Unified service system and method for large model-based agricultural business

Abstract

The application discloses an agricultural business unified service system and method based on a large model, relates to the technical field of large model unified service, accepts a natural language query of a user, generates to-be-processed query information, configures service parameters, generates model calling data based on the to-be-processed query information and the service parameters, receives the model calling data, calls a corresponding inference model based on the model calling data to perform inference, generates a structured inference result, matches a callback interface based on the structured inference result, outputs callback data based on the callback interface that is successfully matched, receives the callback data, performs page operation and data processing based on the callback data, and outputs a business presentation result. Through construction of a prompt word association field and a double-state prompt word encapsulation body, the application realizes on-demand activation of associated prompt words, and effectively improves the response efficiency of multi-round business handling in an agricultural scene.

Description

Technical Field

[0001] This invention relates to the field of unified service technology for large-scale models, and specifically to a unified service system and method for agricultural business based on large-scale models. Background Technology

[0002] In the existing technology, agricultural intelligent question answering systems are usually developed and deployed independently for a single agricultural business type, such as building agricultural technology consultation systems, pest and disease identification systems or agricultural product price query systems, with each system independently maintaining its own prompt word templates and model calling logic; However, existing technologies have certain shortcomings. First, agricultural business types are diverse, and existing technologies lack unified service registration and scheduling for different businesses. Each business system is developed and deployed independently. When new business types need to be added or existing business configurations need to be adjusted, multiple systems need to be modified separately, resulting in difficulties in function reuse, high maintenance costs, and difficulty in achieving intelligent services in multiple business scenarios. At the same time, each model call requires the independent construction of prompt words, and there is no correlation mechanism between prompt words. When users conduct multiple rounds of continuous business processing, it causes a decrease in reasoning efficiency. If the complexity of prompt words is increased, the reasoning complexity will increase, causing delays in model reasoning response and affecting user experience and business processing efficiency. Therefore, developing a unified agricultural business service system based on a large model is of great significance. Summary of the Invention

[0003] The purpose of this invention is to provide a unified service system and method for agricultural operations based on a large model, in order to solve the problems in the prior art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a unified agricultural business service system based on a large model, comprising: User input module: Used to receive natural language queries from users and generate query information to be processed; Service registration module: Connected to the user input module, it is used to configure service parameters, including service type, prompt word template, callback interface identifier, and model call parameters. Based on the query information to be processed and the service parameters, it generates model call data. LLM module: Connected to the service registration module, it receives model call data, calls the corresponding inference model based on the model call data to perform inference, and generates structured inference results; Callback matching module: Connects to the LLM module, matches callback interfaces based on structured inference results, and outputs callback data based on successfully matched callback interfaces; Inference Display Module: Connected to the callback matching module, it receives callback data, performs page operations and data processing based on the callback data, and outputs business presentation results.

[0005] In a preferred embodiment, the user input module includes: An input receiving unit is used to receive natural language queries input by the user; The preprocessing unit is used to perform structured processing on natural language queries, remove redundant information, complete the context, and generate query information to be processed.

[0006] In a preferred embodiment, the service registration module includes: The service definition unit is used to define service types and label the corresponding service identifiers. Service types include three-capital services, transaction services, and agricultural services. The prompt word management unit is used to store system prompt words and user prompt word templates. System prompt words are used to define LLM roles, capability boundaries, and output format constraints, while user prompt word templates contain variable placeholders. The business matching unit matches service identifiers based on the query information to be processed and generates corresponding service business to be processed based on the matching results. Based on the pending service business, a search is performed in the prompt word management unit to obtain the corresponding matching system prompt words and user prompt word templates; Fill the variable placeholders in the user prompt word template with the query information to be processed to generate the corresponding user prompt word. Combine the system prompt word and the corresponding user prompt word to generate the initial prompt word for the sonic logging pipe. Map the initial prompt words to the prompt word association field to generate prompt word cells; The interface configuration unit is used to store the callback interface identifier and model call parameters corresponding to each service identifier, and generate interface configuration information based on the corresponding callback interface identifier and model call parameters of the service to be processed. The prompt word cell body and interface configuration information are encapsulated to generate model call data.

[0007] In a preferred embodiment, the step of mapping the initial prompt word to the prompt word association field and generating the prompt word cell body is as follows: The initial prompt word is input into the prompt word association field, which contains multiple prompt word fragments and the business semantic association relationships between each prompt word fragment; Based on business semantic association, retrieve prompt word fragments that have business relevance to the initial prompt word, encapsulate the retrieved prompt word fragments as latent prompt word clusters, and label each prompt word fragment with trigger condition tags; Construct a cue word cell body space, which has a two-layer storage structure including a recessive kernel and a dominant outer kernel, and the storage space of the dominant outer kernel is larger than that of the recessive kernel; Establish condition-triggered interfaces between the implicit kernel and the explicit outer kernel, and establish normally-pass interfaces on the explicit outer kernel; The latent prompt word cluster is stored in the latent kernel, and the initial prompt word is stored in the explicit outer kernel; The stored implicit kernel, explicit outer kernel, conditional trigger interface, and normally accessible interface are encapsulated into a prompt word cell.

[0008] In a preferred embodiment, the LLM module includes: The model adaptation unit is used to access general-purpose large language models and configure model call interfaces corresponding to service types; The parameter parsing unit is used to receive model call data, match the corresponding model call interface based on the model call data, connect with the service registration module, and activate the corresponding large language model. After determining the large language model, the large language model extracts model call parameters from the model call data. The model call parameters include output length limit parameters and response timeout parameters. The interface connection unit corresponds to the common interface of the large language model to identify the prompt word cell body, and extracts the initial prompt word based on the common interface, performs inference based on the model call parameters, and generates inference results; Once the reasoning is complete, the user continues to input updated query information, and the corresponding large model receives the updated query information through the model call interface. Based on the activation of the condition trigger interface for updating the query information to be processed, the updated query information to be processed is transmitted to the hidden kernel through the constant-pass interface and the condition trigger interface. The trigger condition labels of the latent cue word clusters are matched, and the successfully matched cue word fragments are fed back to the large language model to continue the update inference and generate the update inference results. The reasoning results and the updated reasoning results are output as structured reasoning results.

[0009] In a preferred embodiment, the callback matching module includes: The callback configuration storage unit is used to store the callback interface configuration corresponding to each service identifier. The callback interface configuration includes the callback address, request method, data format and authentication information. The callback matching unit is used to receive the structured inference result and match the corresponding callback interface configuration from the callback configuration storage unit according to the service identifier corresponding to the structured inference result. The data encapsulation unit encapsulates the structured inference results into callback data based on the data format in the matched callback interface configuration. The callback data includes the action type and action parameters. The callback push unit is used to push callback data to the inference display module based on the callback address and authentication information.

[0010] In a preferred embodiment, the inference display module includes: The callback data receiving unit is used to receive callback data and parse the action type and action parameters in the callback data. The execution unit performs corresponding page operations and data processing based on the action type and action parameters, and outputs the page presentation results.

[0011] This invention also provides a unified agricultural business service method based on a large model, including: Receive natural language queries from users and generate query information to be processed; Configure service parameters, including service type, prompt word template, callback interface identifier, and model call parameters. Generate model call data based on the query information to be processed and the service parameters. Receive model call data, call the corresponding inference model based on the model call data to perform inference, and generate structured inference results; Based on the structured reasoning results, the callback interfaces are matched, and the callback data is output based on the successfully matched callback interfaces. Receive callback data, perform page operations and data processing based on the callback data, and generate business presentation results.

[0012] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention manages service parameters for various agricultural business types uniformly through a service registration module, decoupling business logic from model calls. The service registration module pre-configures service parameters for various agricultural business types, such as asset management, transaction services, and agricultural services. A business matching unit automatically identifies the business type of a user query, extracts the corresponding prompt word template identifier, callback interface identifier, and model call parameters, and generates standardized model call data, which is then sent to the LLM module. When a new business type needs to be added, only the corresponding configuration needs to be modified in the service registration module; the processing logic of the LLM module and callback matching module does not need to be modified. Based on the unified encapsulation and centralized scheduling of business definitions, prompt word management, interface configuration, and model call parameters, this invention effectively reduces the coupling between system layers, improves system scalability and maintenance efficiency, and enhances the intelligence level of agricultural digital services to a certain extent. 2. This invention achieves dynamic association and on-demand exposure of prompt words in agricultural business scenarios by constructing a prompt word association field and a bi-state prompt word encapsulation. The invention maps initial prompt words to the prompt word association field, automatically retrieves related prompt word fragments using pre-defined business semantic relationships, and generates a bi-state prompt word encapsulation containing a implicit kernel and an explicit outer kernel. By default, the LLM module only obtains the initial prompt words from the explicit outer kernel through the common interface, maintaining the simplicity and efficiency of reasoning. When the user's subsequent input query information matches the trigger condition label of a prompt word fragment in the implicit kernel, the corresponding associated prompt word is retrieved from the implicit kernel and combined with the initial prompt word, providing enhanced prompt words for the LLM module. Based on this mechanism, intelligent reuse and on-demand activation of business context in multiple interactions are achieved, effectively improving the coherence and accuracy of multi-round question answering in complex agricultural business scenarios. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0014] Figure 1 This is a system flowchart of the present invention.

[0015] Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0017] Example 1, please refer to Figure 1 As shown in this embodiment, the unified agricultural business service system based on a large model includes: User input module: Used to receive natural language queries from users and generate query information to be processed; Service registration module: Connected to the user input module, it is used to configure service parameters, including service type, prompt word template, callback interface identifier, and model call parameters. Based on the query information to be processed and the service parameters, it generates model call data. LLM module: Connected to the service registration module, it receives model call data, calls the corresponding inference model based on the model call data to perform inference, and generates structured inference results; Callback matching module: Connects to the LLM module, matches callback interfaces based on structured inference results, and outputs callback data based on successfully matched callback interfaces; Inference Display Module: Connected to the callback matching module, it receives callback data, performs page operations and data processing based on the callback data, and outputs business presentation results; Furthermore, in existing technologies, agricultural intelligent question-answering systems are usually developed and deployed independently for a single type of agricultural business. For example, separate agricultural technology consultation systems, pest and disease identification systems, or agricultural product price query systems are built, and each system independently maintains its own prompt word templates and model calling logic. However, existing technologies have certain shortcomings. First, agricultural business types are diverse, and existing technologies lack unified service registration and scheduling for different businesses. Each business system is developed and deployed independently. When new business types need to be added or existing business configurations need to be adjusted, multiple systems need to be modified separately, resulting in difficulties in function reuse, high maintenance costs, and difficulty in achieving intelligent services in multiple business scenarios. At the same time, each model call requires the independent construction of prompt words, and there is no correlation mechanism between prompt words. When users conduct multiple rounds of continuous business processing, it causes a decrease in reasoning efficiency. If the complexity of prompt words is increased, the reasoning complexity will increase, causing delays in model reasoning response and affecting user experience and business processing efficiency. This invention manages service parameters for various agricultural business types uniformly through a service registration module, decoupling business logic from model calls. The service registration module pre-configures service parameters for different agricultural business types, such as asset management, transaction services, and agricultural services. A business matching unit automatically identifies the business type of a user query, extracts the corresponding prompt word template identifier, callback interface identifier, and model call parameters, and generates standardized model call data, which is then sent to the LLM module. When a new business type needs to be added, only the corresponding configuration needs to be modified in the service registration module; the processing logic of the LLM module and callback matching module does not need to be changed. Based on the unified encapsulation and centralized scheduling of business definitions, prompt word management, interface configuration, and model call parameters, this invention effectively reduces the coupling between system layers, improves system scalability and maintenance efficiency, and enhances the intelligence level of agricultural digital services to a certain extent. By constructing a prompt word association field and a bi-state prompt word encapsulation, dynamic association and on-demand exposure of prompt words in agricultural business scenarios are realized. This invention maps initial prompt words to the prompt word association field, automatically retrieves related prompt word fragments using pre-set business semantic association relationships, and generates a bi-state prompt word encapsulation containing a implicit kernel and an explicit outer kernel. By default, the LLM module only obtains the initial prompt words in the explicit outer kernel through the common interface, keeping the reasoning simple and efficient. When the user's subsequent input query information matches the trigger condition label of a prompt word fragment in the implicit kernel, the corresponding related prompt words are retrieved from the implicit kernel and combined with the initial prompt words to provide enhanced prompt words for the LLM module. Based on this mechanism, intelligent reuse and on-demand activation of business context in multiple interactions are realized, effectively improving the coherence and accuracy of multi-round question answering in complex agricultural business scenarios.

[0018] In one embodiment, the user input module includes: An input receiving unit is used to receive natural language queries input by the user; The preprocessing unit is used to perform structured processing on natural language queries, remove redundant information, complete the context, and generate query information to be processed. Furthermore, the input receiving unit receives natural language queries input by the user, providing both text and voice input interfaces. The text input interface collects the text content typed by the user through an input box component, while the voice input interface acquires the user's voice signal and converts it into text content by calling the voice acquisition device of the terminal device. The preprocessing unit performs structured processing on the natural language queries, removing redundant information and completing the context to generate query information to be processed. It filters the original query content using a preset agricultural stop word list and regular expression template, removing modal particles, repetitive expressions, and non-business-related additional content, while retaining core semantic elements related to agricultural business. Additionally, it maintains the user session state. The system records the user identity information and business scenario information confirmed in the current session. When a newly received natural language query contains referential or ellipsis expressions, the preprocessing unit extracts the corresponding context information from the user session state record to complete the query content. For example, it replaces referential expressions with confirmed specific entity names and expands ellipsis expressions into complete query statements. Furthermore, the preprocessing unit introduces an automatic time information completion mechanism. When a natural language query involves time-related expressions but lacks a specific time range, it automatically obtains the current system time and completes and standardizes the time expressions according to the time cycle characteristics of agricultural business, generating semantically complete and formatted query information to be processed.

[0019] In one embodiment, the service registration module includes: The service definition unit is used to define service types and label the corresponding service identifiers. Service types include three-capital services, transaction services, and agricultural services. The prompt word management unit is used to store system prompt words and user prompt word templates. System prompt words are used to define LLM roles, capability boundaries, and output format constraints, while user prompt word templates contain variable placeholders. The business matching unit matches service identifiers based on the query information to be processed and generates corresponding service business to be processed based on the matching results. Based on the pending service business, a search is performed in the prompt word management unit to obtain the corresponding matching system prompt words and user prompt word templates; Fill the variable placeholders in the user prompt word template with the query information to be processed to generate the corresponding user prompt word. Combine the system prompt word and the corresponding user prompt word to generate the initial prompt word for the sonic logging pipe. Map the initial prompt words to the prompt word association field to generate prompt word cells; The interface configuration unit is used to store the callback interface identifier and model call parameters corresponding to each service identifier, and generate interface configuration information based on the corresponding callback interface identifier and model call parameters of the service to be processed. The prompt word cell body and interface configuration information are encapsulated to generate model call data; Furthermore, the service definition unit predefines all agricultural business types supported by the system through configuration files or database tables. Each business type is assigned a unique service identifier. Specifically, the service identifiers for the three-capital services correspond to four sub-businesses: fund management, asset management, resource management, and supervision and auditing. The service identifiers for transaction services correspond to four sub-businesses: transaction security verification, transaction contract services, price analysis, and transaction matching. The service identifiers for agricultural services correspond to four sub-businesses: agricultural input procurement, agricultural technology services, planting guidance, and pest and disease control. The mapping relationship between each service identifier and its corresponding sub-business is stored in the service definition table. The prompt word management unit uses a key-value pair storage structure to maintain the system prompt word library and user prompt word templates respectively. The system prompt word library stores each system prompt word using a service identifier as the index key. The content of each system prompt word defines the role the LLM should play in the corresponding business type, its required capabilities, and the structured format requirements for its output. Similarly, each user prompt word template in the user prompt word template library is stored using a service identifier as the index key. Each user prompt word template contains variable placeholders marked with specific tags. These placeholders are filled based on the user's input query information, transforming the user input into a unified data structure. The business matching unit performs word segmentation on the query information to be processed and extracts keyword features. The extracted keyword features are then compared with the service identifiers in the service definition unit. Semantic similarity is calculated for the corresponding sub-business description text. This calculation uses a pre-trained word vector model based on the agricultural domain. After mapping keyword features and sub-business description text to the same vector space, cosine similarity is calculated. The service identifier corresponding to the sub-business with the highest similarity is selected as the matching result. Based on the matching result, a service business object to be processed is generated, containing the service identifier and the corresponding sub-business type information. After matching, the service identifier in the service business object to be processed is used as the search key to perform precise key-value searches in the system prompt word library and user prompt word template library of the prompt word management unit, respectively, to obtain the corresponding system prompt words and user prompt word templates. If the search fails, it will revert to the default general suggestion word template. After obtaining the suggestion word template, the query information to be processed will be filled and replaced according to the position and type of the variable placeholders in the user suggestion word template. The variable placeholders include query content placeholders, time placeholders and entity name placeholders. After filling, the user suggestion word corresponding to the current query will be generated. The system suggestion word and the user suggestion word will be concatenated and combined according to the preset concatenation order. When concatenating, the system suggestion word is placed in the first part as the behavior constraint prefix of LLM, and the user suggestion word is placed in the second part as the specific query input. The initial suggestion word will be generated and mapped to the suggestion word association field. The suggestion word cell will be generated based on the suggestion word association.The interface configuration unit stores the callback interface identifiers and model call parameters corresponding to each service identifier in the interface configuration table. Each record in the interface configuration table uses the service identifier as the primary key and includes a callback interface identifier field and a model call parameter field. The model call parameters specifically include output length limit parameters and response timeout parameters. Using the service identifier in the business object to be processed as the query condition, an exact match query is performed in the interface configuration table to obtain the corresponding callback interface identifier and model call parameters. The callback interface identifier and model call parameters are then encapsulated into interface configuration information. The prompt word cell body and the interface configuration information are encapsulated into model call data and output to the LLM module to activate the corresponding large language model.

[0020] In one embodiment, the step of mapping the initial prompt word to the prompt word association field and generating the prompt word cell body is as follows: The initial prompt word is input into the prompt word association field, which contains multiple prompt word fragments and the business semantic association relationships between each prompt word fragment; Based on business semantic association, retrieve prompt word fragments that have business relevance to the initial prompt word, encapsulate the retrieved prompt word fragments as latent prompt word clusters, and label each prompt word fragment with trigger condition tags; Construct a cue word cell body space, which has a two-layer storage structure including a recessive kernel and a dominant outer kernel, and the storage space of the dominant outer kernel is larger than that of the recessive kernel; Establish condition-triggered interfaces between the implicit kernel and the explicit outer kernel, and establish normally-pass interfaces on the explicit outer kernel; The latent prompt word cluster is stored in the latent kernel, and the initial prompt word is stored in the explicit outer kernel; The stored implicit kernel, explicit outer kernel, conditional trigger interface, and normally accessible interface are encapsulated into prompt word cells; Furthermore, the prompt word association field is implemented using a graph database. Nodes in the graph database correspond to prompt word fragments, and each node's attributes include a unique identifier for the prompt word fragment, its text content, and the business type label it belongs to. Edges in the graph database correspond to the business semantic relationships between prompt word fragments, and the edge attributes include the association type and association strength value. Association types include pre-dependency, complementarity, and extended description relationships. The association strength value is calculated based on the co-occurrence frequency and the degree of business logical dependence between corresponding business entities in the agricultural business knowledge graph. The text content of the initial prompt words is mapped to a semantic vector of a preset dimension, and then the cosine similarity of the semantic vectors is used across all nodes in the graph database. Nearest neighbor search is performed, and the node with the highest similarity exceeding a preset anchoring threshold is identified as the anchor node. If no node exceeding the preset anchoring threshold exists in the search results, the node with the highest similarity is selected as the anchor node. After determining the anchor node, a breadth-first search is performed along the edges in the graph database, starting from the anchor node. The maximum number of hops for the expansion search is limited to two hops. During the expansion process, the association strength value of each edge is judged, and only the nodes connected to edges with association strength values ​​greater than the preset association strength threshold are included in the candidate association node set. The nodes in the candidate association node set are sorted in descending order according to their association strength values, and a preset number of nodes at the top of the sorting are selected as the final set of associated prompt word nodes. The text content of the prompt word fragments from each node in the prompt word node set is extracted and encapsulated into a latent prompt word cluster. The latent prompt word cluster is stored in an ordered list structure. Each element in the list contains the text content of the prompt word fragment and its corresponding trigger condition label. The trigger condition label is based on the business entity type and intent type associated with the prompt word fragment in the agricultural business knowledge graph. Specifically, the corresponding entity type and intent type are extracted from the business type label attribute of the prompt word fragment node as the content of the trigger condition label. After encapsulating the latent prompt word cluster, the system constructs a prompt word cell space in memory. The prompt word cell space adopts a two-layer storage structure, including a latent kernel storage... The system consists of a primary kernel storage area and a secondary kernel storage area. The pre-allocated memory space of the primary kernel storage area is set to an integer multiple of the pre-allocated memory space of the secondary kernel storage area to ensure that the primary kernel has enough space to accommodate the complete initial prompt word. Furthermore, the storage address of the secondary kernel is physically contained within the primary kernel; that is, an address block is allocated within the primary kernel's storage space as the storage address of the secondary kernel. A condition-triggered interface is constructed between the secondary kernel storage area and the primary kernel storage area. This interface is an event-driven matching controller. The matching controller has an internal read handle pointing to the secondary kernel storage area and a write handle pointing to the primary kernel storage area, and its initial state is sleep mode.A constant-access interface is built on the explicit outer kernel storage area. This constant-access interface is implemented as a lightweight access interface, exposing only the initial prompt word content stored in the explicit outer kernel storage area to external callers. Based on the condition-triggered interface, the content in the implicit kernel storage area is hidden from external callers. The implicit prompt word cluster is serialized and written to the implicit kernel storage area, maintaining the order of each prompt word fragment and the correspondence between the trigger condition tags during writing. The initial prompt word is written to the explicit outer kernel storage area, along with its metadata information, including the service identifier and generation timestamp. After writing data to both the implicit and explicit outer kernel storage areas, the memory address references of the implicit and explicit outer kernel storage areas, the matching controller reference of the condition-triggered interface, and the access interface reference of the constant-access interface are encapsulated into an independent prompt word cell object. This prompt word cell object exposes the constant-access interface for the LLM module to call during default inference to obtain the initial prompt word.

[0021] In one embodiment, the LLM module includes: The model adaptation unit is used to access general-purpose large language models and configure model call interfaces corresponding to service types; The parameter parsing unit is used to receive model call data, match the corresponding model call interface based on the model call data, connect with the service registration module, and activate the corresponding large language model. After determining the large language model, the large language model extracts model call parameters from the model call data. The model call parameters include output length limit parameters and response timeout parameters. The interface connection unit corresponds to the common interface of the large language model to identify the prompt word cell body, and extracts the initial prompt word based on the common interface, performs inference based on the model call parameters, and generates inference results; Once the reasoning is complete, the user continues to input updated query information, and the corresponding large model receives the updated query information through the model call interface. Based on the activation of the condition trigger interface for updating the query information to be processed, the updated query information to be processed is transmitted to the hidden kernel through the constant-pass interface and the condition trigger interface. The trigger condition labels of the latent cue word clusters are matched, and the successfully matched cue word fragments are fed back to the large language model to continue the update inference and generate the update inference results. The reasoning results and the updated reasoning results are output as structured reasoning results; Furthermore, the model adaptation unit maintains a model adaptation registry, which records the list of general-purpose large language models accessed by the system and the corresponding call interface configurations for these models. The call interface configuration includes the interface address, authentication credentials, and communication protocol type. During system initialization, the model adaptation registry is read, and connections to each general-purpose large language model are established and their validity verified. Simultaneously, the model adaptation unit constructs a mapping relationship between service identifiers and model types, configuring a corresponding model call interface handle for each service identifier and generating a call interface handle mapping table. The parameter parsing unit receives model call data, extracts the callback interface identifier and model call parameters from the interface configuration fields, and searches for the corresponding model call interface handle in the interface handle mapping table based on the callback interface identifier. If the search is successful, the corresponding large language model instance is activated using this handle. The activation operation includes initiating a session initialization request to the large language model server and obtaining the session identifier. If the search fails, it reverts to the default general-purpose large language model. After determining the large language model, the system reads the model... The model call data extracts output length limit parameters and response timeout parameters. The output length limit parameter constrains the maximum number of characters or tokens output by the large language model in a single operation, while the response timeout parameter sets the maximum waiting time for this inference call. The interface connection unit parses the prompt word cell body object reference from the prompt word cell body and obtains the access handles of the normally accessible interface and the conditional trigger interface through the prompt word cell body object. During the first inference, the interface connection unit accesses the initial prompt word in the explicit outer kernel storage area through the normally accessible interface and passes the initial prompt word as the prompt word to be inferred to the inference execution unit. The inference execution unit assembles the prompt word to be inferred and the model call parameters into an inference request. The inference request includes a prompt word text field, an output length limit field, and a response timeout field. The inference execution unit sends the inference request to the large language model server through the model call interface handle corresponding to the activated large language model. The large language model server performs inference calculations based on the prompt word text and returns the inference result. In a single-round inference scenario, this inference result is the final output of the LLM module.In multi-turn inference scenarios, when the user continues to input updated query information, the updated query information does not need to be regenerated by the service registration module to generate model call data. Instead, it is directly passed to the interface connection unit of the LLM module. The interface connection unit obtains the reference of the bound prompt word cell object from the session cache, uses the updated query information as the input for trigger condition matching, and passes the updated query information to the implicit kernel storage area through the condition trigger interface. The matching controller of the condition trigger interface switches from a dormant state to an active state, and performs matching calculations on the business features of the updated query information with each trigger condition label. The matching calculation uses a combination of string inclusion matching and semantic similarity matching. If the updated query information is updated... If the query information to be processed contains the entity type or intent feature marked by the trigger condition label, a successful match is determined. The condition trigger interface reads the successfully matched prompt word fragment from the implicit kernel storage area and concatenates it with the initial prompt word. During concatenation, the initial prompt word is placed at the beginning as the basic context, and the prompt word fragment is placed at the end as the supplementary context, generating an enhanced prompt word as the updated inference prompt word. The interface connection unit assembles the updated inference prompt word and the model call parameters corresponding to the current session into an update inference request and sends it to the large language model server. The large language model server performs inference based on the enhanced prompt word and returns the updated inference result, outputting the inference result or the updated inference result as the final structured inference result.

[0022] In one embodiment, the callback matching module includes: The callback configuration storage unit is used to store the callback interface configuration corresponding to each service identifier. The callback interface configuration includes the callback address, request method, data format and authentication information. The callback matching unit is used to receive the structured inference result and match the corresponding callback interface configuration from the callback configuration storage unit according to the service identifier corresponding to the structured inference result. The data encapsulation unit encapsulates the structured inference results into callback data based on the data format in the matched callback interface configuration. The callback data includes the action type and action parameters. The callback push unit is used to push callback data to the inference display module based on the callback address and authentication information; Furthermore, the callback configuration storage unit uses a relational database table to store the callback interface configurations corresponding to each service identifier. Each record in the callback configuration table uses the callback interface identifier as the primary key and includes a callback address field, a request method field, a data format field, and an authentication information field. The request method field stores the request method type of the Hypertext Transfer Protocol, the data format field stores the serialization format type of the callback data, and the authentication information field stores the credential token used to verify the legitimacy of the callback request source. The callback matching unit receives the structured inference results, extracts the service identifier field from the structured inference results, and performs a matching query in the callback configuration table of the callback configuration storage unit using the service identifier as the query condition. If a match is successful, the callback address, request method, data format, and authentication information in that record are obtained as the callback interface configuration. If no record is matched, the preset default callback interface configuration is used. The data encapsulation unit receives the callback interface configurations and the structured inference results. The system extracts the intent field from the structured inference results as the action type, and the answer field and confidence field as the main content of the action parameters. Based on the serialization format specified in the data format field of the callback interface configuration, the action type and action parameters are encapsulated to generate callback data conforming to the specified data format. The callback push unit receives the callback data, callback address, and authentication information. It constructs a Hypertext Transfer Protocol request with the callback address as the target endpoint, injects the value of the authentication information field into the request header as authentication credentials, and the request body payload is the encapsulated callback data. The callback push unit sends this request to the inference display module asynchronously and non-blockingly. Furthermore, if the push fails, it automatically retryes according to an exponential backoff retry strategy, with a maximum of three retries. The waiting time interval for each retry increases sequentially. If the push still fails after three retries, the callback failure event is recorded in the system log and an alarm notification is triggered.

[0023] In one embodiment, the inference display module includes: The callback data receiving unit is used to receive callback data and parse the action type and action parameters in the callback data. The execution unit performs corresponding page operations and data processing based on the action type and action parameters, and generates the page presentation result; Furthermore, the callback data receiving unit starts an asynchronous data listener bound to the network port corresponding to the callback address, continuously listening for callback data pushed by the callback push unit. Upon receiving the callback data, it parses the data, extracting the values ​​of the action type field and the action parameter field. The action type field indicates the type of operation to be performed, including page navigation, data refresh, pop-up prompts, and table filling. The action parameter field stores the specific parameters required to perform the operation in the form of a key-value pair set. The presentation execution unit matches the corresponding execution logic according to the action type. For example, when navigating to a page, it extracts the target page route address from the action parameters and calls the page routing component to complete the navigation. When refreshing data, it extracts the data area identifier and update content from the action parameters and calls the interface rendering component to complete the partial update. When displaying a pop-up prompt, it extracts the prompt title and prompt content from the action parameters and calls the pop-up component to complete the display. After execution, the operation results are integrated into the page presentation results and output to the user interface for display.

[0024] Example 2, please refer to Figure 2 As shown in this embodiment, the unified agricultural business service method based on a large model includes: Receive natural language queries from users and generate query information to be processed; Configure service parameters, including service type, prompt word template, callback interface identifier, and model call parameters. Generate model call data based on the query information to be processed and the service parameters. Receive model call data, call the corresponding inference model based on the model call data to perform inference, and generate structured inference results; Based on the structured reasoning results, the callback interfaces are matched, and the callback data is output based on the successfully matched callback interfaces. Receive callback data, perform page operations and data processing based on the callback data, and output the business presentation results.

[0025] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A unified agricultural business service system based on a large model, characterized in that: include: User input module: Used to receive natural language queries from users and generate query information to be processed; Service registration module: Connected to the user input module, it is used to configure service parameters, including service type, prompt word template, callback interface identifier, and model call parameters. Based on the query information to be processed and the service parameters, it generates model call data. LLM module: Connected to the service registration module, it receives model call data, calls the corresponding inference model based on the model call data to perform inference, and generates structured inference results; Callback matching module: Connects to the LLM module, matches callback interfaces based on structured inference results, and outputs callback data based on successfully matched callback interfaces; Inference Display Module: Connected to the callback matching module, it receives callback data, performs page operations and data processing based on the callback data, and outputs business presentation results.

2. The unified agricultural business service system based on a large model according to claim 1, characterized in that, The user input module includes: An input receiving unit is used to receive natural language queries input by the user; The preprocessing unit is used to perform structured processing on natural language queries, remove redundant information, complete the context, and generate query information to be processed.

3. The unified agricultural business service system based on a large model according to claim 1, characterized in that, The service registration module includes: The service definition unit is used to define service types and label the corresponding service identifiers. Service types include three-capital services, transaction services, and agricultural services. The prompt word management unit is used to store system prompt words and user prompt word templates. System prompt words are used to define LLM roles, capability boundaries, and output format constraints, while user prompt word templates contain variable placeholders. The business matching unit matches service identifiers based on the query information to be processed and generates corresponding service business to be processed based on the matching results. Based on the pending service business, a search is performed in the prompt word management unit to obtain the corresponding matching system prompt words and user prompt word templates; Fill the variable placeholders in the user prompt word template with the query information to be processed to generate the corresponding user prompt word. Combine the system prompt word and the corresponding user prompt word to generate the initial prompt word for the sonic logging pipe. Map the initial prompt words to the prompt word association field to generate prompt word cells; The interface configuration unit is used to store the callback interface identifier and model call parameters corresponding to each service identifier, and generate interface configuration information based on the corresponding callback interface identifier and model call parameters of the service to be processed. The prompt word cell body and interface configuration information are encapsulated to generate model call data.

4. The unified agricultural business service system based on a large model according to claim 3, characterized in that, The step of mapping the initial prompt word to the prompt word association field and generating the prompt word cell body is as follows: The initial prompt word is input into the prompt word association field, which contains multiple prompt word fragments and the business semantic association relationships between each prompt word fragment; Based on business semantic association, retrieve prompt word fragments that have business relevance to the initial prompt word, encapsulate the retrieved prompt word fragments as latent prompt word clusters, and label each prompt word fragment with trigger condition tags; Construct a cue word cell body space, which has a two-layer storage structure including a recessive kernel and a dominant outer kernel, and the storage space of the dominant outer kernel is larger than that of the recessive kernel; Establish condition-triggered interfaces between the implicit kernel and the explicit outer kernel, and establish normally-pass interfaces on the explicit outer kernel; The latent prompt word cluster is stored in the latent kernel, and the initial prompt word is stored in the explicit outer kernel; The stored implicit kernel, explicit outer kernel, conditional trigger interface, and normally accessible interface are encapsulated into a prompt word cell.

5. The unified agricultural business service system based on a large model according to claim 1, characterized in that, The LLM module includes: The model adaptation unit is used to access general-purpose large language models and configure model call interfaces corresponding to service types; The parameter parsing unit is used to receive model call data, match the corresponding model call interface based on the model call data, connect with the service registration module, and activate the corresponding large language model. After determining the large language model, the large language model extracts model call parameters from the model call data. The model call parameters include output length limit parameters and response timeout parameters. The interface connection unit corresponds to the common interface of the large language model recognition prompt word cell body, and extracts the initial prompt word based on the common interface, performs inference based on the model call parameters, and generates inference results; Once the reasoning is complete, the user continues to input updated query information, and the corresponding large model receives the updated query information through the model call interface. Based on the activation of the condition trigger interface for updating the query information to be processed, the updated query information to be processed is transmitted to the hidden kernel through the constant-pass interface and the condition trigger interface. The trigger condition labels of the latent cue word clusters are matched, and the successfully matched cue word fragments are fed back to the large language model to continue the update inference and generate the update inference results. The reasoning results and the updated reasoning results are output as structured reasoning results.

6. The unified agricultural business service system based on a large model according to claim 1, characterized in that, The callback matching module includes: The callback configuration storage unit is used to store the callback interface configuration corresponding to each service identifier. The callback interface configuration includes the callback address, request method, data format and authentication information. The callback matching unit is used to receive the structured inference result and match the corresponding callback interface configuration from the callback configuration storage unit according to the service identifier corresponding to the structured inference result. The data encapsulation unit encapsulates the structured inference results into callback data based on the data format in the matched callback interface configuration. The callback data includes the action type and action parameters. The callback push unit is used to push callback data to the inference display module based on the callback address and authentication information.

7. The unified agricultural business service system based on a large model according to claim 1, characterized in that, The inference display module includes: The callback data receiving unit is used to receive callback data and parse the action type and action parameters in the callback data. The execution unit performs corresponding page operations and data processing based on the action type and action parameters, and outputs the page presentation results.

8. A unified agricultural business service method based on a large model, used to implement the unified agricultural business service system based on a large model as described in any one of claims 1-7, characterized in that, Receive natural language queries from users and generate query information to be processed; Configure service parameters, including service type, prompt word template, callback interface identifier, and model call parameters. Generate model call data based on the query information to be processed and the service parameters. Receive model call data, call the corresponding inference model based on the model call data to perform inference, and generate structured inference results; Based on the structured reasoning results, the callback interfaces are matched, and the callback data is output based on the successfully matched callback interfaces. Receive callback data, perform page operations and data processing based on the callback data, and generate business presentation results.

Images