AI ​​product ecosystem in a fine-grained manner in the form of nodes and relational edges.

[0031] The knowledge graph's node types include at least the following six categories: Product nodes, used to store basic information such as the AI ​​product's unique identifier, product name, manufacturer, version number, and release date; Manufacturer nodes, used to store information such as the product supplier's company size, establishment date, industry reputation rating, technical support capability level, and service response time commitment; Functional nodes, used to store the product's functional capabilities, such as data preprocessing, automatic feature engineering, model training, hyperparameter tuning, model deployment, online inference, API calls, and visualization dashboards; Technical indicator nodes, used to store the product's quantitative performance parameters, including but not limited to inference latency, maximum query rate per second, maximum number of supported model parameters, a list of supported algorithm types, and training throughput; Deployment method nodes, used to store the deployment forms supported by the product, including SaaS public cloud services, private container deployment, hybrid cloud deployment, and edge deployment; Industry case nodes, used to store information on successful cases of the product in specific industries, including industry category, customer size, application scenario description, and implementation effect data. The relationship types include at least the following five categories: Belonging relationship, connecting a product node to its parent vendor node; Possessing relationship, connecting a product node to its functional nodes; Indicator value relationship, connecting a product node to its corresponding technical indicator nodes, storing specific quantitative values ​​on the relationship edge; Support relationship, connecting a product node to its supported deployment method nodes; Applicable relationship, connecting a product node to its applicable industry case node, storing case details on the relationship edge. Furthermore, the knowledge graph can be expanded to include competitive relationship nodes as needed, connecting product nodes that constitute a competitive relationship to support subsequent comparative analysis functions.

[0032] The knowledge graph employs a combination of automated data collection and manual review in its data sourcing and update mechanism. For automated data collection, the system uses a scheduled crawler to extract publicly available information such as product feature lists, technical parameter tables, and pricing schemes from the official websites of various AI products. Natural language processing is then used to convert unstructured web page text into structured node and relationship data. Simultaneously, the system connects to structured data from third-party product evaluation platforms via API interfaces, importing industry evaluation results and market analysis data. For manual review, the system includes a management backend for operations personnel to verify, supplement, and correct the automatically collected data. Ambiguous or missing data fields are manually annotated, and the creation of nodes and the establishment of relationships for newly launched products are reviewed and confirmed. The data update frequency is configured differently based on the characteristics of different data sources: official website product information is incrementally synchronized daily, price information is fully compared and updated weekly, and evaluation report data is imported in batches quarterly. When a new product node is added to the knowledge graph or existing node data changes, the system automatically triggers the recalculation of the semantic embedding vectors of the relevant product nodes to ensure that the vector space upon which semantic matching depends remains consistent with the latest knowledge state.

[0033] The constraint mapping module is responsible for transforming the structured selection information output by the demand semantic parsing module into product screening conditions and scoring parameters that can be executed by subsequent modules. The core technical challenge of the constraint mapping module lies in the fact that the constraints in the structured selection information have fundamentally different semantic features. Some constraints describe uncompromising bottom-line conditions, some express only preferred expectations, and some contain vague quantifiers or conditional logic. If these constraints are treated uniformly without distinction, either suitable candidate products will be missed due to overly strict uniform filtering, or the product quality will be ineffectively differentiated due to overly lenient uniform weighting. Therefore, the constraint mapping module adopts a dual-track processing mechanism of hard constraint and soft constraint classification mapping.

[0034] The hard constraint mapping submodule is responsible for identifying and classifying constraints containing precise numerical thresholds or uncompromising conditions in the structured selection information as hard constraints. The specific judgment logic is as follows: constraints involving bottom-line conditions, such as budget constraints, deployment method constraints, compliance certification constraints, and data residency constraints, are marked as hard constraints. Hard constraints are processed by generating Boolean logic filtering expressions that can be directly executed within the knowledge graph. For example, "budget not exceeding 500,000 yuan" is transformed into "annual cost ≤ 500,000 yuan," "must be private deployment" is transformed into "deployment method = private," and "requires Level 3 Information Security Protection Certification" is transformed into "compliance certification includes Level 3 Information Security Protection Certification." All hard constraint filtering expressions are combined using conjunctive normal form into a single screening condition; candidate products must simultaneously satisfy all hard constraint conditions. When a hard constraint cannot be mapped to a specific attribute value within the current data range of the knowledge graph, the system marks the hard constraint as "pending confirmation" and requests further information or adjustment of the constraint expression from the user through the user interface.

[0035] The soft constraint quantification submodule is responsible for identifying and classifying constraints containing fuzzy quantifiers, preference descriptions, or conditional modifications in the structured selection information as soft constraints. For example, "fast response speed," "preferably with experience in the financial industry," "budget around 500,000," and "performance first, price second" are all soft constraints. The processing method for soft constraints is to convert them into weight parameters for each scoring dimension. This invention uses a semantic similarity-guided weight allocation mechanism to complete this conversion. Specifically, the system first prepares a semantic description text for each scoring dimension of the multi-dimensional evaluation. For example, the semantic matching dimension is described as "the degree of fit between product functions and user business needs," the technical performance dimension is described as "the performance level of the product's quantitative performance indicators," the adaptability scoring dimension is described as "the compatibility of the product with the company's existing technology stack," and the cost-effectiveness scoring dimension is described as "the ratio between product function coverage and price." Then, the overall semantics of the structured constraint set and the semantic descriptions of each dimension are mapped to the same vector space through an embedding model, and the cosine similarity is calculated. The formula for weight allocation is: w i = Where C is the overall semantic embedding vector of the structured constraint set, desc i Let sim( be the semantic description embedding vector for the i-th rating dimension) , Let be the cosine similarity function, and λ be the temperature coefficient. When user needs show a clear emphasis on a certain dimension, λ takes a larger value, making the weight distribution more concentrated; when the emphasis on each dimension in user needs is unclear, λ takes a smaller value, making the weight distribution more even. The physical meaning of the above formula is: the more similar the semantic description of a rating dimension is to the user's overall needs constraints, the more important that dimension is in the user's selection considerations, and therefore it is assigned a higher weight. This semantically driven approach allows vague expressions such as "fast response speed" to be automatically quantified as high-weight technical performance dimensions without the need for manually preset rule mapping tables.

[0036] For composite constraints containing conditional logic, the constraint mapping module decomposes the composite constraint into multiple sub-constraints and establishes logical relationships using a large language model. In the example above, "budget around 500,000" is identified as a soft constraint and quantified as a weight parameter for the budget dimension, while "if the performance is exceptionally good, it can be relaxed to 800,000" is identified as a conditional soft constraint. When the technical performance score of a candidate product exceeds a certain threshold, the weight of the budget constraint automatically decreases, forming conditional scoring logic. The module ultimately outputs two data objects: one is the filtering condition expression generated by merging all hard constraints, and the other is the weight parameter vector w=[w1,w2,...,w...] for each scoring dimension. d ] TBoth are passed to the product knowledge graph retrieval module and the multi-dimensional evaluation and scoring module, respectively.

[0037] The product knowledge graph retrieval module receives the hard constraint filtering condition expression output by the constraint mapping module and performs a retrieval and matching operation in the AI ​​product knowledge graph. The retrieval process essentially involves performing a conjunction logic judgment on the hard constraints for each product node in the knowledge graph. Only when a product node and its associated attribute nodes satisfy all conjunction terms in the filtering condition expression is the product included in the candidate product set.

[0038] The specific method for executing the above retrieval is as follows: The system transforms the hard constraint filtering condition expression into a graph database query statement, such as the Cypher query language, and performs filtering queries on the product nodes. For example, when the filtering condition is "Annual cost ≤ 500,000 AND Deployment method = Private AND Compliance certification includes Level 3 Information Security Protection", the system performs a graph traversal operation. Starting from all product nodes, it searches for the subset of products whose annual cost attribute value satisfies ≤ 500,000 along the indicator value relationship edge, searches for the subset of products whose deployment method is private along the support relationship edge, and searches for the attribute node whose compliance certification includes Level 3 Information Security Protection along the possession relationship edge. Then, it performs an intersection operation on the three subsets to finally obtain the candidate product set.

[0039] When the candidate product set is empty, it means that no product within the current product knowledge graph coverage can simultaneously meet all the user's hard constraints. At this point, the product knowledge graph retrieval module triggers a relaxation strategy. The execution flow of the relaxation strategy is as follows: First, the system prioritizes all current hard constraints. The priority level is determined by the constraint priority weight assigned during the requirement semantic parsing phase, and within the same level, they are sorted according to the semantic specificity of the constraints. Then, the system attempts to relax each hard constraint one by one, starting with the lowest priority. The relaxation method varies depending on the constraint type: for budget-related numerical constraints, the numerical range is expanded upwards by a preset step size; for deployment method constraints, "private deployment" is relaxed to "private deployment or hybrid cloud deployment"; for compliance certification constraints, "Level 3 Information Security Protection" is relaxed to "Level 2 or above". After relaxing each constraint, the knowledge graph retrieval is re-executed. If the candidate product set is not empty, the relaxation process stops; if it is still empty, the next constraint is relaxed. If the candidate product set is still empty after relaxing all hard constraints, the system selects several products with the highest semantic similarity to the user's needs from the knowledge graph as alternative recommendations, and explicitly informs the user that there are no perfectly matching products in the current knowledge graph. Regardless of how the candidate product set is ultimately obtained, the system outputs the candidate product list along with a list of relaxed or removed constraint tags, and passes it to the multi-dimensional evaluation and scoring module.

[0040] The multi-dimensional evaluation and scoring module performs a comprehensive quantitative evaluation of the candidate product set output by the product knowledge graph retrieval module. The evaluation is carried out from four independent dimensions, and the scoring results of each dimension are weighted and summed to form the final comprehensive score.

[0041] The semantic matching degree is used to measure the semantic fit between the overall product functional description and the user's requirement text. The semantic matching degree is calculated using a dense vector embedding method generated by a large language model. For each product pp in the candidate product set, the system extracts the complete description text of the product from the knowledge graph. This description text is composed of basic product information, a list of functions, a summary of technical indicators, a description of applicable scenarios, and a summary of industry cases, forming a structured comprehensive product description. Simultaneously, the user's original requirement text serves as the query text. Both are input into the same embedding model, yielding dense vectors E. product (p) and E query (T). The formula for calculating semantic matching degree is the cosine similarity between the two: Match sem (p,T)= The cosine similarity value ranges from [-1, 1]. In this invention, since all dimensions of the embedding vector are non-negative, the actual range is mapped to [0, 1]. A larger value indicates a higher degree of semantic matching between the product function and user needs. The key advantage of this calculation method is that the semantic representation capability obtained by the embedding model through pre-training on a large-scale corpus makes expressions such as "call recording analysis" and "speech recognition + sentiment computing" have a high similarity in the vector space, thereby achieving deep semantic matching that goes beyond literal keyword matching.

[0042] The technical performance score reflects the product's performance level on performance indicators that users care about. This dimension's calculation relies on product technical indicator node data stored in a knowledge graph. First, a list of performance-related keywords is extracted from the structured information of user needs. By semantically matching these keywords with technical indicator nodes in the knowledge graph, the set of performance indicators to be examined in this evaluation is determined. Then, for each selected performance indicator, the quantitative values ​​of each candidate product on that indicator are normalized. Finally, the normalized scores of each indicator are multiplied by their corresponding importance weights and summed to obtain the technical performance score.

[0043] The compatibility score assesses the compatibility level between the product and the enterprise's existing technology stack. This dimension is calculated based on the relationships between product nodes and the enterprise's technology stack in the knowledge graph. When user requirements include information about the existing technology stack, the system searches the knowledge graph for compatibility relationships between each candidate product and these technology stack components. The compatibility score specifically includes: programming language compatibility sub-scores, data format compatibility sub-scores, API specification compatibility sub-scores, cloud platform compatibility sub-scores, etc. Each sub-score is weighted and summed according to the user's emphasis on technology stack compatibility to obtain the compatibility score.

[0044] The cost-effectiveness score comprehensively considers the product's functional coverage, performance, and price ratio. This dimension is calculated by obtaining pricing information for each candidate product from a knowledge graph, extracting the product's comprehensive capability score from the dimensions of functional matching and technical performance, and dividing the comprehensive capability score by the normalized annual total cost as the cost-effectiveness score.

[0045] After obtaining the original scores of candidate products in the above four dimensions, the system uses the weighted TOPSIS method to calculate the comprehensive weighted score. TOPSIS is a classic multi-objective decision-making algorithm. Its core idea is that the optimal solution should be closest to the positive ideal solution and furthest from the negative ideal solution.

[0046] The specific calculation steps for weighted TOPSIS are as follows.

[0047] The first step is to construct the original decision matrix X=[x ij ] N×4 Where N is the number of candidate products, x ij The original score of the i-th candidate product on the j-th scoring dimension.

[0048] The second step is to perform vector normalization on the original decision matrix. The normalization formula is as follows: r ij = , i=1..N, j=1..4 The third step involves using the soft constraint weight vector w=[w1,w2,w3,w4] output by the constraint mapping module. T By weighting the normalized matrix, we obtain the weighted normalized matrix v. ij =w j r ij .

[0049] The fourth step is to determine the positive and negative ideal solutions for each dimension. For the three benefit-oriented indicators—semantic matching degree, technical performance score, and fit score—the positive ideal solution takes the maximum value in each dimension, and the negative ideal solution takes the minimum value in each dimension; the cost-effectiveness score is also treated as a benefit-oriented indicator.

[0050] Step 5: Calculate the Euclidean distance from each candidate product to the positive and negative ideal solutions: D i + = D i - = Step 6: Calculate the relative similarity of each candidate product: Score total (pi)= The relative proximity score ranges from [0,1]. A larger value indicates that the candidate product is closer to the positive ideal solution and further away from the negative ideal solution, meaning it performs better overall. The system sorts the candidate products from largest to smallest relative proximity score to form the final comprehensive score ranking result.

[0051] The interpretable recommendation output module transforms the comprehensive scoring and ranking results generated by the multi-dimensional evaluation and scoring module into a recommendation report that can be directly used and reviewed by the enterprise decision-making team. The interpretable recommendation output module solves the problem of black-box output in existing selection recommendation systems, making the recommendation results operable for formal adoption in the enterprise procurement process.

[0052] The core content of the aforementioned recommendation output comprises four layers. The first layer is a recommendation ranking list, which clearly displays the overall ranking, product name, manufacturer, and overall score of each candidate product, with top-ranked products highlighted visually. The second layer is a detailed score breakdown by dimension, visually displaying the scores of each candidate product in four dimensions—semantic matching, technical performance, suitability, and cost-effectiveness—using radar charts or bar graphs. This allows the decision-making team to clearly understand the strengths and weaknesses of each product across different dimensions. The third layer is a natural language explanation, automatically generated by the large language model based on the matching relationship between the recommendation results and user needs. This explanation is not a mechanical filling of a preset template, but rather uses the specific characteristics of the candidate products, the score results of each dimension, and the user's original demand text as context input to the large language model, which then generates a customized natural language description for that specific recommendation scenario. For example, for the top-ranked product, the generated explanation includes: in which functionalities does the product highly match user needs, in which technical indicators does it excel, in which dimensions does it have shortcomings, and why it is suitable for the user's business scenario overall; the fourth layer is a horizontal comparative analysis report, which allows users to select two to three candidate products for side-by-side comparison, presenting the specific numerical comparison of each product in various detailed indicators in tabular form, to help the decision-making team make the final weighing and judgment.

[0053] At the system operation level, this invention also incorporates a built-in active learning mechanism based on user feedback. After a user receives recommendations and makes a final selection decision, the system records the decision result—that is, which product the user actually chose to purchase and its ranking in the recommendation list. The system periodically performs statistical analysis on the accumulated user feedback data: for frequently selected products, the scoring parameters of their relevant attribute nodes are increased; for frequently abandoned top-ranked products, their common disadvantages are analyzed, and the penalty weights for these features in the scoring model are adjusted accordingly; simultaneously, the system automatically adjusts the temperature coefficient λ in the semantic similarity-guided weight allocation mechanism based on feedback data. When the consistency rate between the user's actual selected product and the system's top-ranked recommended product is high, the current λ value remains unchanged; when the consistency rate is low, the λ value is appropriately reduced to make the weight distribution more even, avoiding overfitting of specific dimension preferences. This closed-loop learning mechanism allows the system's scoring model to continuously evolve with the accumulation of usage data, constantly improving the accuracy of subsequent recommendations and user satisfaction.

[0054] The large language model in this invention adopts a general large model based on the Transformer architecture. Specifically, mature commercial or open source models such as the GPT series, DeepSeek series, and Qwen series can be selected and integrated into the requirement semantic parsing module and the interpretable recommendation output module through API.

[0055] In the demand semantic parsing module, the system configures specific system prompts for each large language model call. These prompts consist of three parts: First, domain knowledge, including common AI product classification systems, key evaluation indicators for each product category, and industry-standard technical terms and conceptual definitions, to help the large model understand user needs within the domain context. Second, output format constraints, explicitly specifying the data structure and field types of each subtask's output in JSON Schema format, ensuring the model output can be directly parsed programmatically by subsequent modules without additional manual post-processing. Third, inference step guidance instructions, employing a thought chain prompting method to guide the large model to generate intermediate steps in the inference process before outputting the final result, improving the accuracy and logical consistency of the parsing results. In the interpretable recommendation output module, the system prompts focus on guiding the large model to generate targeted, complete, and fluent natural language explanations. The prompts explicitly require that the explanations include four essential elements: matching points, advantages, shortcomings, and comprehensive recommendation reasons, ensuring the output content has substantial decision-making reference value for the decision-making team.

[0056] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit the technical solutions. Although the applicant has described the present invention in detail with reference to preferred embodiments, those skilled in the art should understand that any modifications or equivalent substitutions made to the technical solutions of the present invention cannot depart from the spirit and scope of the present invention and should be covered within the scope of the claims of the present invention.

Claims

1. A large-scale model-based intelligent selection system for enterprise AI products, characterized in that, include: The requirement semantic parsing module is used to receive natural language requirement description text input by enterprise users, perform semantic parsing on the requirement text through a large language model, and extract structured selection information. The structured selection information includes at least product category tags, explicit constraint set, implicit constraint set obtained by reasoning, and priority weight of each constraint. The constraint mapping module, connected to the requirement semantic parsing module, is used to classify each constraint item in the structured selection information into hard constraints and soft constraints, and generate executable product screening conditions and scoring function parameters. The product knowledge graph retrieval module is connected to the constraint mapping module. It is used to perform retrieval and matching in the pre-built AI product knowledge graph based on the hard constraints, filter out the candidate product set that meets all hard constraints, and trigger a relaxation strategy when the candidate product set is empty. The multi-dimensional evaluation and scoring module is connected to the product knowledge graph retrieval module. It is used to calculate the scores of four dimensions for each product in the candidate product set: semantic matching degree, technical performance score, adaptability score and cost-effectiveness score, and calculate the comprehensive weighted score according to the weight parameters of the soft constraints. The interpretable recommendation output module is connected to the multi-dimensional evaluation and scoring module. It is used to sort the candidate products in descending order according to the comprehensive weighted score to generate a recommendation list and output the recommendation results including recommendation ranking, sub-dimensional score details and natural language explanation. Specifically, the semantic parsing module, when performing semantic parsing using a large language model, employs a chain-like prompting strategy to decompose the semantic parsing task into three sub-tasks: intent recognition, entity extraction, and implicit constraint reasoning. The implicit constraint reasoning sub-task, based on the common-sense reasoning capabilities of the large language model, completes the selection constraints not explicitly mentioned by the user. The constraint mapping module calculates the weight parameters for each scoring dimension for soft constraints containing fuzzy quantifiers or preference descriptions through a semantic similarity-guided weight allocation mechanism. The multi-dimensional evaluation and scoring module uses a weighted TOPSIS multi-objective optimization algorithm to calculate a comprehensive weighted score, thus forming a fully automated selection architecture from unstructured requirement input to interpretable recommendation output.

2. The enterprise AI product intelligent selection system based on a large model as described in claim 1, characterized in that, In the demand semantic parsing module, the intent recognition subtask is used to determine the AI ​​product category to which the user demand belongs, the entity extraction subtask is used to extract quantitative parameters and condition descriptions from the demand text, and the implicit constraint reasoning subtask is used to automatically reason about implicit selection conditions such as data scale, concurrent demand, and security compliance level based on the user demand context.

3. The enterprise AI product intelligent selection system based on a large model as described in claim 1, characterized in that, The node types of the AI ​​product knowledge graph include at least product nodes, manufacturer nodes, function nodes, technical indicator nodes, deployment method nodes, and industry case nodes; the relationship types include at least belonging relationships, possessing relationships, indicator value relationships, supporting relationships, and applicable relationships; the knowledge graph is stored in a graph database and is dynamically updated and maintained through a combination of automated data collection and manual review.

4. The enterprise AI product intelligent selection system based on a large model according to claim 1, characterized in that, In the constraint mapping module, constraints containing precise numerical thresholds or uncompromising condition descriptions are classified as hard constraints, which are used to perform Boolean logic filtering in the knowledge graph; constraints containing fuzzy quantifiers, preference descriptions, or conditional modifications are classified as soft constraints. The semantic similarity-guided weight allocation mechanism is expressed as follows: w i = Where C is the set of structured constraints, desc i Let sim( be the semantic description embedding vector for the i-th rating dimension) , ) is the cosine similarity function, and λ is the temperature coefficient.

5. The enterprise AI product intelligent selection system based on a large model according to claim 1, characterized in that, In the multi-dimensional evaluation and scoring module, the semantic matching degree is calculated using dense vector embedding generated by a large language model, and the calculation formula is as follows: Match sem (p,T)= Among them, E product (p) is the vector obtained by embedding the complete description text of product p through the embedding model mapping, E query (T) is the query vector obtained by mapping the user's request text T through the same embedding model.

6. The enterprise AI product intelligent selection system based on a large model according to claim 1, characterized in that, The multi-dimensional evaluation and scoring module uses the weighted TOPSIS multi-objective optimization algorithm to calculate the comprehensive weighted score. Specifically, it includes: performing vector normalization on the original scores of each product in the candidate product set in four dimensions: semantic matching degree, technical performance score, suitability score, and cost-effectiveness score; weighting the scores of each dimension with the weight parameters of the soft constraints; determining the positive ideal solution and negative ideal solution for each dimension; calculating the distance of each candidate product to the positive ideal solution and negative ideal solution; and using the relative proximity as the final comprehensive score.

7. The enterprise AI product intelligent selection system based on a large model according to claim 1, characterized in that, In the product knowledge graph retrieval module, the relaxation strategy includes: when the candidate product set that satisfies all hard constraints is empty, the system identifies the constraint items that lead to zero results, relaxes or removes the constraints one by one from low to high priority, until the candidate product set is not empty, and marks the relaxed or removed constraint items and returns them to the user along with the prompt information.

8. The enterprise AI product intelligent selection system based on a large model according to claim 1, characterized in that, In the interpretable recommendation output module, the natural language explanation is automatically generated by the large language model based on the matching relationship between the recommendation results and user needs. The detailed subdivision scoring is visualized in the form of radar charts or bar charts. The recommendation results support the generation of a horizontal comparative analysis report of two to three candidate products.

9. The enterprise AI product intelligent selection system based on a large model according to claim 1, characterized in that, The AI ​​product knowledge graph supports an active learning mechanism based on user selection decision feedback. After the system records the user's final selection decision, it automatically adjusts the temperature coefficient and product rating parameters in the semantic similarity-guided weight allocation mechanism according to the decision feedback, so as to achieve continuous optimization of the rating model.

10. The enterprise AI product intelligent selection system based on a large model according to any one of claims 1 to 9, characterized in that, The large language model adopts a general large model based on the Transformer architecture and is integrated into the requirement semantic parsing module and the interpretable recommendation output module via API. The requirement semantic parsing module is configured with system prompt words that include selection domain knowledge, output format constraints, and reasoning step guidance instructions.

Images