Intelligent guide algorithm system and intelligent search guide method for AI configuration tool

Abstract

The invention provides an intelligent guide algorithm system and an intelligent search guide method for an AI configuration tool. Comprising an intention recognition module used for receiving a demand text input by a user and performing multi-modal vectorization processing on the demand text; the intelligent question and answer module is used for generating function description information, operation example information and module mark information according to knowledge entries matched with the demand text when the function category label indicates that the platform has the function category; the AI auxiliary programming module retrieves function description information in an external information source based on the demand text when the function category label indicates that the platform does not have the function category; and the AI configuration judgment module is used for judging the legality of each connection line according to a pre-trained configuration rule model in the process of configuring and dragging the connection lines by the user, and outputting configuration error prompt information when the judgment result is illegal. According to the method, rapid and accurate retrieval, dynamic knowledge expansion, automatic code generation and configuration error real-time verification can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to an intelligent guide algorithm system and an intelligent search guide method for an AI configuration tool. BACKGROUND

[0002] An AI configuration tool (also referred to as an "AI low-code / no-code configuration platform") generally faces business developers or operation and maintenance engineers, and provides capabilities such as a visual component library, a drag-and-drop process arrangement, and a running environment management, so as to reduce the development and deployment threshold of an artificial intelligence model. In order to reduce the learning cost and guide the user to complete component selection, parameter configuration, and connection operation, a "wizard" function module is generally matched on the platform side, so as to provide auxiliary services such as function retrieval, step prompting, and error checking in the interactive interface.

[0003] The existing wizard scheme is generally divided into the following two categories:

[0004] The first category: non-intelligent wizard. It is presented in the form of a "user manual", an FAQ list, or a static web page. The user needs to manually query a keyword in a large amount of documents, and then compare the operation steps to complete the operation steps according to the actual scene, and the retrieval efficiency is low.

[0005] The second category: intelligent wizard. The operation process and the common problem set of the platform module are constructed based on a knowledge graph, the intelligent question-answering subsystem (text or voice interface) is used to perform semantic matching on the user's question and return the corresponding answer. This scheme can provide more intuitive operation guidance within the existing function range of the platform, but the reply of the intelligent question-answering system completely depends on the pre-recorded knowledge base entries.

[0006] However, the existing technical scheme still has the following problems: 1. The retrieval efficiency is limited: the non-intelligent wizard needs to manually review the documents, and it is difficult to quickly locate the step description that matches the actual demand. 2. The knowledge coverage is insufficient: the intelligent wizard can only answer the functions recorded in the knowledge base, and cannot give effective responses to the functions that are not yet included or recorded in the platform. 3. Lack of dynamic expansion capability: when the user's demand exceeds the existing knowledge base, the existing system cannot automatically expand the function description or generate executable code to meet the new function call. 4. The configuration verification is incomplete: the existing scheme focuses on static guidance, and the support for real-time legality verification during the user's drag-and-connect process is insufficient, resulting in the need for manual troubleshooting of the error connection. SUMMARY

[0007] Therefore, the embodiments of the present application provide an intelligent guide algorithm system and an intelligent search guide method for an AI configuration tool, so as to solve the problems of slow retrieval, knowledge loss, difficulty in dynamically expanding functions, and difficulty in real-time verification of configuration errors in the prior art.

[0008] The first aspect of the embodiment of the application provides an intelligent guide algorithm system for an AI configuration tool, comprising: an intention recognition module, configured to receive a requirement text input by a user, perform multi-modal vectorization processing on the requirement text, and generate a function category label based on a vector similarity matching result, the function category label indicating a platform existing function category or a platform non-existing function category corresponding to the requirement text; an intelligent question and answer module, in communication with the intention recognition module, configured to, when the function category label indicates the platform existing function category, call a knowledge base retrieval interface, generate function description information, operation example information and module marking information corresponding to internal components of the AI configuration tool according to a knowledge item matched by the requirement text, and output the function description information, the operation example information and the module marking information; an AI assisted programming module, in communication with the intention recognition module, configured to, when the function category label indicates the platform non-existing function category, retrieve function description information in an external information source based on the requirement text, generate a target programming language code file according to a running environment template of the AI configuration tool, and output the code file; and an AI configuration judgment module, connected with a configuration operation interface of the AI configuration tool, configured to, during a configuration drag and connection process performed by the user, determine the legality of each connection according to a pre-trained configuration rule model, and output a configuration error prompt information when the determination result is illegal.

[0009] The second aspect of the embodiment of the application provides an intelligent search guide method based on the intelligent guide algorithm system of the AI configuration tool, comprising: receiving a requirement text input by a user, sending the requirement text to an intention recognition module, performing multi-modal vectorization processing and outputting a function category label; when the function category label indicates a platform existing function category, performing semantic retrieval on the requirement text vector in an intelligent question and answer module, and sequentially generating function description information, operation example information and module marking information corresponding to internal components of the AI configuration tool according to the matched knowledge item; when the function category label indicates a platform non-existing function category, grabbing external information source data in an AI assisted programming module, generating a code file conforming to the syntax of a target programming language, and completing dependent library references, entry functions and running parameters; calling a running verification subunit in the AI assisted programming module to execute the code file in a sandbox running environment, and obtaining a compilation log and a running result; in an AI configuration judgment module, real-time monitoring configuration drag and connection operations performed by the user on an AI configuration tool interface, determining the legality of the connection based on a pre-trained configuration rule model, and generating a configuration error prompt information containing an error connection identifier and associated with the module marking information when the connection is illegal; and encapsulating the function description information, the operation example information, the module marking information, the code file, the compilation log, the running result and the configuration error prompt information as structured response data and sending the structured response data to an AI configuration tool front-end interface for component positioning and interactive display by an interface rendering module.

[0010] The at least one technical solution adopted by the embodiments of the present application can achieve the following beneficial effects:

[0011] By means of the intention recognition module, the demand text input by the user is received, the demand text is subjected to multi-modal vectorization processing, and a function category label is generated based on a vector similarity matching result, the function category label indicating a platform existing function category or a platform non-existing function category corresponding to the demand text; the intelligent question and answer module communicates with the intention recognition module, when the function category label indicates the platform existing function category, a knowledge base retrieval interface is called, function description information, operation example information and module marking information corresponding to internal components of the AI configuration tool are generated according to the knowledge item matched by the demand text, and the function description information, the operation example information and the module marking information are output; the AI assisted programming module communicates with the intention recognition module, when the function category label indicates the platform non-existing function category, function description information is retrieved in an external information source based on the demand text, a target programming language code file is generated according to a running environment template of the AI configuration tool, and the code file is output; the AI configuration judgment module is connected with a configuration operation interface of the AI configuration tool, during the configuration drag and connection process of the user, the legality of each connection is judged according to a pre-trained configuration rule model, and when the judgment result is illegal, a configuration error prompt information is output. The present application can realize fast and accurate retrieval, dynamic knowledge expansion, automatic code generation and real-time configuration error checking. BRIEF DESCRIPTION OF DRAWINGS

[0012] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0013] Figure 1 is a working process schematic diagram of the intelligent guide algorithm system for the AI configuration tool provided by the embodiments of the present application;

[0014] Figure 2 is a structural composition schematic diagram of the intelligent guide algorithm system for the AI configuration tool provided by the embodiments of the present application;

[0015] Figure 3 is a user text input box schematic diagram in an actual scene provided by the embodiments of the present application;

[0016] Figure 4 is a flow schematic diagram of the intelligent search guide method of the intelligent guide algorithm system based on the AI configuration tool provided by the embodiments of the present application. DETAILED DESCRIPTION

[0017] In the following description, for purposes of explanation and not limitation, specific details are set forth such as particular architectures, technologies, techniques, etc. in order to provide a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application can be practiced in other embodiments that depart from these specific details. In other instances, detailed descriptions of well-known systems, devices, circuits, and methods are omitted so as not to obscure the description of the present application with unnecessary detail.

[0018] Currently, the guide of AI configuration tools mainly adopts two modes: one is static user manual or FAQ list, and the user needs to manually search and compare the operation steps; the other is an intelligent question and answer system based on knowledge graph, which can provide semantic retrieval type guidance within the platform's recorded function range. However, the above solutions only cover known knowledge items and lack real-time verification of the legality of the connection in the drag-and-drop configuration process.

[0019] The existing guide solutions have the following shortcomings in terms of retrieval efficiency, knowledge coverage, dynamic function expansion, and real-time configuration verification: slow retrieval speed, difficulty in quickly locating the corresponding steps; unable to provide effective answers when the user's demand exceeds the knowledge base; unable to automatically generate executable code according to new requirements and verify it in the platform environment; configuration connection errors need to be manually checked and lack immediate prompts.

[0020] Therefore, the present application proposes an intelligent guide algorithm system for AI configuration tools, which constructs a four-level collaborative architecture of "intention recognition module-intelligent question and answer module-AI assisted programming module-AI configuration judgment module", wherein:

[0021] The intention recognition module performs multi-modal vectorization on user text and outputs "existing function / non-existing function" labels;

[0022] The intelligent question and answer module calls the vector database to retrieve knowledge items under the existing function branch, generates function descriptions, operation examples, and component labels;

[0023] The AI assisted programming module retrieves external data under the non-existing function branch, generates target language code with the help of a large language model, and completes dependency completion and sandbox verification combined with platform running templates;

[0024] The AI configuration judgment module real-time monitors the drag-and-drop connection operation, determines the legality of the connection based on a pre-trained rule model, and outputs prompt information containing error identification, and is associated with the aforementioned component labels to locate the problem node.

[0025] Through the above scheme, the present application realizes fast and accurate retrieval, dynamic knowledge expansion, automatic code generation, and real-time verification of configuration errors, thereby significantly improving the ease of use and functional completeness of AI configuration tools.

[0026] The working process and principle of the intelligent guide algorithm system of the present application will be described in detail below in combination with the accompanying drawings and embodiments, Figure 1 is a working flow diagram of the intelligent guide algorithm system for the AI configuration tool provided by the embodiments of the present application, as Figure 1 indicated, the working process of the system can specifically include the following contents:

[0027] The working process and principle of the intelligent guide algorithm system of the present application will be described step by step below in combination with the flow shown in Figure 1 . The flow starts with user input of a requirement text, and finally presents the multi-dimensional results retrieved or generated in the AI configuration tool front end in synchronization, and checks the legality of the connection in real time when the user actually configures the operation. The specific steps are as follows:

[0028] S1 User inputs requirement text: input natural language requirement in the text input box of the AI configuration tool, and the system calls the intent recognition module to receive the text.

[0029] S2 Intent recognition: the intent recognition module performs multi-modal vectorization on the requirement text, encodes the text segments and picture label content into embedded vectors, and compares the similarity with the existing corpus in the vector database, and outputs the function category label. If the similarity exceeds the threshold, it is determined that the platform has a function label, otherwise it is determined that the platform does not have a function label.

[0030] S3 Category shunt: the system selects different branches according to the function category label. If it is a platform function, go to S4; if it is a platform function, go to S6.

[0031] S4 Existing function retrieval: the intelligent question and answer module calls the semantic retrieval subunit to retrieve the most matched knowledge items in the vector database with the requirement text, and then generates operation example information according to the component dependency relationship in the items by the example generation subunit, and gathers function description information and extracts module marker information of the corresponding components.

[0032] S5 result formatting and output ①: the formatting subunit encapsulates the function description information, operation example information and module marker information according to the preset prompt template, forms structured response data and pushes it to the front end interface, and the front end highlights the related components and displays interactive step prompts accordingly.

[0033] S6 No function retrieval: the AI assisted programming module first retrieves external public documents and example codes by the networked retrieval subunit, and then generates target programming language code files by the code generation subunit calling large language model.

[0034] S7 Code Completion and Verification: The template mapping subunit completes the dependency library references, entry functions, and runtime parameters based on the runtime environment template, and then passes the code file to the runtime verification subunit; the runtime verification subunit compiles or interprets the code in the sandbox environment and collects compilation logs and runtime results.

[0035] S8 Result Formatting and Output ②: The AI-assisted programming module encapsulates function descriptions, code files, compilation logs, and running results into structured response data, which is also pushed to the front-end interface for users to view or directly access.

[0036] S9 Configuration Real-time Verification: During the process of dragging and dropping connections as guided by the user, the real-time monitoring subunit of the AI ​​configuration judgment module captures the connection information and inputs it into the pre-trained configuration rule model. When the model determines that the connection is invalid, the error prompt subunit generates a prompt containing the error connection identifier and the associated module marking information, which is displayed on the front-end interface in real time and located to the corresponding component to help the user correct it in time.

[0037] Through the closed-loop process from S1 to S9 described above, the system achieves continuous collaboration in requirements analysis, knowledge retrieval, automatic code generation, result verification, and configuration validation, ensuring that users can receive synchronous guidance and immediate feedback whether they are querying existing functions or expanding unknown functions.

[0038] The specific structure and function of the intelligent wizard algorithm system for AI configuration tools provided in this application will be described in detail below with reference to the accompanying drawings and specific embodiments. Figure 2 This is a schematic diagram of the structural composition of the intelligent wizard algorithm system for AI configuration tools provided in the embodiments of this application, as shown below. Figure 2 As shown, the intelligent wizard algorithm system for AI configuration tools may specifically include the following modules:

[0039] The intent recognition module 201 is used to receive the user's input request text, perform multimodal vectorization processing on the request text, and generate a function category label based on the vector similarity matching result. The function category label indicates whether the request text corresponds to a function category that the platform already has or does not have a function category.

[0040] The intelligent question answering module 202 communicates with the intent recognition module. When the function category label indicates that the platform already has a function category, it calls the knowledge base retrieval interface, generates function description information, operation example information, and module tag information corresponding to the internal components of the AI ​​configuration tool based on the knowledge entries matched with the requirement text, and outputs the function description information, operation example information, and module tag information.

[0041] The AI-assisted programming module 203 communicates with the intent recognition module, retrieves function description information in an external information source based on the demand text when the function category label indicates that the platform does not have a function category, and generates a target programming language code file according to a running environment template of the AI configuration tool and outputs the code file;

[0042] The AI configuration judgment module 204 is connected with a configuration operation interface of the AI configuration tool, and judges the legality of each connection according to a pre-trained configuration rule model during the configuration drag-and-connect process of the user, and outputs a configuration error prompt information when the judgment result is illegal.

[0043] In some embodiments, the intent recognition module comprises:

[0044] The text segmentation subunit is configured to split the demand text into a plurality of text segments according to a preset byte length before vectorization;

[0045] The picture preprocessing subunit is configured to perform noise removal and size normalization on the pictures in the platform document;

[0046] The vector encoding subunit is configured to call the same vectorization model to generate embedding vectors for the text segments and the preprocessed picture label content, respectively;

[0047] The similarity judgment subunit is configured to perform cosine similarity calculation on the user input vector and the embedding vector, and output the function category label according to the matching result that is not lower than a preset threshold.

[0048] Specifically, in the present embodiment, the intent recognition module (also referred to as "intent recognition intelligent agent" in the following embodiments of the present application) is deployed in the backend service cluster of the AI configuration tool, and is responsible for classifying the demand text as soon as possible after the user submits it, so as to determine the subsequent branch process. The module is composed of a text segmentation subunit, a picture preprocessing subunit, a vector encoding subunit and a similarity judgment subunit in sequence, and the overall working process is as follows.

[0049] First, after the text segmentation subunit receives the original demand text transmitted by the front end, the text is sequentially divided according to a preset byte length threshold (1000 bytes in the present embodiment) to avoid exceeding the model limit of the single vectorization input length. The natural paragraph boundary is preserved during the division, and if a single paragraph exceeds the threshold, it is further divided according to the semantic pause point. Finally, a set of text segment collections that do not exceed the threshold are output.

[0050] Subsequently, the picture preprocessing subunit performs offline processing on the jpg, jpeg, gif format pictures contained in the knowledge base document at the platform startup. The subunit first calls the denoising algorithm to remove scanning noise, and then performs size normalization according to the unified long side 256 pixel rule to generate standardized images; then the HSV chroma threshold method is used to eliminate the pure background area, reducing the subsequent OCR calculation amount. The preprocessed images are written into the intermediate cache and read by the subsequent steps.

[0051] The vector encoding subunit simultaneously processes two types of inputs at runtime: one is the text segment from the text segmentation subunit, and the other is the standardized image output by the picture preprocessing subunit. For text segments, the subunit calls the same preset vectorization model (such as a 768-dimensional sentence vector model based on the Transformer architecture) to generate corresponding text embedding vectors; for images, first, the text and table information are obtained through OCR recognition, and then a multi-modal semantic model is used to extract image semantic features. The two are spliced and input into the same vectorization model to obtain image label embedding vectors. All embedding vectors are stored in the vector database and marked with source identification for retrieval.

[0052] The similarity determination subunit performs vectorization simultaneously when receiving the user's original demand text, and maintains the same model and dimension configuration as the vector encoding subunit. Then the cosine similarity function is used to calculate the similarity between the user input vector and the text embedding vector and the image label embedding vector in the vector database, respectively. In this embodiment, the threshold is set to 0.5. When any similarity is greater than or equal to 0.5, it is determined that the user's demand belongs to the platform's existing function category; if all similarities are below the threshold, the platform does not have a function category label. The label is written into the message queue, triggering Figure 2 The intelligent question answering system branch or the AI-assisted programming intelligent agent branch shown.

[0053] It should be noted that, in order to ensure real-time performance, the vector encoding subunit completes the embedding operation of all knowledge base texts and images in the system initialization stage, and periodically listens to document addition and deletion events for incremental updates; the similarity determination subunit uses an asynchronous non-blocking method to write logs, avoiding becoming a bottleneck in high-concurrency scenarios. Through the above specific implementation, each subunit forms a flow pipeline in the data interface layer, with upstream output being downstream input, ensuring the rapid decision-making ability of the intent recognition module in the entire intelligent guide process.

[0054] In some examples, the building process of the intent recognition module (i.e., the intent recognition intelligent agent) of the present application can include the following steps:

[0055] Step 1: Organize the materials of the platform usage method and example program, and save the format of the materials as txt, pdf, picture (jpg, jpeg, gif, etc.), and html static web page;

[0056] Step 2: According to the length limit of the vectorization model (such as 1000), divide the files of the platform usage method and example program (except pictures) into a limit size (such as 1000);

[0057] Step 3: Select a vectorization model, and vectorize the above files, marked as A;

[0058] Step 4: For picture (jpg, jpeg, gif, etc.) format files, perform picture preprocessing;

[0059] Step 5: Select a multi-modal model and an OCR model, use the OCR model to recognize the text and tables in the picture, and then use the multi-modal model to perform semantic recognition of the picture, and comprehensively analyze the semantic recognition results and the OCR recognition results of the same picture to obtain the label content of the picture;

[0060] Step 6: Select a vectorization model, and vectorize the label content of the picture, marked as B;

[0061] Step 7: Select a large language base model, mark the contents of A and B as “platform existing function”; collect materials from external network, marked as “platform does not have function”;

[0062] Step 8: Train the selected large language base model to obtain model D;

[0063] Step 9: Input the text input by the user into the model D, and analyze the platform existing function and the platform not having function respectively.

[0064] In some embodiments, a picture preprocessing subunit and a vector encoding subunit are provided between:

[0065] An OCR recognition subunit is configured to recognize characters and tables in the picture and generate a text result;

[0066] A semantic fusion subunit is configured to perform feature fusion on the text result and the picture semantic vector based on a multi-modal model to obtain picture label content.

[0067] Specifically, in this embodiment, after the picture preprocessing subunit completes denoising and size normalization, it will submit the standardized image together with a unique identifier to the OCR recognition subunit located in the same process pool. The functions and implementation methods of the OCR recognition subunit and the semantic fusion subunit are described in detail below.

[0068] Firstly, the OCR recognition subunit sequentially performs three-stage processing of text detection, character recognition, and structured parsing. The text detection stage uses an end-to-end detection network based on a convolutional feature pyramid to generate a number of candidate text boxes on a 256x256 pixel input, and outputs a coordinate set after non-maximum suppression. The character recognition stage calls a bidirectional LSTM+CTC decoding model for each text box to directly output a UTF-8 encoded string. When the detection box contains obvious table line segments, the subunit further triggers the table structured parsing process, applies an improved segmentation-merging algorithm to locate cells at row-column intersections, and writes the recognized characters into a two-dimensional array according to row-column indexes. Finally, the OCR recognition subunit generates a JSON text result containing "position-content" key-value pairs and returns it to the upstream caller.

[0069] Subsequently, the semantic fusion subunit reads the JSON text result and the visual semantic vector of the corresponding image, and performs cross-modal feature fusion. In the specific implementation process, the subunit first uses the Tokenizer consistent with the text vectorization to split the character sequence in the JSON into a Token sequence, and outputs a 512-dimensional text vector through the Transformer text encoder; at the same time, the visual semantic vector is extracted in advance by the multi-modal model in the image encoding branch, and the dimension is also 512. The semantic fusion subunit constructs a bidirectional cross-attention layer on the two types of vectors: the first path weight is calculated with the text vector as Query and the visual vector as Key and Value, the second path weight is calculated with the visual vector as Query and the text vector as Key and Value, the two path results are element-level weighted and summed, then concatenated, and then reduced to a unified 768-dimensional vector through a fully connected layer. To ensure vector space consistency, the semantic fusion subunit performs L2 normalization processing at the output end, and writes the obtained picture label content together with the image identifier into the internal cache for subsequent retrieval and comparison by the vector encoding subunit.

[0070] Through the above OCR recognition and semantic fusion process, the system not only accurately extracts the text and table information in the picture, but also generates more discriminative multi-modal label content combined with the visual context, significantly improving the matching accuracy of subsequent similarity determination.

[0071] In some embodiments, the intelligent question and answer module includes:

[0072] A knowledge base construction subunit for writing the text embedding vector and the picture embedding vector processed by the vector encoding subunit into a vector database;

[0073] A semantic retrieval subunit for retrieving the closest knowledge items in the vector database based on the demand text vector;

[0074] An example generation subunit is configured to automatically generate corresponding operation example information according to the component dependency relationship in the knowledge item.

[0075] Specifically, in the present embodiment, the intelligent question and answer module is deployed on the retrieval service node of the AI configuration tool backend, and carries out three-stage flow of the knowledge base construction subunit, the semantic retrieval subunit and the example generation subunit. The working process thereof forms a closed loop with the aforementioned intent recognition module, the OCR recognition subunit and the semantic fusion subunit, and ensures that the user can obtain structured guidance information when querying the existing functions of the platform. The specific implementation is as follows.

[0076] Firstly, the knowledge base construction subunit receives the text embedding vectors and the picture embedding vectors output by the vector encoding subunit in batches in the system initialization stage. The subunit generates a three-element index field for each vector: document identifier, paragraph serial number and vector type, and completes the database writing through the batch writing interface of the vector database. In order to balance the retrieval efficiency and incremental update capability, the knowledge base construction subunit adopts a partitioned inverted strategy: the main partition is divided according to the functional domain, and the sub-partition index is maintained according to the update timestamp; when the platform document is added or deleted, only the affected sub-partition is rewritten to avoid full reconstruction.

[0077] Secondly, after receiving the "existing function" trigger signal of the intent recognition module, the semantic retrieval subunit first performs consistent vectorization operation on the user demand text to obtain the demand text vector. Then, the K nearest neighbor algorithm is called in the vector database to retrieve the k most similar knowledge items according to the cosine distance. In the present embodiment, k is set to 5, and a dynamic threshold is introduced to filter out low-confidence items: when the similarity difference between the kth item and the first item exceeds 0.2, the number of returned items is automatically reduced. The semantic retrieval subunit also analyzes the item metadata to extract the component identifier list and the dependency relationship description, providing input for the example generation subunit.

[0078] Finally, the example generation subunit looks up the interface specification, parameter template and visualization type in the platform component metadata table according to the retrieved component identifier list, and constructs a component dependency directed graph. The subunit arranges the nodes in topological order, generates three-stage operation examples in combination with default parameters and common connection paradigms: (1) component addition explanation, which clearly indicates the dragging position of each node in the AI configuration tool canvas; (2) connection step explanation, which lists the connection direction and constraint conditions; (3) parameter configuration explanation, which lists the key parameters and suggested values in the order of components. The example generation subunit encapsulates the function explanation information, operation example information and module marker information at the output end, and pushes the display request to the front end, which accordingly superimposes a highlighted outline and a bubble prompt on the corresponding component to help the user quickly complete the configuration setup. Through the cooperation of the above three subunits, the intelligent question and answer module realizes efficient retrieval of knowledge items and automatic generation of operable steps, significantly improving the self-learning efficiency of the existing functions of the platform.

[0079] In some examples, the building process of the intelligent question-answering module of the present application can include the following steps:

[0080] Step 1: Organize the materials of the platform usage method and example program, and save the format of the materials as txt format, pdf format, picture (jpg, jpeg, gif, etc.) format, and html static web page format;

[0081] Step 2: According to the size limit (such as 1000) of the vectorization model length, divide the files (except pictures) of the platform usage method and example program into a size limit (such as 1000) or less;

[0082] Step 3: Select a vectorization model, and vectorize the above files, marked as A;

[0083] Step 4: For files in picture (jpg, jpeg, gif, etc.) format, perform picture preprocessing;

[0084] Step 5: Select a multi-modal model and an OCR model, use the OCR model to recognize the text and tables in the picture, and then use the multi-modal model to perform semantic recognition of the picture, and comprehensively analyze the semantic recognition results and the OCR recognition results of the same picture to obtain the label content of the picture;

[0085] Step 6: Select a vectorization model, and vectorize the label content of the picture, marked as B;

[0086] Step 7: Select an intelligent question-answering model (such as deepseek, tongyiqianwen model, etc.), and construct a knowledge base of the intelligent question-answering model using the vectorized content of the files and the vectorized content of the picture;

[0087] Step 8: Vectorize the text content input by the user, marked as C;

[0088] Step 9: Compare C with A and C with B respectively, and design a threshold size (such as 0.5) as the threshold for accepting the comparison results; when the comparison results of C with A and C with B exceed 0.5, the knowledge is taken as part of the return result, marked as D;

[0089] Step 10: Adjust D to the format of the prompt, and return it to the user;

[0090] Step 11: According to the content returned to the user, form a marked display for the AI configuration tools that may be used.

[0091] In some embodiments, the intelligent question-answering module further includes:

[0092] The formatting subunit is configured to combine the function description information, the operation example information and the module mark information into structured response data according to a preset prompt specification.

[0093] Specifically, in the present embodiment, the formatting subunit is arranged at the end of the intelligent question and answer module process, and is mainly responsible for mapping the function description information, the operation example information and the module mark information output by the upstream example generation subunit into unified structured response data according to a preset prompt specification, and pushing the unified structured response data to the front end. The implementation steps of the subunit are as follows.

[0094] Firstly, the formatting subunit is preset with a set of prompt templates for front-end analysis of the AI configuration tool. The template is composed of three sequential placeholders: ① function description segment placeholder <ft>, for inserting function description information; ② operation example segment placeholder <st>, for inserting operation example information generated in topological order; ③ component marker segment placeholder <mt>, which is used to insert module mark information. When the system starts, the formatting sub-unit loads the template into memory and builds a placeholder index table in memory to facilitate quick replacement.

[0095] Secondly, after receiving the three types of information from the example generation sub-unit, the formatting sub-unit first performs paragraph justification on the function description information: removes redundant symbols, merges consecutive blank lines, and deletes automatic page breaks between Chinese punctuation and English characters to ensure the consistency of the layout when output is presented on multiple terminals. Then, the formatting sub-unit performs numbering and normalization on the operation example information: automatically generates "StepX" labels in the order of component dependency topology, and adds component ID annotations to each step to enable bidirectional jumping to the corresponding component of the canvas when rendered by the front end.

[0096] Thirdly, the formatting sub-unit appends UI rendering parameters to the module mark information. According to the component ID, coordinates, and component type in the module mark information, the formatting sub-unit adds the fields of highlight style name, display duration, and prompt offset. For example, when the component type is "data input module", the highlight style is set to "info-blue", the display duration is set to 4s, and the prompt offset is set to top offset 20px. All style parameters come from the front-end style configuration table, ensuring cross-version compatibility.

[0097] Subsequently, the formatting sub-unit calls the placeholder replacement engine to replace the <ft> 、 <st> 、 <mt>The placeholder generates an intermediate text string that conforms to the prompt specification. To facilitate decoupling of the backend and the frontend, the text string is further encapsulated as a key-value structure, where the "prompt" field stores the complete text, the "componentMap" field stores a component ID and UI parameter mapping table, and the "linkFlag" field is used to indicate whether an anchor relationship needs to be established between the step label and the component highlight in the operation example.

[0098] Finally, the formatting subunit writes the structured response data to the frontend message queue through the message push interface. After the frontend receives the data, it first displays the card-style text content according to the "prompt" field, then reads the "componentMap" field to add a highlight border to the corresponding component and display a bubble prompt. If the user clicks the "StepX" label in the operation example, the canvas immediately scrolls and focuses on the corresponding component, achieving synchronous linkage between the text guide and the configuration canvas.

[0099] Through this formatting process, the embodiment ensures that the function description, operation example, and component label are tightly coupled at the structural level, while still maintaining flexible configuration at the rendering level, thereby meeting the visualization needs of AI configuration tools across platforms and resolutions, and further improving the interactive friendliness and ease of learning of the intelligent guide algorithm system.

[0100] In some embodiments, the AI-assisted programming module includes:

[0101] The networking retrieval subunit is configured to retrieve open-source documents and example code related to the demand text from external information sources;

[0102] The code generation subunit is configured to call a large language model to generate a code file that conforms to the syntax of the target programming language based on the open-source documents and example code;

[0103] The template mapping subunit is configured to complete the dependent library reference, entry function, and running parameters according to the running environment template of the AI configuration tool.

[0104] Specifically, in the present embodiment, the AI-assisted programming module (also referred to as "AI-assisted programming agent" in the following embodiments of the present application) is automatically called after the intent recognition module outputs the "platform has no function" label, and its internal structure is composed of a networking retrieval subunit, a code generation subunit, and a template mapping subunit in a flow pipeline order, and the specific implementation process is as follows.

[0105] Firstly, the network retrieval subunit receives the requirement text and the target programming language identifier. The subunit constructs a composite search formula containing the three elements of "function keyword + language keyword + open source", accesses external code hosting sites and technical document portals through the REST interface, and prioritizes the open source repositories under the OSI license and official sample documents in the returned results. Subsequently, the network retrieval subunit extracts core passages with a matching degree of function keywords not less than 0.7 using a fast summary algorithm, and removes sensitive information and platform-independent hardware instructions through regular rules; Finally, the knowledge package for the code generation subunit is generated, which contains function signature, core algorithm idea, dependent library list and usage example fragment.

[0106] Subsequently, the code generation subunit loads the large language model fine-tuning weights corresponding to the target programming language, assembles the requirement text, knowledge package and platform API index file into multiple context prompts. The subunit generates multiple candidate code fragments through temperature sampling control, and performs syntax checking and security scanning in the local static analysis engine, retaining only the warning-free candidates. For scenarios with optional algorithm paths, the code generation subunit establishes priorities based on execution efficiency and external dependencies, selects the highest priority code fragment as the main output, and marks the algorithm version and model generation timestamp in the metadata field.

[0107] Finally, after receiving the output of the code generation subunit, the template mapping subunit first reads the running environment template of the AI configuration tool, which defines the dependent library reference placeholder, entry function placeholder and running parameter placeholder in advance. The template mapping subunit automatically parses the third-party library list according to the import statement in the code fragment, and completes version alignment and image tag supplementation by comparing with the platform image configuration table; If the entry function is missing, a uniform entry function is inserted according to the language specification and the user's main logic is encapsulated as an independent method; For running parameters, the template mapping subunit maps them to command line arguments or configuration file key-value pairs according to the value range in the requirement text and the platform environment variable rules. After revision, the template mapping subunit generates code files and dependent description files that conform to the platform directory structure, and writes file paths, dependency summaries and generation algorithm fingerprints into the message queue for subsequent running verification subunit or front-end request scheduling.

[0108] Through the above linkage, the AI-assisted programming module realizes a full-automatic closed loop from external knowledge retrieval to platformized code delivery, ensuring that users can quickly obtain high-consistency code files that can be directly run even in the absence of target functions on the platform.

[0109] In some examples, the building process of the AI-assisted programming module (i.e., AI-assisted programming intelligent agent) of the present application can include the following steps:

[0110] Step 1: Organize the running environment of existing programming languages on the platform, such as Python, Java, C++, etc.

[0111] Step 2: Organize various code program segment contents and function description tags for Python, Java, and C++ respectively.

[0112] Step 3: Select a large language base model, train the function problem, code segment content, and function description content, and obtain the AI-assisted programming agent model E.

[0113] Step 4: When the user inputs the text content, call the AI-assisted programming agent model E, and output the corresponding code and function description.

[0114] In some embodiments, the AI-assisted programming module further includes:

[0115] The running verification sub-unit is used to execute the code file in the sandbox running environment and return the compilation log and running result to the front-end interface after execution is completed.

[0116] Specifically, in this embodiment, the running verification sub-unit is located in the terminal process of the AI-assisted programming module, and is used to perform sandbox-level automated compilation or interpretation running on the code file generated by the template mapping sub-unit. The complete implementation steps are as follows.

[0117] First, the running verification sub-unit dynamically selects a sandbox image according to the target language identifier of the code file. All images are pre-installed with runtime libraries consistent with the platform environment template and enable seccomp and AppArmor dual-layer policies to limit file system and network access range. After calling the container orchestration service, the sub-unit starts an independent container instance with resource quotas of CPU≤1 core, memory≤512MiB, and time limit≤30s, and injects the code file and dependency description file into the container working directory through read-only mounting.

[0118] Second, the running verification sub-unit distinguishes between two types of execution processes within the container: if the target language is C++ or Java, it first calls the corresponding compiler to perform full compilation and redirects the compilation stdout and stderr to the log buffer; if it is Python or other script languages, it skips the compilation phase and directly executes the main entry script, while enabling a timeout daemon thread to monitor the execution time. After the script or binary is started, the sub-unit captures the standard output, standard error, and process exit code, and transmits them to the host shared memory in real time through the built-in log driver of the container.

[0119] Subsequently, the running verification sub-unit closes the network namespace and destroys the container immediately after the container exits, reads the log entries and exit code in the shared memory, and forms a result object containing three fields of "compileLog", "runLog", and "exitCode". If the exitCode is non-zero, an "errorCategory" field is appended and mapped to an enumerated value of "syntax error", "missing dependency", or "runtime exception" according to the std err content for front-end highlighting.

[0120] Finally, the running verification sub-unit sends the result object to the front-end interface through the message push channel. After receiving the object, the front-end renders the logs according to the "compileLog" and "runLog" columns, and displays the execution time and exitCode at the edge of the canvas; if there is an errorCategory, a red warning bubble is popped up next to the corresponding module marker. Through the above process, the running verification sub-unit ensures that the code file is executed safely in a controlled environment and provides immediate feedback of the compilation and running results to the user, realizing the closed loop of the AI-assisted programming module.

[0121] In some embodiments, the AI configuration judgment module comprises:

[0122] A rule training sub-unit configured to construct training samples based on the input data type, the output data type, and the connection restriction, and train a binary classification judgment model;

[0123] A real-time monitoring sub-unit configured to collect connection information and input the binary classification judgment model when a user drag-and-connect event occurs;

[0124] An error prompt sub-unit configured to generate prompt information containing an error connection identifier and send it to the front-end interface when it is determined that the connection is illegal;

[0125] The prompt information generated by the error prompt sub-unit contains component identifiers corresponding to the module marker information output by the intelligent question-answering module or the AI-assisted programming module, so as to automatically locate the related components on the front-end interface.

[0126] Specifically, in the present embodiment, the AI configuration judgment module (also referred to as "AI configuration judgment agent" in the following embodiments of the present application) is deployed in the real-time verification service process of the AI configuration tool, and is used to make instant judgments and feedback on the connection legality during the whole process of user visual drag-and-connect. The module is composed of a rule training sub-unit, a real-time monitoring sub-unit, and an error prompt sub-unit, and its implementation is as follows.

[0127] Firstly, the rule training subunit extracts features from historical project samples during the platform offline training phase. The platform database records the input data type, output data type, connection restriction label, and component pair information for each legal and illegal connection. The subunit concatenates the feature vector according to the "input type + output type + restriction label" and marks the legal connection as "1" and the illegal connection as "0" to construct a binary classification training sample. Then, the random forest algorithm is called for model training, and the best tree depth and tree number parameters are selected through five-fold cross-validation to finally solidify the binary classification judgment model file. This model file is deployed with the service image and triggers an incremental training process after each rule library update to ensure that the verification rules remain synchronized with the latest component changes.

[0128] Secondly, the real-time monitoring subunit runs in the WebSocket channel listener between the front end and the back end. When the user triggers the drag-and-drop connection event on the AI configuration tool canvas, the front end immediately sends the connection start component ID, end component ID, start output type, and end input type information packaged as a JSON object to the listener. The real-time monitoring subunit parses the object to generate a feature vector in the same format as the rule training phase and calls the binary classification judgment model loaded in memory for inference. To ensure smooth interaction, the real-time monitoring subunit uses a thread pool locking method to handle concurrent connection requests, with a single inference delay controlled within 5ms.

[0129] Finally, the error prompt subunit starts the prompt generation process when it receives the binary classification judgment model output result as "illegal". The subunit first creates an error prompt entity containing the error connection identification, connection start and end node IDs, error type enumeration, and timestamp fields, then queries the latest module label information mapping table output by the intelligent question and answer module or AI assisted programming module to locate the coordinates of the start or end component on the canvas. If the mapping table contains the corresponding component ID, the error prompt subunit appends the coordinates and highlight color code fields to the returned message to ensure that the front end can automatically focus and flash the component. Subsequently, the error prompt subunit pushes the prompt entity to the front end through WebSocket, and the front end draws a red cross icon in the middle of the connection and a flashing red box around the corresponding component, while adding a "illegal connection" text entry to the side panel log panel.

[0130] Through the above rule training, real-time monitoring, and error prompt processes, the AI configuration judgment module achieves millisecond-level detection of the legality of user drag-and-drop connections and associates the judgment results with the component label information output by the intelligent question and answer module or AI assisted programming module, enabling the error prompt to accurately locate the relevant component position on the canvas, thereby helping users to instantly discover and correct connection errors during the configuration process.

[0131] In some examples, the building process of the AI configuration judgment module (i.e., the AI configuration judgment intelligent agent) of the present application can include the following steps:

[0132] Step 1: Organize the configuration rules of the platform configuration tool, including input data types, output data types, selection items of connected modules, restriction items, and commonly used items as training data items;

[0133] Step 2: Select a classification model (such as SVM model, decision tree, random forest, etc.), and train it into a binary classification judgment model that can judge true or false;

[0134] Step 3: When the user configures, automatically start the binary classification judgment model, and directly display a prompt error for the configuration connection that does not meet the configuration rules.

[0135] In some embodiments, the outputs of the intelligent question answering module and the AI assisted programming module are configured to be displayed through the front-end interface of the AI configuration tool, and the judgment result of the AI configuration judgment module is associated with the module marker information output by the intelligent question answering module or the AI assisted programming module, so as to locate the relevant components on the front-end interface.

[0136] Specifically, in the present embodiment, the front-end interface of the AI configuration tool adopts a browser-side single-page application architecture, which is composed of three parts of "canvas rendering layer, panel display layer, and message subscription bus". The canvas rendering layer is responsible for drawing component nodes, connections, and highlight overlays in the Canvas2D way; the panel display layer renders the wizard text and log on the right side of the interface in the DOM node way; the message subscription bus realizes event distribution based on the native EventTarget in the browser, and maintains a long connection with the backend through WebSocket. When the front-end starts, it first loads the component metadata, style sheet, and animation key frame definition, and then establishes a WebSocket connection with the backend intelligent wizard service, generates a sessionId corresponding to a single session for subsequent message verification.

[0137] I. Front-end rendering process of the intelligent question answering module and the AI assisted programming module

[0138] 1. When the backend intelligent question answering module completes the encapsulation of the structured response data, it pushes a guideMsg event in JSON format to the WebSocket channel. The payload field of the event contains prompt text, componentMap mapping table, linkFlag indication bit, timeStamp, and msgId. componentMap is a key-value set, with the key being the component ID and the value containing the x, y, w, h coordinates of the component in the canvas coordinate system, the highlight style name styleName, and the duration duration.

[0139] 2. The message bus dispatches guideMsg to two listeners, "PromptRenderer" and "ComponentHighlighter", according to the event type. PromptRenderer first converts prompt to HTML fragment by invoking built-in Markdown escape module, and then dynamically decides whether to insert a fold button according to the content length. If prompt contains heading tags (#, ##), a collapsible table of contents is generated in the side panel.

[0140] 3. ComponentHighlighter reads componentMap and retrieves corresponding node references in the canvas object registry for each component ID. If the retrieval is successful, a semi-transparent overlay instance is created, positioned at the outer edge of the node rectangle by 5px with absolute coordinates, applies the corresponding CSS animation keyframes (e.g. glow-blue or glow-green) according to styleName, and registers a timer to destroy the overlay after duration expires. If linkFlag is true, PromptRenderer adds data-step attributes and writes component IDs for each StepX tag in the guide text, and then uniformly registers click events. When the user clicks StepX, the canvas.scrollToCoordinate(x, y, 300ms) function is called to smoothly scroll the canvas in a quadratic easing curve within 300 milliseconds and trigger the node's own pulse animation.

[0141] The running result message of the AI-assisted programming module is similar in format to guideMsg, but the prompt field in the payload is accompanied by code block explanations, and the front-end rendering process is basically the same. The difference is that ComponentHighlighter generates anchors for file names or function names in example code, so that when the user clicks on a file name, a running log dialog box pops up below the panel.

[0142] II. Error prompt linkage of AI configuration judgment module

[0143] 1. During the user's drag-and-drop connection process, the front-end sends lineMsg containing startId, endId, outType, inType, and lineId to the backend real-time monitoring subunit every time a connectionAttempt event occurs. If the backend binary classification judgment model returns legal, the front-end only draws the default connection path; if it returns illegal, the backend pushes errorMsg event to the front-end.

[0144] 2. The payload field of the errorMsg includes errorType, lineId, startId, endId, componentStyle, errorCategory, timeStamp. The message bus dispatch this event to "LineErrorRenderer" and "ComponentHighlighter". LineErrorRenderer looks up the path node in canvas connection management table according to lineId, overlays a red dashed line of fixed width and draws a red cross at the middle point of the line. The opacity of the cross is animated between 1.0 and 0.2 by CSS to create a flickering effect.

[0145] 3. ComponentHighlighter looks up the canvas component according to startId or endId, creates a red box overlay and sets styleName to error-red, which keeps flickering until the user fixes the connection. To avoid multiple error hints overlapping, Highlighter queries the global overlay stack before creating a new one. If the component is already highlighted, it only updates the flashCount and extends the duration, no need to create a new instance.

[0146] 4. If the user deletes the error connection or reconnects successfully, the frontend receives a resolve Msg from the backend. LineErrorRenderer removes the red line and cross immediately, and notifies ComponentHighlighter to decrease the flashCount of the corresponding overlay. When the flashCount reaches zero and the duration ends, the overlay is automatically destroyed. Meanwhile, the error entry is marked as "resolved" and the font is grayed out in the side panel.

[0147] III. Data Consistency and Fault Tolerance

[0148] To ensure the sequential consistency of guideMsg and errorMsg in a high concurrency environment, the front end punches a time stamp for each event in the subscription bus and performs highlighting and de-highlighting in ascending order of time stamp within the same component ID range to avoid old tips from covering new tips. If the component ID does not exist in componentMap or the canvas node reference is null, the Highlighter degenerates into a "panel positioning mode", that is, an orange warning banner is inserted at the top of the side panel to prompt the user that the current component may not be loaded or hidden. For message loss caused by network jitter, the WebSocket connection enables a heartbeat packet, which sends a ping every 15 seconds. If 3 non-replies are detected, the connection is automatically re-hands and the unconfirmed msgId queue is re-sent.

[0149] Through the above detailed message distribution, DOM rendering, canvas highlighting and error resolution processes, the embodiment realizes deep coupling display of intelligent question and answer information, AI-assisted programming results and real-time configuration verification feedback in the same front-end interface, enabling the user to jump to the related component with one key when reading the guide text, immediately locate and correct the connection error, and thus complete the closed-loop interaction of guide tips and configuration verification.

[0150] The application discloses an intelligent guide algorithm system applied to an AI configuration tool, intelligently identifies and prompts the AI configuration tool module based on an artificial intelligence algorithm, helps users quickly and conveniently learn the use of the AI configuration tool, and also enables the user to conveniently locate the possible error steps; for the functions that have never existed on the platform, an AI-assisted programming intelligent agent is used to form corresponding code and function description according to the text input content of the user, further expand the system functions, and conveniently enable the user to use; in the implementation of the intelligent question and answer system, the method of picture preprocessing, OCR identification and multi-modal model is innovatively used for the picture (jpg, jpeg, gif, etc.) format file, and the accuracy of picture semantic analysis is improved.

[0151] The above embodiment describes the specific module structure and functions of the intelligent guide algorithm system applied to the AI configuration tool, and the implementation process of the intelligent search guide method of the application will be described in detail in combination with specific embodiments. Figure 3 is a schematic diagram of a user text input box in an actual scene provided by an embodiment of the application, Figure 4 is a flowchart of an intelligent search guide method of an intelligent guide algorithm system based on an AI configuration tool provided by an embodiment of the application, as shown in Figures 3 to 4 The intelligent search guide method can specifically include the following steps:

[0152] S401, receiving a demand text input by a user, sending the demand text to an intent recognition module, performing multi-modal vectorization processing and outputting a function category label;

[0153] S402, when the function category label indicates that the platform has a function category, performing semantic retrieval on the demand text vector in an intelligent question and answer module, generating function description information, operation example information, and module marker information corresponding to internal components of an AI configuration tool in sequence according to matched knowledge items;

[0154] S403, when the function category label indicates that the platform does not have a function category, grabbing external information source data in an AI assisted programming module, generating a code file conforming to the syntax of a target programming language, and supplementing dependent library references, entry functions, and running parameters;

[0155] S404, calling a running verification subunit in the AI assisted programming module to execute the code file in a sandbox running environment, and obtaining a compilation log and a running result;

[0156] S405, real-time monitoring configuration drag and connection operations performed by the user on an AI configuration tool interface in the AI configuration judgment module, determining the legality of the connection based on a pre-trained configuration rule model, and generating configuration error prompt information containing error connection identification and associated with module marker information when the connection is illegal;

[0157] S406, encapsulating the function description information, the operation example information, the module marker information, the code file, the compilation log, the running result, and the configuration error prompt information as structured response data and sending them to an AI configuration tool front-end interface for component positioning and interactive display by an interface rendering module.

[0158] Specifically, in the embodiment, the intelligent search guide method relies on an intelligent guide algorithm system that has been constructed, and completes the whole link closed loop of demand analysis, knowledge retrieval, code generation, running verification, and configuration verification around the intelligent search guide method as shown in the flowchart. Figure 4 The following describes key processing details in combination with steps S401 to S406.

[0159] S401 User text access and intent recognition

[0160] When the user submits a natural language requirement in the input box of the AI configuration tool, the system immediately records the sessionld and sends the original text to the intent recognition module. The module first calls the text segmentation subunit to cut the text according to the 1000 byte threshold, and then uses the vectorization model to generate embedding vectors for each segment; at the same time, the picture preprocessing, OCR recognition and semantic fusion subunit generates corresponding label embedding for the knowledge base picture. The similarity determination subunit retrieves the most similar entry in the vector database, and if the highest similarity is ≥0.5, the "existing function" label is output, otherwise the "non-existing function" label is output, and the requirement vector is written to the message bus.

[0161] S402 Knowledge retrieval and example generation of existing function path

[0162] When the function category label is "existing function", the semantic retrieval subunit of the intelligent question and answer module performs cosine K nearest neighbor retrieval according to the requirement text vector, and obtains the most relevant knowledge entry set. The component dependency information between entries is sent to the example generation subunit, which constructs a component directed graph, generates step-by-step operation examples according to topological sorting, and extracts component IDs. Finally, the function description information, operation example information and module label information are written to the structured response body by the formatting subunit, where the module label information binds the coordinate and highlight style parameters for each component ID for direct rendering by the front end.

[0163] S403 Code generation and template mapping of non-existing function path

[0164] When the function category label is "non-existing function", the AI-assisted programming module starts the network retrieval subunit, based on "function keyword + language keyword" combination to search open source code repositories and official documents, extract function signatures and algorithm core paragraphs to form knowledge packages. The code generation subunit assembles the requirement text, knowledge package and platform API index into a prompt, calls the language model to generate multiple candidate code segments, and filters out the code with correct syntax and least dependencies through static analysis. The template mapping subunit reads the running environment template, automatically completes the import statement, entry function and parameter parsing logic for the code, and outputs the code file and dependency description file under the standard project directory.

[0165] S404 Sandbox running verification and result collection

[0166] After template mapping, the verification subunit is run to select a matching seccomp sandbox image based on the language identifier, compile or interpret the execution code in a container with a limit of 1 CPU core, 512 MiB of memory, and a time limit of 30 seconds. During execution, stdout and stderr are returned to the host cache in real time through the log driver. After the container ends, the subunit collects the compileLog, runLog, and exitCode, and if the exitCode is not zero, it analyzes the error type and generates the errorCategory field. Finally, all logs and running results are packaged into the response body.

[0167] S405 Real-time verification and error positioning

[0168] When the user drags the component connection according to the wizard prompt, the front end encapsulates each connectionAttempt event as lineMsg and sends it to the real-time listening subunit. The latter extracts the input and output types and calls the binary classification judgment model to determine whether it is legal. When it is determined to be illegal, the error prompt subunit creates errorMsg, which contains lineId, startId, endId, and the highlight style inherited from the module marker information, to ensure that the front end can provide synchronous flashing prompts and canvas positioning for the error connection and related components.

[0169] S406 Unified packaging and front-end rendering

[0170] The system unifies the function description information, operation example information, module marker information, code file path, compilation log, running result, and possible configuration error prompts into a JSON response, including prompt, componentMap, codePath, compileLog, runLog, errorMsg, and timeStamp. The front end receives the response and renders the prompt text and running log in the side panel. It also highlights the components based on componentMap, and if errorMsg exists, it draws a cross on the corresponding connection and a red box around the component. The user can click the operation step label or error prompt to trigger the canvas to automatically smooth scroll and focus on the component, realizing the bidirectional linkage of the text wizard and visual configuration.

[0171] Through the above six steps, the embodiment realizes a complete closed loop of requirement analysis-knowledge retrieval-code generation-running verification-configuration verification-visual feedback, enabling the AI configuration tool to provide users with intelligent search wizard services that cover both existing platform functions and new functions in a single interface.

[0172] It should be understood that the size of the serial number of each step in the above method embodiment does not mean the order of execution, and the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present disclosure.

[0173] The above embodiments are only used to illustrate the technical solutions of the present application, but not limit them; although the technical solutions of the present application are described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can still be modified, or some technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.< / mt> < / st> < / ft> < / mt> < / st> < / ft>

Claims

1. An intelligent wizard algorithm system for AI configuration tools, characterized in that, include: The intent recognition module is used to receive user-inputted demand text, perform multimodal vectorization processing on the demand text, and generate functional category labels based on vector similarity matching results. The functional category labels indicate whether the platform has existing functional categories or not. The intelligent question answering module communicates with the intent recognition module. When the function category label indicates that the platform already has a function category, it calls the knowledge base retrieval interface, generates function description information, operation example information, and module tag information corresponding to the internal components of the AI ​​configuration tool based on the knowledge entries matched by the requirement text, and outputs the function description information, operation example information, and module tag information. The AI-assisted programming module communicates with the intent recognition module. When the function category label indicates that the platform does not have a function category, it retrieves function description information from external information sources based on the requirement text, generates a target programming language code file according to the running environment template of the AI ​​configuration tool, and outputs the code file. The AI ​​configuration judgment module connects to the configuration operation interface of the AI ​​configuration tool. During the user's configuration drag-and-drop connection process, it judges the legality of each connection according to the pre-trained configuration rule model, and outputs configuration error prompt information when the judgment result is illegal.

2. The system according to claim 1, characterized in that, The intent recognition module includes: The text segmentation subunit is used to split the required text into several text fragments according to a preset byte length before vectorization; The image preprocessing subunit is used to perform noise removal and size normalization on images in the platform documents; The vector encoding subunit is used to call the same vectorization model to generate embedding vectors for the text fragment and the preprocessed image tag content, respectively. The similarity determination subunit is used to calculate the cosine similarity between the user input vector and the embedded vector, and output the function category label based on the matching result that is not lower than a preset threshold.

3. The system according to claim 2, characterized in that, The image preprocessing subunit and the vector encoding subunit are configured with the following: The OCR recognition subunit is used to perform character and table recognition on the image and generate text results; The semantic fusion subunit is used to fuse the text results with the image semantic vector based on a multimodal model to obtain the image label content.

4. The system according to claim 1, characterized in that, The intelligent question-answering module includes: The knowledge base construction subunit is used to write the text embedding vector and image embedding vector processed by the vector encoding subunit into the vector database; The semantic retrieval subunit is used to retrieve the most similar knowledge entries from the vector database based on the required text vector. Example generation subunit is used to automatically generate corresponding operation example information based on the component dependencies in the knowledge entry.

5. The system according to claim 4, characterized in that, The intelligent question-answering module also includes: The formatting subunit is used to combine the functional description information, the operation example information, and the module tagging information into structured response data according to the preset prompt specification.

6. The system according to claim 1, characterized in that, The AI-assisted programming module includes: The network retrieval subunit is used to retrieve open-source documents and sample code related to the required text from external information sources; The code generation subunit is used to call the large language model to generate code files that conform to the syntax of the target programming language based on the open-source documentation and sample code. The template mapping subunit is used to complete the dependency library references, entry functions, and runtime parameters based on the runtime environment template of the AI ​​configuration tool.

7. The system according to claim 6, characterized in that, The AI-assisted programming module also includes: The execution verification subunit is used to execute the code file in the sandbox environment and return the compilation log and execution results to the front-end interface after execution.

8. The system according to claim 1, characterized in that, The AI ​​configuration judgment module includes: The rule training subunit is used to construct training samples based on the input data type, output data type, and connection constraints, and to train a binary classification model. The real-time monitoring subunit is used to collect connection information and input it into the binary classification judgment model when a user drags and connects a line. The error message subunit is used to generate a message containing an error connection identifier and send it to the front-end interface when the connection is determined to be invalid. The error message generated by the error message subunit includes a component identifier corresponding to the module tagging information output by the intelligent question-and-answer module or the AI-assisted programming module, so as to automatically locate the relevant component on the front-end interface.

9. The system according to claim 1, characterized in that, The outputs of the intelligent question-answering module and the AI-assisted programming module are both configured to be displayed through the front-end interface of the AI ​​configuration tool, and the judgment result of the AI ​​configuration judgment module is associated with the module tag information output by the intelligent question-answering module or the AI-assisted programming module, so as to locate the relevant components on the front-end interface.

10. An intelligent search guide method based on an intelligent guide algorithm system for AI configuration tools as described in any one of claims 1 to 9, characterized in that, include: Receive user input request text, send the request text to the intent recognition module, perform multimodal vectorization processing, and output function category labels; When the function category label indicates that the platform already has a function category, semantic retrieval is performed on the requirement text vector in the intelligent question answering module, and function description information, operation example information, and module tag information corresponding to the internal components of the AI ​​configuration tool are generated in sequence according to the matched knowledge entries. When the function category label indicates that the platform does not have a function category, the AI-assisted programming module captures external information source data, generates code files that conform to the syntax of the target programming language, and completes the dependency library references, entry functions and running parameters. The AI-assisted programming module calls the execution verification subunit to execute the code file in the sandbox runtime environment and obtains the compilation log and execution results; The AI ​​configuration judgment module monitors the user's configuration drag-and-drop connection operations on the AI ​​configuration tool interface in real time, determines the legality of the connection based on the pre-trained configuration rule model, and generates a configuration error prompt message containing an error connection identifier and associated with the module's marking information when the connection is invalid. The functional description information, operation example information, module tag information, code files, compilation logs, running results, and configuration error prompts are encapsulated into structured response data and sent to the front-end interface of the AI ​​configuration tool so that the interface rendering module can locate components and display interactions.

Images