AI product intelligent selection assistant system integrating large language models

Abstract

The application discloses an AI product intelligent selection assistant system integrated with a large language model, and relates to the technical fields of artificial intelligence and intelligent recommendation. The system comprises a user intention understanding and demand analysis module, a multi-source product knowledge base construction and management module, a retrieval enhancement generation engine module, a multi-dimensional constraint reasoning and scoring module, a multi-agent collaborative selection module, a synergistic effect analysis module and an interpretable recommendation output module. The user inputs the selection demand in natural language, and a structured demand representation is generated through intention recognition and entity extraction. Candidate product information is retrieved from the knowledge base through a three-way hybrid retrieval strategy. After hard constraint filtering and seven-dimensional comprehensive scoring, demand analysis, candidate screening, deep evaluation, comparative analysis and scheme recommendation are completed by six special agents in collaboration. The synergistic effect analysis module calculates the synergistic effect score of the multi-product combination to generate a combination recommendation scheme. Finally, the user is presented with a visual chart and natural language recommendation reasons.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and intelligent recommendation technology, specifically to an AI product intelligent selection assistant system integrating a large language model, which is particularly suitable for intelligent selection scenarios of complex products such as AI software products and enterprise-level solutions. Background Technology

[0002] Currently, enterprises generally lack efficient and accurate technical support when making selection decisions for AI large-scale models, intelligent development platforms, and various enterprise-level software products. Traditional selection methods often rely on keyword matching, collaborative filtering, or recommendation algorithms based on simple deep learning. While these technologies are adequate for shallow information processing, they struggle with deep intent reasoning and semantic disambiguation when faced with complex business needs expressed by users in natural language. Furthermore, selection decisions are essentially a constraint-satisfying problem influenced by multiple factors, including commercial license compatibility, hardware resource configuration, inference performance metrics, technology stack matching, and cost budget. Existing systems typically simplify these constraints to a single ranking factor or linear weighting function, lacking the ability to hierarchically model and logically reason about hard constraints and soft preferences. This often results in recommendation results failing to meet key performance indicators, severely impacting the usability and reliability of the selection solution.

[0003] At the knowledge update and system architecture levels, existing technologies have significant shortcomings. On the one hand, the AI ​​product field experiences extremely rapid knowledge iteration, with weekly model version updates, monthly performance benchmark breakthroughs, and quarterly price strategy adjustments becoming the norm in the industry. However, existing selection systems generally rely on static knowledge bases or rule engines maintained manually on a regular basis, lacking an automated synchronization mechanism with official product documentation, open-source communities, and authoritative evaluation lists. This results in severely outdated selection criteria, making it difficult to reflect the latest status and actual performance of products. On the other hand, most current selection recommendation schemes treat each product as an independent recommendation, focusing only on the single-item score of a single product, completely ignoring the strong correlation between different products in terms of interface protocol compatibility, data flow continuity, technology stack consistency, and operational management collaboration when building a complete solution. This can lead to data silos or architectural conflicts after users select multiple optimal product combinations, resulting in the actual performance of the overall solution being far below expectations, or even making deployment impossible.

[0004] A review and analysis of existing patent literature reveals that while some technical solutions attempt to address the aforementioned issues from different perspectives, significant gaps in coverage and limitations remain. For instance, patent CN119336590A proposes a step-by-step screening method based on commercial licensing, hardware resources, and inference speed, but it employs a linear filtering strategy and lacks the ability to understand users' natural language intent and provide a flexible, multi-dimensional comprehensive evaluation. Another patent, CN119624592A, attempts to combine Transformer, graph neural networks, and large language models for feature extraction and product recommendation, but its application scenarios are primarily geared towards general e-commerce or content recommendation fields. It neither models the specific technical constraints of AI products nor provides an interpretable selection reasoning process. Furthermore, while patent CN120179897A introduces the concept of a multimodal interactive intelligent assistant, its application is limited to specific vertical fields and lacks the general selection decision-making capability for enterprise-level diversified AI product catalogs. In summary, existing technologies lack a complete technical solution capable of simultaneously achieving deep semantic parsing for natural language needs, unified logical reasoning with multi-dimensional hard and soft constraints, real-time enhanced product knowledge retrieval, evaluation of combined solutions based on synergistic effects, and interpretable recommendations through multi-agent collaboration. To address this technological gap, this invention proposes an AI-powered intelligent product selection assistant system integrating a large language model. This system aims to overcome the aforementioned technical bottlenecks and provide enterprises with efficient, accurate, interpretable, and globally optimized intelligent selection decision support. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies by providing an AI-powered intelligent product selection assistant system that integrates a large language model. It aims to achieve a deep understanding of users' natural language needs, enhanced intelligent retrieval of product knowledge bases, and comprehensive reasoning and optimization based on multi-dimensional constraints through a multi-agent collaborative architecture. This provides users with accurate and interpretable AI-powered product selection suggestions, improving selection efficiency and decision-making quality.

[0006] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: An AI-powered product selection assistant system integrating a large language model includes: The user intent understanding and requirement parsing module is configured to receive product selection requirements input by users in natural language, call a large language model to perform intent recognition and entity extraction, and output a structured requirement representation. The multi-source product knowledge base construction and management module is configured to collect multi-dimensional information of AI products, construct a knowledge graph containing product nodes and relationship edges, establish a vectorized index, and incrementally update the knowledge base at regular intervals. The retrieval enhancement generation engine module is configured to retrieve candidate product information from the knowledge base based on the structured requirement representation through a multi-path retrieval strategy, and to reorder and fuse the retrieval results to generate selection context information. The multi-dimensional constraint reasoning and scoring module is configured to perform hard constraint satisfaction filtering on candidate products based on the constraints in the structured requirement representation, and to calculate scores on seven dimensions: functional matching degree, performance adaptability, resource adaptability, cost-effectiveness, availability, security compliance and scalability, and generate a comprehensive selection score. The multi-agent collaborative selection module includes a demand analysis agent, a candidate screening agent, a deep evaluation agent, a comparative analysis agent, a solution recommendation agent, and a coordination and supervision agent. The coordination and supervision agent is responsible for task decomposition and result summarization. Each agent performs task division and collaboration based on a large language model, and outputs selection results including selection recommendation schemes and recommendation reasons. The synergy analysis module is configured to calculate the synergy score for combination schemes containing multiple candidate products based on the compatibility relationship and functional complementarity between products, and generate a recommended combination product scheme. An interpretable recommendation output module is configured to present the selection results to the user in the form of visual charts, natural language recommendation reasons, and confidence level labels.

[0007] This invention constructs a complete intelligent selection closed loop from demand input to selection result output through the systematic integration of seven core modules. The user intent understanding module transforms natural language requirements into structured representations, solving the bottleneck of traditional systems' inability to deeply understand user semantics. The RAG engine accurately recalls candidate product information through multi-path retrieval, providing high-quality context for reasoning. The multi-dimensional constraint reasoning and scoring module comprehensively scores across seven dimensions, avoiding the one-sidedness of single-dimensional evaluation. The multi-agent collaboration module professionally decomposes complex selection tasks through the division of labor and cooperation of six specialized agents, improving the accuracy and reliability of reasoning. The synergy effect analysis module achieves global optimization of multi-product combination schemes for the first time, and the interpretable output module makes the selection results transparent and traceable. The collaborative work of each module improves selection efficiency.

[0008] The optimized user intent understanding and requirement parsing module includes: The intent recognition unit classifies user input intent using a finely tuned large language model. The entity extraction unit uses a Transformer-based named entity recognition model to extract entity information such as product type, functional requirements, performance indicators, resource constraints, deployment methods, and compliance requirements. Context memory units maintain dialogue coherence across multiple interactions, recording users' historical selection preferences and historical dialogue information; The structured requirements generation unit transforms the parsed results into standardized requirements description documents.

[0009] In the above scheme, the intent recognition unit accurately classifies selection scenarios using a fine-tuned large language model, providing scenario adaptation strategies for subsequent reasoning; the entity extraction unit uses a Transformer model to accurately extract seven core entities from natural language, covering all key information dimensions of selection decisions; the context memory unit maintains dialogue coherence and records historical preferences during multiple interactions, enabling personalized reasoning; and the structured requirement generation unit transforms unstructured input into standardized requirement documents, providing a unified and computable input format for downstream modules. The collaboration of these four units accurately transforms vague, colloquial requirements into structured constraints, ensuring the accuracy and consistency of the selection process from the outset.

[0010] The optimized retrieval enhancement generation engine module employs a three-way hybrid retrieval strategy: The first search method is a structured search based on structured query language, used to accurately match product specifications and parameters; The second retrieval method is semantic retrieval based on vector similarity, used to recall semantically relevant product description text; The third retrieval method is graph retrieval based on knowledge graphs, used to obtain information on relationships between products; The three search results are merged and sorted by the re-sorting unit to output the Top-K results.

[0011] In the above scheme, structured retrieval precisely matches product specifications, meeting the precise retrieval requirements under numerical constraints; semantic retrieval, based on vector similarity, recalls semantically relevant product descriptions without keyword matching, compensating for the semantic blind spots of structured retrieval; graph retrieval, based on knowledge graph traversal, obtains the relationships between products, providing relational data support for synergistic effect analysis. The three results are re-ranked and fused to output Top-K results, balancing recall and precision, and significantly improving retrieval accuracy compared to a single retrieval strategy.

[0012] In the optimized multi-dimensional constraint reasoning and scoring module: the constraints are divided into hard constraints and soft constraints according to their mandatory nature. Hard constraints are mandatory conditions that must be met, while soft constraints are preference conditions that can be optimized through trade-offs. The hard and soft constraints are obtained by parsing the structured requirement representation output by the user intent understanding and requirement parsing module. The multiple dimensions include seven dimensions: functional matching degree, performance adaptability, resource adaptability, cost-effectiveness, usability, security compliance, and scalability. The comprehensive score is the weighted sum of the scores of each dimension, and the weights are dynamically adjusted according to the user's preference for the soft constraints in the structured requirement representation. This solution divides the constraints into hard and soft constraints and provides a clear calculation framework for the comprehensive score. Hard constraints, as mandatory conditions, have priority and filter out unqualified products first, avoiding the structural risk of "shortcomings being masked by high scores" in traditional linear weighting. Soft constraints, as optimizable preferences, have their weights dynamically adjusted according to the user's preference, realizing personalized recommendations. The seven evaluation dimensions cover all key considerations for AI product selection. The comprehensive score is a weighted sum of each dimension, which ensures that the recommendation results meet both mandatory conditions and maximize the satisfaction of user preferences, thereby improving the usability of the solution and user satisfaction.

[0013] In the optimized multi-agent collaborative selection module: the coordinating and supervising agent is responsible for task decomposition, communication coordination between agents, and result summarization; the demand analysis agent is responsible for in-depth analysis of user needs and identification of implicit needs; the candidate screening agent filters the initial candidate product list based on hard constraints; the deep evaluation agent performs fine-grained evaluation from the dimensions of function, performance, cost, and compatibility; the comparative analysis agent is responsible for horizontal comparative analysis between candidate products; and the solution recommendation agent generates the final recommended solution based on the comprehensive evaluation results. In this invention, the coordinating and supervising agent acts as the global scheduling center, ensuring orderly collaboration among agents and avoiding task conflicts and information redundancy; the demand analysis agent focuses on in-depth analysis and identification of implicit needs, discovering key conditions not explicitly expressed by users; the candidate screening agent quickly filters based on hard constraints, narrowing the search space for subsequent evaluations and improving processing efficiency; the deep evaluation agent performs fine-grained evaluation from four dimensions: function, performance, cost, and compatibility, ensuring the professionalism and comprehensiveness of the evaluation; the comparative analysis agent is responsible for horizontal comparison; and the solution recommendation agent integrates the outputs of all parties to generate the final recommendation.

[0014] In the optimized synergy effect analysis module, for candidate product combinations P={p1,p2,...,pk}, the synergy effect score is calculated as follows: Construct a compatibility matrix (Compatibility(i,j)) among candidate products; Calculate the synergistic effect score: Synergy(P) = Complementary(i,j) represents the degree of complementarity of product functions.

[0015] In the above scheme, the compatibility matrix Compatibility(i,j) quantifies the degree of technical compatibility between any two product combinations; the functional complementarity Complementary(i,j) evaluates the functional complementarity between products; and the synergy score Synergy(P) calculates the average of the products of compatibility and complementarity of all product pairs in the combination, thus evaluating the overall integration efficiency of the combination scheme. This enables the system to recommend not only the best single product, but also the best overall product combination scheme, avoiding the selection trap of having the best individual product scores but conflicting architectures after combination.

[0016] The optimized, interpretable recommendation output module includes: The recommendation reason generation unit uses a large language model to generate natural language recommendation reasons; Comparing the visualization units, generate radar charts or bar charts to display the score comparison of candidate products across various dimensions; An interactive exploration interface allows users to adjust the scoring weights or constraints of various dimensions to view changes in the selection results in real time. The confidence score labeling unit labels each recommendation result with a confidence score and its influencing factors.

[0017] The above scheme constructs a complete and explainable recommendation system through four output units: the recommendation reasoning generation unit presents the reasoning process in natural language, making up for the lack of transparency in the decision-making of deep learning models; the comparison visualization unit transforms multi-dimensional abstract data into intuitive visual information, reducing the cognitive burden on users; the interactive exploration interface supports scenario analysis, enhancing users' sense of participation and control; and the confidence labeling unit labels the confidence score and influencing factors, enabling users to objectively evaluate the reliability of the recommendations.

[0018] This invention also introduces an AI product intelligent selection method using the above-mentioned integrated large language model AI product intelligent selection assistant system, wherein each module of the system is executed according to the following steps: S1: The User Intent Understanding and Requirement Parsing module receives the product selection requirements input by the user in natural language, performs intent recognition and entity extraction through a large language model, and generates a structured requirement representation. S2: Multi-source product knowledge base construction and management module: Constructs a multi-source product knowledge base, collects multi-dimensional information of AI products to build knowledge graphs and vector indexes; S3: The retrieval enhancement generation engine module retrieves candidate product information from the knowledge base based on the structured requirement representation through a three-way hybrid retrieval strategy, and generates selection context information after reordering and knowledge fusion; S4: The multi-dimensional constraint reasoning and scoring module performs hard constraint filtering on candidate products based on the constraints in the structured requirement representation and performs comprehensive scoring on seven dimensions: functional matching degree, performance adaptability, resource adaptability, cost-effectiveness, availability, security compliance and scalability. S5: The multi-agent collaborative selection module performs requirement analysis, candidate screening, in-depth evaluation, comparative analysis, and solution recommendation through multi-agent collaboration, generating a selection recommendation solution; S6: The synergy analysis module performs synergy analysis on combination schemes containing multiple candidate products and generates recommended combination product schemes. S7: The explainable recommendation output module presents the selection results to the user in the form of visual charts and natural language recommendation reasons.

[0019] The above solution provides a complete methodological framework for the system. S1 uses a large language model for intent recognition and entity extraction to ensure accurate requirement analysis; S2 constructs a multi-source knowledge base to provide structured, vectorized, and relational data support for the selection process; S3 ensures comprehensive and accurate candidate product information through three-way retrieval and knowledge fusion; S4 achieves scientific evaluation and ranking through hard constraint filtering and seven-dimensional scoring; S5 generates recommendation schemes through multi-agent collaboration to ensure professional and in-depth reasoning; S6 supports combined scheme recommendations through synergy analysis; and S7 presents the results in visualization and natural language to ensure understandability. These seven interconnected steps standardize and streamline complex selection tasks, balancing high precision with efficient automation.

[0020] To further optimize the above-mentioned intelligent selection method for AI products, the multi-agent collaboration in step S5 adopts the following process: 1) The coordinating and supervising agent receives tasks and decomposes and assigns them to the requirements analysis agent; 2) After the requirements analysis agent generates a structured requirements report, it triggers a candidate selection agent; 3) The candidate selection agent performs hard-constraint filtering to generate a candidate list; 4) Deeply evaluate the agent and concurrently perform evaluations across various dimensions; 5) Comparative analysis of the agent's generation of horizontal comparison results; 6) The solution recommendation agent integrates the outputs from all parties to generate a recommended solution; 7) Coordinate and supervise the quality of the intelligent agent's review and recommendation schemes.

[0021] The above scheme uses a coordinating and supervising intelligent agent as the scheduling center, and proceeds in the order of "requirement analysis → hard constraint filtering → parallel deep evaluation → comparative analysis → comprehensive recommendation → quality review". Requirements analysis is executed first to ensure that subsequent steps are based on accurate requirements; candidate screening quickly compresses the product space and reduces the computational overhead of deep evaluation; deep evaluation is performed concurrently for each dimension, shortening the overall processing time; comparative analysis and scheme recommendation are completed in sequence to ensure the progressive reasoning logic; and the coordinating and supervising agent conducts the final review to ensure that the output quality meets the standards.

[0022] To further optimize the aforementioned AI-powered intelligent product selection method, the multi-source product knowledge base undergoes incremental updates via scheduled tasks and event-triggered mechanisms, automatically synchronizing product version updates, performance evaluation changes, and price fluctuations. This solution employs a dual update strategy of scheduled tasks and event triggers, ensuring the system automatically synchronizes product version updates, performance evaluation changes, and price fluctuations, keeping the knowledge base always in sync with the actual product status. Compared to manual periodic maintenance, this mechanism compresses the knowledge timeliness window from days to hours, eliminating the risk of recommendation failure due to outdated knowledge and guaranteeing that selection suggestions are always based on the latest and most accurate information. This provides a guarantee for the long-term stable operation of the system and the continued trust of users.

[0023] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention achieves improved selection efficiency and accuracy through the deep integration of large language model semantic understanding and structured constraint modeling. Automated requirement analysis and multi-agent collaborative reasoning compress the traditional selection cycle of several weeks to the minute level. In terms of accuracy, the multi-level constraint satisfaction check mechanism performs hierarchical modeling and logical reasoning on hard constraints and soft preferences, avoiding the structural defects of traditional linear weighted models. In the AI ​​large model selection standard test set, the recommendation results of this system are higher than those of the baseline system based on keyword matching, which fully demonstrates the substantial progress of this invention in selection accuracy.

[0024] 2. This invention achieves a leapfrog improvement over existing technologies in two dimensions: knowledge timeliness and global recommendation. Through incremental updates and event-triggered synchronization mechanisms, the system achieves near real-time synchronization with official product documentation, open-source communities, and third-party evaluation platforms, eliminating the risks caused by outdated knowledge. At the same time, the original synergy effect analysis module, through compatibility matrix and functional complementarity quantification model, realizes for the first time in the selection system the overall effectiveness evaluation of multi-product combination solutions, avoiding hidden risks and improving the success rate of the implementation and conversion of selection solutions.

[0025] 3. This invention also makes progress in terms of interpretability and system scalability. The system constructs a transparent output system covering the entire chain of results, reasons, and evidence through three major mechanisms: natural language recommendation reason generation, multi-dimensional scoring radar chart visualization, and confidence level labeling. At the architectural level, the multi-agent collaborative framework decomposes complex selection tasks into specialized sub-tasks, with each agent working together. This allows the system to quickly adapt to different product categories or industry scenarios by dynamically adjusting the evaluation focus, demonstrating good versatility and industrialization promotion value. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall architecture of the AI ​​product intelligent selection assistant system integrating a large language model according to the present invention; Figure 2 This is a flowchart of the user intent understanding and requirement parsing module in this invention; Figure 3 This is a flowchart illustrating the construction process of the multi-source product knowledge base in this invention. Figure 4 This is a schematic diagram of the RAG engine module architecture in this invention; Figure 5 This is a flowchart of the multi-dimensional constraint reasoning and scoring module in this invention; Figure 6 This is a schematic diagram of the multi-agent collaborative selection module architecture and interaction in this invention; Figure 7 This is a schematic diagram of the synergistic effect analysis module in this invention; Figure 8 This is a schematic diagram illustrating the recommended output interface in this invention. Detailed Implementation

[0027] 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.

[0028] like Figures 1-8 As shown, the AI-powered intelligent product selection assistant system integrating a large language model of the present invention comprises seven core modules: a user intent understanding and demand parsing module, a multi-source product knowledge base construction and management module, a retrieval enhancement generation engine module, a multi-dimensional constraint reasoning and scoring module, a multi-agent collaborative selection module, a synergy effect analysis module, and an interpretable recommendation output module. These modules are sequentially connected to form a complete intelligent selection closed loop from demand input to selection result output. The technical solutions for each module are described in detail below.

[0029] The user intent understanding and requirement parsing module is configured to receive product selection requirements input by the user in natural language, call a large language model for intent recognition and entity extraction, and output a structured requirement representation. This module is the system's entry point, responsible for transforming the user's vague, colloquial natural language descriptions into precise, computable structured requirement data. Its processing quality directly affects the inference accuracy of all subsequent modules. The user intent understanding and requirement parsing module specifically includes the following four units: The intent recognition unit classifies user input intents using a fine-tuned large language model. Specifically, it first constructs a labeled dataset containing multiple categories of selection intents, with 500 to 1000 typical user questions labeled for each category, covering five intent types: initial selection consultation, product comparison request, upgrade / migration consultation, technical solution verification, and procurement decision support. Then, based on this dataset, the base large language model is fine-tuned to enable accurate intent classification in selection scenarios. After user input, the intent recognition unit outputs intent category labels and corresponding confidence scores, providing scenario-adaptive parsing strategies for subsequent processing. For example, under a product comparison intent, the focus is on extracting the comparison object entity; under an upgrade / migration intent, the focus is on extracting current system information.

[0030] The entity extraction unit employs a named entity recognition model based on the Transformer architecture to extract seven categories of key entity information from user input, specifically including: Product type refers to the category of products required by the user, such as large language models, AI development platforms, intelligent customer service systems, computer vision solutions, etc. Functional requirements refer to the specific functional capabilities that users expect the product to have, such as text generation, image recognition, multimodal understanding, and code assistance. Performance metrics are the specific quantitative requirements of users for product performance, such as inference speed, accuracy, concurrency processing capability, and inference latency. Resource constraints are information limiting the resources available to the user, such as GPU model and quantity, video memory capacity, storage space, budget limit, etc. Deployment method, i.e. the product deployment form expected by users, such as public cloud SaaS, private deployment, hybrid deployment, etc. Compliance requirements refer to users' specific needs for product compliance, such as data security levels, commercial license types, and open-source protocol requirements.

[0031] The entity extraction unit is fine-tuned using a domain-annotated corpus based on a general pre-trained model, enabling the model to accurately identify the seven types of entities and their semantic relationships. To improve the recall rate of entity recognition, the entity extraction unit also employs few-shot extraction based on a large language model as an auxiliary pathway, generating multiple extraction results for the same input, which are then fused through voting to output the final entity set.

[0032] The context memory unit is responsible for maintaining the coherence and consistency of the dialogue across multiple interactions, recording the user's historical selection preferences and historical dialogue information. Specifically, the context memory unit maintains a session state cache, storing user preference information identified in the current session in key-value pairs, such as "prefers open-source solutions," "budget not exceeding 500,000," and "technology stack is the Python ecosystem," as well as a complete summary of the dialogue history before this round. When the user engages in multiple rounds of interaction, the context memory unit concatenates historical preferences and historical dialogue information into the input context of the current round, enabling the large language model to fully consider the user's long-term preferences and dialogue context when understanding the current needs, and avoiding the user repeating mentioned needs and conditions in long dialogues.

[0033] The structured requirements generation unit aggregates and integrates the outputs of the intent recognition, entity extraction, and contextual memory units, transforming them into a standardized requirements description document. This document is organized in a structured format such as JSON or XML and includes user intent tags, entity lists, historical preference summaries, and a list of constraints automatically generated based on this information. This standardized requirements description document serves as the final output of this module, providing a unified, complete, and computable input for downstream knowledge retrieval, constraint reasoning, and multi-agent collaboration.

[0034] The multi-source product knowledge base construction and management module is configured to collect multi-dimensional information about AI products, construct a knowledge graph containing product nodes and relationship edges, establish a vectorized index, and incrementally update the knowledge base at regular intervals. This module serves as the system's data infrastructure, providing comprehensive, accurate, and real-time product knowledge support for selection and reasoning.

[0035] At the data acquisition level, the multi-source product knowledge base construction and management module collects product information through the following channels: official product documentation and API documentation, technical evaluation reports and benchmark test data, open-source community information, commercial license agreement texts, user reviews and case studies, pricing information and billing models, etc. The acquisition methods include two types: automated acquisition based on web crawlers and near real-time synchronization based on API interfaces. The former is used to batch acquire static document-type information, while the latter is used to synchronize dynamic data such as pricing information and version status.

[0036] At the knowledge graph construction level, the multi-source product knowledge base construction and management module unifies the collected structured and unstructured data into a product knowledge graph. Node types in the graph include product entities, functional entities, technical indicator entities, application scenario entities, and license type entities. Relationship edge types in the graph include compatibility relationships, dependency relationships, competition relationships, version evolution relationships, and applicable scenario associations. The knowledge graph is stored using a graph database, supporting efficient graph traversal queries and relationship reasoning.

[0037] At the vectorized index construction level, the multi-source product knowledge base construction and management module segments unstructured texts such as product description documents, technical white papers, and user cases, and uses the Embedding model to generate a high-dimensional vector representation of each text block to build a vector database index for subsequent semantic similarity retrieval.

[0038] At the knowledge update level, the multi-source product knowledge base construction and management module adopts a dual update strategy combining scheduled tasks and event-triggered mechanisms. Scheduled tasks perform full or incremental synchronization at preset intervals, retrieving the latest product information from various data sources and updating the knowledge graph and vector index. The event-triggered mechanism monitors official product release channels and triggers incremental updates immediately upon detecting key events such as product version updates, performance evaluation releases, or price changes. Through these mechanisms, the knowledge base's timeliness window is compressed from days to hours, eliminating the risk of recommendation failure due to outdated knowledge.

[0039] The retrieval enhancement generation engine module is configured to retrieve candidate product information from the knowledge base based on the structured requirement representation, using a multi-path retrieval strategy. It then reorders and fuses the retrieval results to generate selection context information. This retrieval enhancement generation engine module acts as a bridge between the knowledge base and the inference module, responsible for transforming user requirements into precise knowledge retrieval and organizing it into context information that can be directly used by a large language model. Specifically, the retrieval enhancement generation engine module employs the following three-path hybrid retrieval strategy: The first-way search is a structured search based on Structured Query Language (SCL), used for precise matching of product specification parameters. This search transforms numerical and categorical constraints in the structured requirement representation into SQL or Cypher query statements, performing precise matching in the knowledge graph and structured database. For example, when user requirements include precise conditions such as "GPU requirement is at least A100 40GB," "inference latency is less than 100 milliseconds," and "commercial license type is Apache 2.0," the first-way search can directly locate product entities that meet all precise conditions. The advantage of structured search lies in its accuracy and determinism, ensuring that hard constraints are met during the retrieval stage.

[0040] The second-path retrieval is a semantic retrieval based on vector similarity, used to recall semantically relevant product description text. This second-path retrieval combines the functional and scenario descriptions from the structured requirement representation into query text, encodes it into query vectors using an embedding model, and then performs a nearest neighbor search in a vector database to recall the product description text fragments with the highest vector similarity. The advantage of semantic retrieval lies in its ability to discover product information that is semantically relevant to user needs but does not have precise keyword matches, compensating for the semantic understanding blind spots of structured retrieval and improving the retrieval recall rate.

[0041] The third-path retrieval is a knowledge graph-based graph retrieval used to obtain information about relationships between products. Starting with the candidate product entities returned by the first and second-path retrievals, the third-path retrieval performs a graph traversal query within the knowledge graph to obtain information such as compatibility, dependencies, competition, and version evolution relationships among the candidate products. The advantage of graph retrieval lies in its ability to discover relational information that cannot be obtained through isolated retrieval, providing relational data support for subsequent synergy analysis and combination scheme recommendations.

[0042] The three search results are first deduplicated and then aggregated into the re-ranking unit. The re-ranking unit adopts a re-ranking model based on the Cross-Encoder architecture. The user's query text is used as the query and each search result is used as a candidate document. The model scores the relevance of each query-document pair. The original similarity scores of the three searches and the relevance scores of the re-ranking model are combined to finally rank the candidate documents and select the Top-K most relevant search results.

[0043] The multi-dimensional constraint reasoning and scoring module is configured to perform hard constraint satisfaction filtering on candidate products based on the constraints in the structured requirement representation, and to calculate scores on seven dimensions: functional matching degree, performance adaptability, resource adaptability, cost-effectiveness, availability, security compliance and scalability, to generate a comprehensive selection score.

[0044] The multi-dimensional constraint reasoning and scoring module first parses all constraints from the structured requirement representation to identify all user requirements, categorizing them into hard constraints and soft constraints based on their degree of enforceability. Hard constraints are mandatory conditions that must be met, such as commercial license types complying with enterprise compliance requirements, GPU models being compatible with existing hardware, and data not being uploaded to the cloud for deployment. If a candidate product fails to meet any hard constraint, it is excluded from the candidate set, leaving no room for compromise or trade-offs. Soft constraints are preference conditions that can be optimized through trade-offs, such as prioritizing inference speed, minimizing costs, and favoring open-source solutions. The degree to which soft constraints are met is allowed to vary, allowing different users to assign different priorities to different soft constraints based on their preferences.

[0045] The hard constraint satisfaction filtering unit executes according to the following logic: each hard constraint is represented as a decision function, all candidate products are traversed sequentially, and only products that satisfy all hard constraint decision functions are retained, forming a candidate set. This filtering logic ensures that every product in the candidate set meets all the user's mandatory requirements in terms of basic feasibility, avoiding the problem in traditional recommendation systems where a low score is masked by other high scores, leading to the failure to meet the basic conditions of the recommendation scheme.

[0046] For each product in the candidate set, the multi-dimensional constraint reasoning and scoring module calculates a score across the following seven dimensions, with each dimension having a maximum score of 100: The functional matching score is based on the semantic matching degree between the product functional list recorded in the product knowledge base and the functional requirement entities in the user's requirements. The matching score is calculated using a large language model or semantic similarity model to reflect whether the product function meets the user's core business needs.

[0047] The performance fit score is based on benchmark test data in the product knowledge base, such as performance indicators like inference speed, accuracy, and concurrency capabilities. It is compared item by item with the performance indicators required by the user's needs, and the compliance rate and the degree of exceeding the performance indicators are calculated to reflect whether the product meets or exceeds the user's expectations in terms of performance.

[0048] Resource compatibility score is calculated based on the degree of matching between the hardware resources required for product deployment, such as GPU model and quantity, video memory capacity, and storage space, as well as software environment requirements, such as operating system version and dependent library version, and the user's existing environment, reflecting the product's deployability on the user's existing technical infrastructure.

[0049] Cost-benefit scoring comprehensively considers the cost-effectiveness relationship between the product's procurement cost, usage cost, maintenance cost and product functional value, as well as the cost elasticity of the product under different scales of use, reflecting the economic rationality of the product under the user's budget constraints.

[0050] Usability scores are calculated based on a combination of factors, including the completeness of product documentation (coverage, API documentation quality, etc.), community activity (GitHub stars, issue response speed, technical Q&A coverage, etc.), and ease of use evaluation, reflecting the product's ease of use and daily user experience.

[0051] The security compliance score is calculated based on a comprehensive assessment of factors such as whether the product has passed relevant security certifications, including Information Security Level Protection and ISO 27001, data security protection mechanisms, including data encryption, access control, and audit logs, and the degree of matching between commercial license terms and corporate compliance requirements. It reflects the product's usability within the corporate compliance framework.

[0052] The scalability score is calculated based on a comprehensive evaluation of factors such as whether the product provides rich API interfaces and SDK support, whether it supports plug-in extensions and custom development, whether it has a clear version upgrade path and ecosystem integration capabilities, reflecting the product's ability to support future business growth and functional evolution.

[0053] The comprehensive scoring calculation unit uses a weighted summation method to calculate the comprehensive selection score for each candidate product. The calculation formula is as follows: Score total = w F · F score + w P · P score + w R · R score + w C · C score + w U · U score + w S ·S score + w E · E score

[0054] Each dimension's weight is a non-negative number, and their sum equals 1. The initial values ​​of the weights are set based on domain expert experience. For example, in performance-sensitive scenarios, the weight of performance fit score is increased, and in cost-sensitive scenarios, the weight of cost-effectiveness score is increased. At the same time, the weights are dynamically adjusted according to the user's preference for various soft constraints in the structured requirement representation. For example, if the user explicitly states "limited budget, cost control is the priority," then the weight of the cost-effectiveness score is increased accordingly. Users can also adjust the weights of each dimension themselves through the subsequent interactive interface to explore the selection results under different preferences.

[0055] The multi-agent collaborative selection module is the core inference engine of this system, comprising six dedicated agents: a needs analysis agent, a candidate screening agent, a deep evaluation agent, a comparative analysis agent, a solution recommendation agent, and a coordination and supervision agent. Each agent is built upon a large language model and collaborates through unified scheduling by the coordination and supervision agent, ultimately outputting selection results including recommended solutions and their rationale. The specific functions and collaboration processes of each agent are as follows: The coordinating and supervising agent, acting as the global scheduling center, is responsible for task decomposition, communication coordination among agents, and result aggregation. Specifically, this includes: receiving user requirements and determining the task's complexity and type; decomposing the overall selection task into independently executable sub-tasks; assigning each sub-task to the corresponding specialized agent; monitoring the execution progress and output quality of each agent; adjudicating conflicts or inconsistencies in the conclusions output by different agents; and finally aggregating the outputs from all parties, reviewing the overall quality of the recommended solution, and deciding whether further evaluation or re-execution is necessary.

[0056] The requirements analysis agent is responsible for deeply analyzing user needs and identifying implicit needs. After receiving structured requirements representations, the agent not only understands the explicitly stated needs but also infers logically implicit requirements that are not explicitly expressed. For example, when a user states, "We need to process a large number of legal documents," the agent can infer implicit needs such as "support for long context windows" and "the ability to parse documents." When a user states, "The team only has 3 AI engineers," the agent can infer implicit needs such as "the product needs to have good usability and complete documentation." The results of identifying implicit needs are added to the requirements analysis report for subsequent use by the agent.

[0057] The candidate selection agent filters the initial candidate product list based on hard constraints. After receiving the requirements analysis report, the agent applies all hard constraints, such as "must support private deployment," "license agreement must meet commercial requirements," and "cost must be below the budget limit," to all products in the product knowledge base, performing the aforementioned hard constraint satisfaction filtering, and outputting an initial candidate product list that passes all hard constraint checks. The agent can also perform preliminary sorting and labeling of the candidate list, such as marking the specific constraints that each candidate product passes and fails.

[0058] The deep evaluation agent performs fine-grained evaluations from the dimensions of functionality, performance, cost, and compatibility. After receiving the initial list of candidate products, the deep evaluation agent retrieves detailed technical documents, evaluation data, and user cases from the knowledge base for each candidate product in the list. It conducts in-depth qualitative and quantitative analysis according to four dimensions: functional completeness, performance benchmark compliance rate, total cost of ownership, and technology stack compatibility, and generates a detailed evaluation report for each candidate product.

[0059] The comparative analysis agent is responsible for the horizontal comparative analysis between candidate products. After receiving the evaluation reports of each candidate product output by the deep evaluation agent, the comparative analysis agent organizes the comparative information in a unified comparative framework, including a comparison table of basic product information, a radar chart comparing scores in each dimension, a bar chart comparing key performance indicators, a SWOT analysis of strengths and weaknesses, etc., to generate horizontal comparative results between candidate products and clearly identify the relative strengths and weaknesses of each product in different dimensions.

[0060] The solution recommendation agent generates the final recommended solution based on the comprehensive evaluation results. The solution recommendation agent integrates the comprehensive requirements analysis report, in-depth evaluation report, horizontal comparison results and the comprehensive score output by the multi-dimensional constraint reasoning and scoring module. It adopts the reasoning ability of the large language model and generates the final selection recommendation solution according to the three-layer structure of "the preferred solution is the main product and core reasons, the alternative solution is the second best product and applicable conditions, and the unrecommended product is the reason for explicit exclusion".

[0061] The multi-agent collaborative selection module executes according to the following sequence: The coordinating and supervising agent receives user requirements, makes an overall judgment on the task, decomposes the task, and distributes the requirement data to the requirement analysis agent; the requirement analysis agent performs in-depth requirement analysis, generates a structured requirement report containing explicit and implicit requirements, and returns it to the coordinating and supervising agent; the coordinating and supervising agent forwards the requirement report to the candidate screening agent, which performs hard constraint filtering, generates an initial candidate product list, and returns it to the coordinating and supervising agent; the coordinating and supervising agent distributes the requirement report and candidate product list to the deep evaluation agent, which concurrently performs fine-grained evaluations on each candidate product, generating... Detailed evaluation reports for each product are compiled and returned to the coordinating and supervising agent. The coordinating and supervising agent distributes the evaluation reports to the comparative analysis agent, which generates horizontal comparison results and returns them to the coordinating and supervising agent. The coordinating and supervising agent distributes all the above intermediate results to the solution recommendation agent, which integrates the outputs from all parties to generate the final recommended solution and returns it to the coordinating and supervising agent. The coordinating and supervising agent conducts a quality review of the recommended solution, and outputs the approved solution as the selection result. If problems are found during the review, such as insufficient evidence to support the recommendation reasons or logical contradictions in the conclusions of the various agents, the coordinating and supervising agent instructs the relevant agents to conduct supplementary evaluations or re-execute the evaluation.

[0062] The synergy analysis module is configured to calculate synergy scores for combination schemes containing multiple candidate products based on the compatibility and functional complementarity between products, and generate recommended combination product schemes. This synergy analysis module addresses typical scenarios where enterprise users often need to combine multiple AI products, such as large models, vector databases, and application development frameworks, to form complete solutions, enabling an overall performance evaluation of multi-product combination schemes.

[0063] Specifically, for the candidate product combination P = {p1, p2, ..., p...} k}, where k≥2, the synergy score is calculated as follows: First, construct the product compatibility matrix. Based on the product compatibility relationships stored in the knowledge graph, extract the compatibility coefficient Compatibility(i,j) of each product pair in the combination, representing product p i With product p j The compatibility level of the combined technologies ranges from 0 to 1. The compatibility coefficient is determined based on factors such as: consistency of API interface standards, interoperability of data formats, uniformity of operation and maintenance management tools, and whether official compatibility certification has been obtained. When two products are known to have compatibility issues, the compatibility coefficient tends towards 0; when two products have been fully verified to be seamlessly integrated, the compatibility coefficient tends towards 1.

[0064] Secondly, calculate the functional complementarity (Complementary(i,j)) between the products, with a value ranging from 0 to 1. The criteria for determining complementarity include: whether the two products have overlapping or complementary functions; whether their combined use can form a complete functional loop from data input to output; and whether they occupy different layers in the solution architecture, such as infrastructure, platform, or application layers. When the two products have highly overlapping functions, like being large models of the same type, the complementarity value is low; when the two products belong to different functional layers and cooperate to form a complete solution, the complementarity value is high.

[0065] Then, the synergistic effect score of the combination scheme P is calculated according to the following formula: Synergy(P) = This formula means that the synergy effect of a combination scheme is equal to the average of the products of compatibility and complementarity of all product pairs in the combination. When the products in the combination have good compatibility and complementary functions, the synergy effect score is high, indicating that the combination scheme can form a complete solution with functional closed loop, data connectivity, and unified interface standards. When there are compatibility issues or functional overlap and redundancy in the combination, the synergy effect score is low, indicating that the combination scheme may have architectural conflicts or functional waste.

[0066] Finally, based on the optimal recommendation of a single product, the synergy effect analysis module enumerates various combinations of candidate products. Depending on the combination size k, it enumerates two-product combinations, three-product combinations, etc., calculates the synergy effect score for each combination, and comprehensively considers the independent scores of each product and the synergy effect score of the combination. According to the fusion ranking logic that the comprehensive effectiveness equals the weighted sum of the independent scores plus the synergy effect score, the top-ranked combination product recommendation schemes are generated for enterprise users to make complete technology stack selection decisions.

[0067] The explainable recommendation output module is configured to present the selection results to the user in the form of visual charts, natural language recommendation reasons, and confidence level annotations. The explainable recommendation output module is the final output layer of the system, responsible for transforming the selection results generated by the multi-agent collaborative selection module and the synergy effect analysis module into recommendations that users can intuitively understand, fully trust, and effectively use. The explainable recommendation output module specifically includes the following four units: The recommendation reason generation unit utilizes a large language model to generate natural language recommendation reasons. Based on the selection recommendation scheme as input, the unit extracts the key advantages of each recommended product, such as "leading inference speed among similar products," "commercial licensed under Apache 2.0, fully compliant with enterprise compliance requirements," and "possesses excellent Chinese processing capabilities, suitable for legal text analysis scenarios." Each advantage is mapped one-to-one with the corresponding constraints in the user's needs; for example, "Your low latency requirement—this product's inference latency is only 80 milliseconds, fully met," generating a structured recommendation reason explanation. For each recommended product, the recommendation reason includes at least three parts: a description of key advantages, i.e., in which dimensions the product excels; a description of matching user needs, i.e., how the product meets the user's specific needs; and a hint of potential limitations, i.e. in which dimensions the product performs moderately or has limitations, allowing users to make informed decisions.

[0068] The comparative visualization unit generates radar charts or bar charts to display the score comparisons of candidate products across various dimensions. Specifically, the comparative visualization unit overlays the score data of each candidate product across seven dimensions—functionality, performance, resource suitability, cost-effectiveness, usability, security compliance, and scalability—in radar chart format on the same coordinate system, allowing users to intuitively compare the overall performance of each product in seven dimensions and their respective strengths and weaknesses. Simultaneously, key performance indicators, such as inference speed, accuracy, and cost, are compared intuitively using grouped bar charts. All visualization charts support interactive operations such as zooming and hovering to view values ​​within the user interface.

[0069] The interactive exploration interface allows users to adjust the weights or constraints of various dimensions to view changes in selection results in real time. While displaying recommendation results, the interface provides users with visual weight adjustment sliders corresponding to the weight adjustments of the seven dimensions, as well as constraint switches, such as "adjusting the budget limit from 500,000 to 300,000" or "adding compatibility requirements for domestically produced chips." When a user adjusts any weight or constraint, the system re-executes constraint filtering and comprehensive score calculation in the background, and refreshes the recommendation ranking and comparison charts in real time on the interface. This supports users in conducting scenario analysis, exploring the differences in selection results under different preferences and constraints, and assisting users in making their final selection decisions.

[0070] The confidence score annotation unit labels each recommendation result with a confidence score and its influencing factors. The confidence score is a value ranging from 0 to 1, comprehensively reflecting the reliability of the recommendation result. Factors influencing the confidence score include: the amount of information explicitly stated in the user's needs (the more complete the information, the higher the confidence score); the completeness of the candidate product information in the knowledge base (the more detailed the product documentation and the more comprehensive the evaluation data, the higher the confidence score); and the timeliness of the knowledge base information (the more recently the product information was updated, the higher the confidence score). When outputting recommendation results, the confidence score annotation unit simultaneously presents the confidence score and detailed explanations of each influencing factor, such as "This recommendation has a confidence score of 0.87, indicating sufficient information support; the lowering factor is that the price information for candidate product A was updated 90 days ago, it is recommended to verify the latest pricing," enabling users to objectively assess the reliability of the recommendation results and to manually verify parts with lower confidence scores.

[0071] As shown in the figure, the present invention also provides an AI product intelligent selection method using the above system. This method takes each module of the system as the execution subject and executes the following seven standardized steps in sequence according to the complete logical link from demand input to solution output, so as to transform the complex product selection decision process into a standardized process that can be automated, traceable and optimized.

[0072] Step S1 is executed by the user intent understanding and requirement parsing module. It receives the user's product selection requirements input in natural language, performs intent recognition and entity extraction using a large language model, and generates a structured requirement representation. Specifically, Step S1 first classifies the user input intent using a fine-tuned large language model, identifying which type of user's current selection scenario it belongs to: initial selection consultation, product comparison request, upgrade / migration consultation, technical solution verification, or procurement decision support. This provides a scenario-adaptive parsing strategy for subsequent processing. Then, a named entity recognition model based on the Transformer architecture is used to extract seven key entity information categories from the user input, including product type, functional requirements, performance indicators, resource constraints, deployment methods, and compliance requirements. This is combined with few-sample extraction based on the large language model as an auxiliary pathway to improve recall. Subsequently, a context memory unit maintains dialogue coherence across multiple interactions and records the user's historical selection preferences and historical dialogue information. Finally, all the above parsing results are summarized and integrated, transforming them into a standardized requirement description document organized in a structured format such as JSON or XML for use by downstream modules. Step S1 transforms vague, colloquial natural language requirements into precise, computable structured data from the source. The quality of this processing directly affects the inference accuracy of all subsequent steps.

[0073] Step S2 is executed by the multi-source product knowledge base construction and management module, which is used to build a multi-source product knowledge base and collect multi-dimensional information of AI products to construct a knowledge graph and vector index. Specifically, step S2 collects product information through multiple channels, including official product documents and API documents, technical evaluation reports and benchmark test data, open source community information, commercial license agreement texts, user reviews and case studies, pricing information and billing models, etc. The collection methods include automated collection based on web crawlers and near real-time synchronization based on API interfaces. The collected structured and unstructured data are uniformly constructed into a product knowledge graph. The graph nodes include product entities, functional entities, technical indicator entities, application scenario entities, license type entities, etc., and the relationship edges include compatibility relationships, dependency relationships, competition relationships, version evolution relationships, applicable scenario associations, etc. A graph database is used for storage to support efficient graph traversal query and relationship reasoning. At the same time, unstructured texts such as product description documents, technical white papers, and user cases are segmented into text, and a high-dimensional vector representation of each text block is generated using an embedding model to construct a vector database index for subsequent semantic similarity retrieval. The construction of the knowledge base is a continuous iterative process. For details on its incremental update mechanism, please refer to the subsequent explanation of knowledge updates. Step S2 provides a comprehensive, accurate, and real-time product knowledge infrastructure for the entire selection process.

[0074] Step S3 is executed by the retrieval enhancement generation engine module, which retrieves candidate product information from the knowledge base based on the structured requirement representation using a three-way hybrid retrieval strategy, and generates selection context information after reordering and knowledge fusion. Step S3 is the crucial bridge connecting the knowledge base and the reasoning step. It employs a three-pronged retrieval strategy executed in parallel: The first prong is structured retrieval based on a structured query language, transforming numerical and categorical constraints in the structured requirement representation into SQL or Cypher query statements. This is then precisely matched in the knowledge graph and structured database to ensure that hard constraints are met during the retrieval phase. The second prong is semantic retrieval based on vector similarity, combining functional and scenario descriptions from the structured requirement representation into query text. This text is then encoded into query vectors using an embedding model and searched for nearest neighbors in the vector database to retrieve semantically relevant product description text fragments, compensating for the semantic blind spots of structured retrieval. The third prong is graph retrieval based on the knowledge graph, using the candidate product entities returned by the first two prongs as a starting point to perform graph traversal queries in the knowledge graph, obtaining information such as compatibility, dependency, competition, and version evolution relationships among candidate products. After deduplication, the three search results are aggregated into the re-ranking unit. The re-ranking model based on the Cross-Encoder architecture scores the relevance of each query-document pair. The final ranking is performed by combining the original similarity scores of the three searches and the relevance scores of the re-ranking model, and the Top-K most relevant results are selected. Finally, the knowledge fusion unit extracts key attributes from the Top-K search results to form structured product cards, summarizes the relationships between products to form a relationship list, and organizes key paragraphs to form supporting evidence. The selection context information is output in the form of a structured document. Step S3 takes into account both the recall and precision of the search by leveraging the complementary advantages of multi-path retrieval and the precise selection of re-ranking.

[0075] Step S4 is executed by the multi-dimensional constraint reasoning and scoring module. Based on the constraints in the structured requirement representation, it performs hard constraint filtering on candidate products and provides a comprehensive score across seven dimensions: functional matching, performance adaptability, resource adaptability, cost-effectiveness, usability, security compliance, and scalability. Step S4 first parses all constraints from the structured requirement representation, categorizing them into hard and soft constraints according to their degree of enforceability. Hard constraints are mandatory conditions that must be met, such as commercial license types meeting enterprise compliance requirements or GPU models being compatible with existing hardware. Soft constraints are preference conditions that can be optimized through trade-offs, such as maximizing inference speed or minimizing cost. Then, each hard constraint is represented as a decision function. All candidate products are iterated through sequentially, and only products satisfying all hard constraint decision functions are retained to form a candidate set. This ensures that each product in the candidate set meets all the user's mandatory requirements in terms of basic feasibility, avoiding the problem in traditional recommendation systems where a low score is masked by other high scores, leading to unmet basic conditions in the recommended solution. Then, each product in the candidate set is scored across seven dimensions, with each dimension having a maximum score of 100: Functionality matching score is calculated based on the semantic matching between the product's feature list and user functional requirements; performance adaptability score is calculated by comparing product benchmark test data with user performance requirements item by item; resource adaptability score is calculated based on the matching between the hardware resources and software environment required for product deployment and the user's existing environment; cost-effectiveness score comprehensively considers the cost-effectiveness relationship between procurement costs, usage costs, maintenance costs, and functional value; usability score is calculated based on a comprehensive assessment of documentation completeness, community activity, and ease of use; security compliance score is calculated based on a comprehensive assessment of the matching between security certification, data security protection mechanisms, and commercial licensing terms; and scalability score is calculated based on a comprehensive assessment of API richness, plug-in extension capabilities, version upgrade paths, and ecosystem integration capabilities. Finally, a weighted summation method is used to calculate the overall selection score. The initial weights for each dimension are set based on domain expert experience and dynamically adjusted according to user preferences for various soft constraints. Step S4 uses a hierarchical evaluation mechanism combining hard constraint filtering and soft constraint weighting to ensure that the recommendation results satisfy both mandatory conditions and maximize the satisfaction of user preferences.

[0076] Step S5 is executed by the multi-agent collaborative selection module, which performs requirement analysis, candidate screening, in-depth evaluation, comparative analysis, and solution recommendation through multi-agent collaboration to generate a selection recommendation solution. Step S5 adopts an architecture of six dedicated agents working together, each agent is built based on a large language model, and proceeds in an orderly manner according to a time sequence through the unified scheduling of the coordinating and supervising agents. This step specifically includes the following seven sub-processes: First, the coordinating and supervising agent receives the task, judges and decomposes the overall task, and distributes the requirement data to the requirement analysis agent; Second, the requirement analysis agent performs deep requirement analysis, not only understanding the user's explicitly stated requirements, but also further reasoning about the user's unexpressed but logically implicit requirements, such as "processing a large number of legal documents" and "needing to have document parsing capabilities," generating a structured requirement report containing explicit and implicit requirements, and returning it to the coordinating and supervising agent; Third, the coordinating and supervising agent forwards the requirement report to the candidate screening agent, which applies all the hard constraints to all products in the product knowledge base, performs hard constraint satisfaction filtering, generates an initial candidate product list, and returns it to the coordinating and supervising agent; Fourth, the coordinating and supervising agent distributes the requirement report and candidate product list to the deep evaluation agent, which concurrently performs fine-grained evaluations on each candidate product, conducting in-depth qualitative and quantitative analysis from four dimensions: functional completeness, performance benchmark compliance rate, total cost of ownership, and technology stack compatibility, generating... After generating detailed evaluation reports for each product, the reports are compiled and returned to the coordinating and supervising agent. Fifth, the coordinating and supervising agent distributes the evaluation reports to the comparative analysis agent. The comparative analysis agent generates horizontal comparison results using a unified comparison framework, including a comparison table of basic product information, radar charts comparing scores across dimensions, bar charts comparing key performance indicators, and SWOT analysis of strengths and weaknesses, and returns the results. Sixth, the coordinating and supervising agent distributes all the above intermediate results to the solution recommendation agent. The solution recommendation agent integrates the comprehensive score output from the demand analysis report, in-depth evaluation report, horizontal comparison results, and multi-dimensional constraint reasoning and scoring modules. Using the reasoning ability of a large language model, it generates the final selection recommendation solution according to a three-layer structure: the preferred solution (i.e., the main recommended product and its core reasons), the alternative solution (i.e., the second-best product and its applicable conditions), and the unrecommended product (i.e., the explicitly excluded product), and returns the result. Seventh, the coordinating and supervising agent conducts a quality review of the recommended solution. Once the review is passed, it is output as the selection result. If problems are found during the review, such as insufficient evidence to support the recommendation reasons or logical contradictions in the conclusions of different agents, the relevant agents are instructed to supplement the evaluation or re-execute the solution. Step S5 decomposes the complex selection reasoning task into multiple independently executable specialized sub-tasks through task specialization and sequential collaboration among intelligent agents, balancing reasoning depth and processing efficiency.

[0077] Step S6 is executed by the synergy effect analysis module, which performs synergy effect analysis on combination schemes containing multiple candidate products and generates recommended combination product schemes. Step S6 addresses the typical scenario where enterprise users often need to combine multiple AI products to form a complete solution, and achieves an overall performance evaluation of multi-product combination schemes. Specifically, for candidate product combinations P={p1,p2,...,p...} k}, where k≥2, firstly, based on the compatibility relationships between products stored in the knowledge graph, the compatibility coefficient Compatibility(i,j) of each product pair in the combination is extracted, representing the compatibility coefficient of product p. i With product p j The technical compatibility of the combined products is calculated, with a value ranging from 0 to 1. This is determined based on factors such as whether API interface standards are consistent, data formats are interoperable, operation and maintenance management tools are unified, and whether official compatibility certification has been obtained. Then, the functional complementarity between the products is calculated, with a value ranging from 0 to 1. This is determined based on factors such as whether the two products have overlapping or complementary functions, whether their combined use can form a complete functional loop from data input to result output, and whether they occupy different levels in the solution architecture. Finally, the formula Synergy(P) = ... The synergy score of the combination scheme is calculated. This synergy score is equal to the average of the products of compatibility and complementarity of all product pairs in the combination. A high score indicates that the combination scheme can form a complete solution with functional closed loop, data connectivity, and unified interface standards. A low score indicates that the combination scheme may have architectural conflicts or functional waste. Finally, based on the optimal recommendation of individual products, multiple combination methods between candidate products are enumerated. Depending on the combination size k, two-product combinations, three-product combinations, etc., are enumerated, and the synergy score of each combination is calculated. Taking into account the independent scores of each product and the combination synergy score, the top-ranked combination product recommendation schemes are generated according to the fusion ranking logic that the overall effectiveness equals the weighted sum of independent scores plus the synergy score. Step S6 uses synergy analysis to make up for the shortcomings of traditional selection methods that only focus on the independent indicators of individual products, and avoids the selection trap of optimal individual product scores but architectural conflicts after combination.

[0078] Step S7 is executed by the interpretable recommendation output module, which presents the selection results to the user in the form of visual charts and natural language recommendation reasons. Step S7 is the final output stage of the selection process, responsible for transforming the selection results generated by the multi-agent collaborative selection module and the synergy analysis module into a recommendation presentation that users can intuitively understand, fully trust, and effectively use. Specifically, Step S7 first uses a large language model as the input based on the selection recommendation scheme to extract the key advantages of each recommended product, mapping each advantage feature to the corresponding constraints in the user's needs, generating a structured explanation of the recommendation reasons. For each recommended product, it should include at least three parts: a description of key advantages (i.e., in which dimensions the product performs outstandingly), a description of matching user needs (i.e., how the product meets the user's specific needs), and a hint of potential limitations (i.e., in which dimensions the product performs poorly or has limitations) to inform the user's informed decision-making. Then, the scoring data of each candidate product in seven dimensions are displayed in the form of a radar chart overlaid on the same coordinate system, and key performance indicators are compared intuitively in the form of grouped bar charts. All visual charts support zooming. The system allows for interactive operations such as hovering to view numerical values; it also provides users with visual weight adjustment sliders corresponding to the weight adjustments of seven dimensions and constraint condition switches. When a user adjusts any weight or constraint condition, the system re-executes constraint filtering and comprehensive score calculation in the background and refreshes the recommendation ranking and comparison charts in real time on the interface. This supports users in conducting scenario analysis to explore the differences in selection results under different preferences and constraints. Finally, each recommendation result is labeled with a confidence score in the range of 0 to 1, comprehensively considering factors such as the amount of information explicitly stated in the user's needs, the completeness of candidate product information in the knowledge base, and the timeliness of the knowledge base information. Detailed explanations of each influencing factor are also presented, enabling users to objectively assess the reliability of the recommendation results and manually verify parts with low confidence scores. Step S7 constructs a transparent output system covering the entire chain of "result-reason-basis-credibility," bridging the cognitive gap between intelligent systems and human decision-makers.

[0079] The multi-source product knowledge base described in step S2 above is incrementally updated through scheduled tasks and event-triggered mechanisms, automatically synchronizing product version updates, performance evaluation changes, and price changes. Continuous updates to the knowledge base are a key mechanism to ensure the long-term effective operation of the system. Specifically, this incremental update mechanism employs a dual-trigger strategy: on the one hand, scheduled tasks perform full or incremental synchronization at preset intervals, obtaining the latest product information from various data sources and updating the knowledge graph and vector index. For example, the latest version number is retrieved from the official product website every morning, and the latest benchmark test data is synchronized from the technical evaluation platform weekly. On the other hand, an event-triggered mechanism monitors official product release channels, including RSS subscriptions, version release webhooks, and technical community announcements. Incremental updates are triggered immediately when key events such as product version updates, performance evaluation releases, or price changes are detected. For example, when a major model releases a new version, the system immediately triggers the synchronization update of that product information. The updated content covers multiple dimensions, including product version number and release date, new features and performance changes, price adjustment information, benchmark data updates, and changes in open-source community activity. Through the combined efforts of the aforementioned dual update strategies, the timeliness window of the knowledge base is compressed from the daily level to the hourly level, eliminating the risk of recommendation failure due to outdated knowledge. For example, it avoids recommending discontinued versions or making incorrect judgments based on outdated performance data, ensuring that selection suggestions are always based on the latest and most accurate product information, thus providing a guarantee for the long-term stable operation of the system and the continued trust of users.

[0080] In summary, this invention provides an AI-powered intelligent product selection assistant system integrating a large language model. It comprises seven modules forming a complete closed loop: user intent understanding, a multi-source knowledge base, a RAG retrieval engine, multi-dimensional constraint reasoning and scoring, multi-agent collaborative selection, synergy effect analysis, and explainable recommendation output. The system achieves deep analysis of natural language requirements through a large language model, employs a three-way hybrid retrieval strategy for accurate product knowledge retrieval, completes complex selection reasoning through the division of labor and collaboration of six dedicated intelligent agents, and introduces synergy effect scoring to achieve global optimization of multi-product combination solutions. This invention shortens the traditional manual selection cycle from weeks to minutes, solving technical challenges such as information overload, complex constraints, knowledge lag, and lack of collaboration in AI product selection, and possesses promising prospects for industrial application.

[0081] 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. An AI-powered product intelligent selection assistant system integrating a large language model, characterized in that: include: The user intent understanding and requirement parsing module is configured to receive product selection requirements input by users in natural language, call a large language model to perform intent recognition and entity extraction, and output a structured requirement representation. The multi-source product knowledge base construction and management module is configured to collect multi-dimensional information of AI products, construct a knowledge graph containing product nodes and relationship edges, establish a vectorized index, and incrementally update the knowledge base at regular intervals. The retrieval enhancement generation engine module is configured to retrieve candidate product information from the knowledge base based on the structured requirement representation through a multi-path retrieval strategy, and to reorder and fuse the retrieval results to generate selection context information. The multi-dimensional constraint reasoning and scoring module is configured to perform hard constraint satisfaction filtering on candidate products based on the constraints in the structured requirement representation, and to calculate scores on seven dimensions: functional matching degree, performance adaptability, resource adaptability, cost-effectiveness, availability, security compliance and scalability, and generate a comprehensive selection score. The multi-agent collaborative selection module includes a demand analysis agent, a candidate screening agent, a deep evaluation agent, a comparative analysis agent, a solution recommendation agent, and a coordination and supervision agent. The coordination and supervision agent is responsible for task decomposition and result summarization. Each agent performs task division and collaboration based on a large language model, and outputs selection results including selection recommendation schemes and recommendation reasons. The synergy analysis module is configured to calculate the synergy score for combination schemes containing multiple candidate products based on the compatibility relationship and functional complementarity between products, and generate a recommended combination product scheme. An interpretable recommendation output module is configured to present the selection results to the user in the form of visual charts, natural language recommendation reasons, and confidence level labels.

2. The AI ​​product intelligent selection assistant system integrating a large language model according to claim 1, characterized in that, The user intent understanding and requirement parsing module includes: The intent recognition unit classifies user input intent using a finely tuned large language model. The entity extraction unit uses a Transformer-based named entity recognition model to extract entity information such as product type, functional requirements, performance indicators, resource constraints, deployment methods, and compliance requirements. Context memory units maintain dialogue coherence across multiple interactions, recording users' historical selection preferences and historical dialogue information; The structured requirements generation unit transforms the parsed results into standardized requirements description documents.

3. The AI ​​product intelligent selection assistant system integrating a large language model according to claim 1, characterized in that, The retrieval enhancement generation engine module employs a three-way hybrid retrieval strategy: The first search method is a structured search based on structured query language, used to accurately match product specifications and parameters; The second retrieval method is semantic retrieval based on vector similarity, used to recall semantically relevant product description text; The third retrieval method is graph retrieval based on knowledge graphs, used to obtain information on relationships between products; The three search results are merged and sorted by the re-sorting unit to output the Top-K results.

4. The AI ​​product intelligent selection assistant system integrating a large language model according to claim 1, characterized in that, In the multi-dimensional constraint reasoning and scoring module: the constraints are divided into hard constraints and soft constraints according to their mandatory nature. Hard constraints are mandatory conditions that must be met, while soft constraints are preference conditions that can be optimized by trade-offs. The hard and soft constraints are obtained by parsing the structured requirement representation output by the user intent understanding and requirement parsing module. The multiple dimensions include seven dimensions: functional matching degree, performance adaptability, resource adaptability, cost-effectiveness, availability, security compliance, and scalability. The comprehensive score is the weighted sum of the scores of each dimension, and the weights are dynamically adjusted according to the user's preference for the soft constraints in the structured requirement representation.

5. The AI ​​product intelligent selection assistant system integrating a large language model according to claim 1, characterized in that, In the multi-agent collaboration selection module: the coordinating and supervising agent is responsible for task decomposition, communication coordination between agents, and result summarization; the demand analysis agent is responsible for in-depth analysis of user needs and identification of implicit needs. The candidate selection agent filters the initial list of candidate products based on hard constraints; the deep evaluation agent performs fine-grained evaluations from the dimensions of functionality, performance, cost, and compatibility; the comparative analysis agent is responsible for the horizontal comparative analysis between candidate products; and the solution recommendation agent generates the final recommended solution based on the comprehensive evaluation results.

6. The AI ​​product intelligent selection assistant system integrating a large language model according to claim 1, characterized in that, In the synergy effect analysis module, for the candidate product combination P={p1, p2, ..., p...} k The synergy score is calculated as follows: Construct a compatibility matrix (Compatibility(i,j)) among candidate products; Calculate the synergistic effect score: Synergy(P)= Complementary(i,j) represents the degree of complementarity of product functions.

7. The AI ​​product intelligent selection assistant system integrating a large language model according to claim 1, characterized in that, The explainable recommendation output module includes: The recommendation reason generation unit uses a large language model to generate natural language recommendation reasons; Comparing the visualization units, generate radar charts or bar charts to display the score comparison of candidate products across various dimensions; An interactive exploration interface allows users to adjust the scoring weights or constraints of various dimensions to view changes in the selection results in real time. The confidence score labeling unit labels each recommendation result with a confidence score and its influencing factors.

8. An AI product intelligent selection method using the system described in any one of claims 1-7, characterized in that, Each module of the system executes the following steps: S1: The User Intent Understanding and Requirement Parsing module receives the product selection requirements input by the user in natural language, performs intent recognition and entity extraction through a large language model, and generates a structured requirement representation. S2: Multi-source product knowledge base construction and management module: Constructs a multi-source product knowledge base, collects multi-dimensional information of AI products to build knowledge graphs and vector indexes; S3: The retrieval enhancement generation engine module retrieves candidate product information from the knowledge base based on the structured requirement representation through a three-way hybrid retrieval strategy, and generates selection context information after reordering and knowledge fusion; S4: The multi-dimensional constraint reasoning and scoring module performs hard constraint filtering on candidate products based on the constraints in the structured requirement representation and performs comprehensive scoring on seven dimensions: functional matching degree, performance adaptability, resource adaptability, cost-effectiveness, availability, security compliance and scalability. S5: The multi-agent collaborative selection module performs requirement analysis, candidate screening, in-depth evaluation, comparative analysis, and solution recommendation through multi-agent collaboration, generating a selection recommendation solution; S6: The synergy analysis module performs synergy analysis on combination schemes containing multiple candidate products and generates recommended combination product schemes. S7: The explainable recommendation output module presents the selection results to the user in the form of visual charts and natural language recommendation reasons.

9. The AI ​​product intelligent selection method according to claim 8, characterized in that, The multi-agent cooperation described in step S5 adopts the following process: 1) The coordinating and supervising agent receives tasks and decomposes and assigns them to the requirements analysis agent; 2) After the requirements analysis agent generates a structured requirements report, it triggers a candidate selection agent; 3) The candidate selection agent performs hard-constraint filtering to generate a candidate list; 4) Deeply evaluate the agent and concurrently perform evaluations across various dimensions; 5) Comparative analysis of the agent's generation of horizontal comparison results; 6) The solution recommendation agent integrates the outputs from all parties to generate a recommended solution; 7) Coordinate and supervise the quality of the intelligent agent's review and recommendation schemes.

10. The AI ​​product intelligent selection method according to claim 8, characterized in that, The multi-source product knowledge base is incrementally updated through scheduled tasks and event triggering mechanisms, automatically synchronizing product version updates, performance evaluation changes, and price changes.

Images