Information generation system, method and apparatus

Abstract

Embodiments of the present specification provide an information generation system, method and device, wherein the information generation system comprises: an information processing module configured to determine inquiry information and inquiry memory information in response to an inquiry request submitted by a target user, and write the inquiry information and the inquiry memory information into a state object; a slot extraction module configured to read target inquiry information from the state object, perform slot extraction on the target inquiry information, and determine standard slot information according to a result of the slot extraction, under the condition that the target user is determined to have an inquiry intention according to the state object; an information query module configured to query commodity information corresponding to the standard slot information by calling a commodity query tool, and determine commodity transaction information according to the commodity information; and an information feedback module configured to generate initial inquiry feedback information according to the inquiry intention, the standard slot information and the commodity transaction information, update the initial inquiry feedback information into target inquiry feedback information by using a large language model, and send the target inquiry feedback information to the target user.

Description

Technical Field

[0001] The embodiments in this specification relate to the field of information processing technology, and in particular to information generation systems, methods and apparatus. Background Technology

[0002] In existing technologies, most trade industries exhibit significant characteristics such as highly dispersed product categories, complex combinations of specifications and parameters, and rapidly changing market prices. Customer communication heavily relies on instant messaging tools for real-time interaction. In these scenarios, customers rarely use standardized inquiry forms, instead employing fragmented expressions, including short sentences, line breaks, multiple rounds of dialogue to supplement details, industry abbreviations, and even typos to convey their needs. The content of the communication is often scattered across different points in the same conversation, forming unstructured and incomplete inquiry fragments. For this type of information, existing technologies primarily rely on manual parsing of customer needs and manual inventory and price checks by customer service or sales personnel; or automated routing systems based on keyword matching and regular expressions; or directly calling general-purpose language models to generate chat responses. While these methods have basic coverage capabilities in some standardized scenarios, they reveal systemic flaws in the actual business loop of trade. For example, they cannot accurately associate industry abbreviations with material types, nor can they intelligently link and complete product categories and specifications, resulting in low key parameter recognition rates. Meanwhile, the system design treats customer messages as isolated events, failing to maintain session context state. It cannot save intermediate state information awaiting follow-up, nor can it automatically integrate subsequent material or specification data provided by the customer into historical inquiry requests, resulting in information fragmentation and duplicate inquiries. Therefore, an effective solution is urgently needed to address these issues. Summary of the Invention

[0003] In view of the above, embodiments of this specification provide an information generation system. One or more embodiments of this specification also relate to an information generation system method, an information generation system apparatus, a computing device, a computer-readable storage medium, and a computer program product, to address the technical deficiencies existing in the prior art.

[0004] According to a first aspect of the embodiments of this specification, an information generation system is provided, comprising: The information processing module is used to determine the inquiry information and inquiry memory information in response to the inquiry request submitted by the target user, and write the inquiry information and the inquiry memory information into the status object; The slot extraction module is used to read target inquiry information from the status object when the target user is determined to have an inquiry intention based on the status object, extract slots from the target inquiry information, and determine standard slot information based on the slot extraction result. The information query module is used to call the product query tool to query the product information corresponding to the standard slot information, and determine the product transaction information based on the product information; The information feedback module is used to generate initial inquiry feedback information based on the inquiry intent, the standard slot information, and the commodity transaction information, update the initial inquiry feedback information to target inquiry feedback information using a large language model, and send the target inquiry feedback information to the target user.

[0005] According to a second aspect of the embodiments of this specification, an information generation method is provided, comprising: In response to a request for quotes submitted by a target user, determine the quote information and quote memory information, and write the quote information and quote memory information into a status object; If the target user is determined to be inquiring about pricing based on the status object, the target pricing information is read from the status object, the target pricing information is extracted into slots, and the standard slot information is determined based on the slot extraction result. Use the product query tool to query the product information corresponding to the standard slot information, and determine the product transaction information based on the product information; An initial inquiry feedback is generated based on the inquiry intent, the standard slot information, and the commodity transaction information. The initial inquiry feedback is then updated to a target inquiry feedback using a large language model, and the target inquiry feedback is sent to the target user.

[0006] According to a third aspect of the embodiments of this specification, an information generation apparatus is provided, comprising: The determining unit is configured to determine inquiry information and inquiry memory information in response to an inquiry request submitted by a target user, and write the inquiry information and the inquiry memory information into a status object; The extraction unit is configured to, when the target user is determined to have an inquiry intent based on the status object, read the target inquiry information from the status object, extract the slot from the target inquiry information, and determine the standard slot information based on the slot extraction result; The query unit is configured to call a product query tool to query the product information corresponding to the standard slot information, and determine the product transaction information based on the product information; The feedback unit is configured to generate initial inquiry feedback information based on the inquiry intent, the standard slot information, and the commodity transaction information, update the initial inquiry feedback information to target inquiry feedback information using a large language model, and send the target inquiry feedback information to the target user.

[0007] According to a third aspect of the embodiments of this specification, a computing device is provided, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the above-described information generation method.

[0008] According to a fourth aspect of the embodiments of this specification, a computer-readable storage medium is provided that stores computer-executable instructions, which, when executed by a processor, implement the steps of the information generation method described above.

[0009] According to a fifth aspect of the embodiments of this specification, a computer program product is provided, including a computer program or instructions that, when executed by a processor, implement the steps of the information generation method described above.

[0010] The information generation system provided in this embodiment responds to a price inquiry request submitted by a target user through an information processing module, determining price inquiry information and price inquiry memory information, and writing the price inquiry information and price inquiry memory information into a state object. Then, using a slot extraction module, if the target user is determined to have a price inquiry intent based on the state object, the system reads the target price inquiry information from the state object, performs slot extraction on the target price inquiry information, and determines standard slot information based on the slot extraction result. On this basis, an information query module calls a product query tool to query the product information corresponding to the standard slot information, and determines the product transaction information based on the product information. Finally, an information feedback module generates initial price inquiry feedback information based on the price inquiry intent, the standard slot information, and the product transaction information. Using a large language model, the initial price inquiry feedback information is updated to target price inquiry feedback information, and the target price inquiry feedback information is sent to the target user. This system integrates inquiry information with stored data through an information processing module, achieves accurate slot identification and completion through a slot extraction module, ensures the real-time and accuracy of commodity transaction information through an information query module, and generates reliable feedback based on a large language model through an information feedback module. This effectively solves the problems of difficult processing of fragmented inquiries, missing session context, and unreliable feedback information in trade. It can efficiently process fragmented inquiry requests in trade, accurately extract key slot information, maintain session context to avoid repeated inquiries, and generate reliable feedback information based on real-time queries, thereby improving the automation level of the inquiry process and the reliability of transactions. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the structure of an information generation system provided in one embodiment of this specification; Figure 2 This is a schematic diagram of the structure of another information generation system provided in one embodiment of this specification; Figure 3This is a flowchart illustrating an information generation method provided in one embodiment of this specification; Figure 4 This is a flowchart illustrating the processing procedure of an information generation system provided in one embodiment of this specification; Figure 5 This is a schematic diagram of the structure of an information generation device provided in one embodiment of this specification; Figure 6 This is a structural block diagram of a computing device provided in one embodiment of this specification. Detailed Implementation

[0012] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0013] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a,” “described,” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0014] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0015] Furthermore, it should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in one or more embodiments of this specification are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0016] This specification provides an information generation system. One or more embodiments of this specification also relate to an information generation system method, an information generation system apparatus, a computing device, a computer-readable storage medium, and a computer program product, which will be described in detail in the following embodiments.

[0017] In practical applications, existing steel trade inquiry processing often exhibits non-standardized expression in customer inquiry requests, including short sentences, line breaks, multiple rounds of supplementation, industry abbreviations, and typos. Existing technical solutions lack contextual association mechanisms and domain-specific slot completion capabilities, leading to ineffective handling of information fragmentation. Specifically, customer inquiry information submitted at different times is difficult for the system to integrate, resulting in inaccurate identification of inquiry intent. Furthermore, the lack of fact boundary control allows the system to generate responses without calling real-time data interfaces or provide false information when the interface yields no results, thus affecting the accuracy and automation level of inquiry processing. For example, in a steel trading platform, a customer submits multiple rounds of inquiry requests via instant messaging. The first message, "Pattern 3.5*1500 4.5*1500, check the price," is received by the system, but "pattern" is not completed to form patterned steel plate, and the pending follow-up status is not saved. Subsequently, the second message, "Does Liugang have 1.2*1250?", was isolated and failed to be linked to the specification information of the previous message, resulting in an incomplete query request. When the third message, "Q235B", was received, the material information was not merged into the previous inquiry, thus generating an incorrect inventory query. The system generated a reply without calling the real-time price interface, which may have claimed that the item was in stock when the actual inventory was insufficient, causing misunderstanding for the customer.

[0018] In view of this, see Figure 1 and Figure 2 , Figure 1 This specification shows a schematic diagram of the structure of an information generation system according to one embodiment. Figure 2 This specification illustrates a schematic diagram of another information generation system provided in one embodiment. The information generation system 100 includes an information processing module 110, a slot extraction module 120, an information query module 130, and an information feedback module 140, specifically comprising: Information processing module 110 is used to determine inquiry information and inquiry memory information in response to an inquiry request submitted by a target user, and write the inquiry information and the inquiry memory information into a status object; The slot extraction module 120 is used to read target inquiry information from the status object when the target user is determined to have an inquiry intention based on the status object, extract slots from the target inquiry information, and determine standard slot information based on the slot extraction result. The information query module 130 is used to call the product query tool to query the product information corresponding to the standard slot information, and determine the product transaction information based on the product information; The information feedback module 140 is used to generate initial inquiry feedback information based on the inquiry intent, the standard slot information and the commodity transaction information, update the initial inquiry feedback information to target inquiry feedback information using a large language model, and send the target inquiry feedback information to the target user.

[0019] The information generation system provided in this embodiment can be applied to any business transaction scenario, such as steel trading, furniture trading, and office supplies trading. It is used to solve the problems of difficulty in processing fragmented inquiries, lack of session context, and unreliable feedback information in existing technologies. It enables the generation of reliable feedback information based on real-time queries, thereby improving the automation level of the inquiry process and the reliability of transactions.

[0020] This embodiment uses the application of an information generation system in a steel trading scenario as an example for illustration. The same or corresponding descriptions in other scenarios can be found in the descriptions in this embodiment, and will not be elaborated further here.

[0021] Specifically, the information processing module is configured as the entry point for the system to receive user inquiry requests. It is responsible for the initial processing of received requests to identify and extract core inquiry content and construct context by combining it with the user's historical information. An inquiry request represents the input submitted by a target user to the system, containing information such as the price, specifications, and inventory of goods or services. This request can be in text form, such as a text message sent via instant messaging. Inquiry information represents the specific requirements for goods or services extracted from the user's inquiry request, such as product name, specifications, and materials. Inquiry memory information represents historical interaction data or preset user preference information related to the target user, used to provide context and personalized support for the current inquiry request. The state object is configured as a unified data structure within the system for storing and transmitting all relevant information for the current inquiry session. This object is read and updated at various stages of the inquiry processing flow to ensure the continuity and integrity of the information.

[0022] Accordingly, the slot extraction module is configured as a component of the system used to identify and extract structured data (i.e., "slots") from user inquiry information. These slots correspond to specific attributes of goods or services, such as product name, material, specifications, quantity, etc. Inquiry intent represents the system's judgment of the target user's current session purpose, such as whether the user wants to inquire about prices, check inventory, or perform other operations. Target inquiry information represents the user inquiry content read from the state object that requires slot extraction. Slot extraction refers to the process of converting unstructured text information into structured key-value pairs, for example, extracting "patterned steel plate 3.5mm" as "Product Name: Patterned Steel Plate, Thickness: 3.5mm". Standard slot information represents structured product attribute information that conforms to the system's internal query specifications after slot extraction and possible data standardization.

[0023] Correspondingly, the information query module is configured as a component of the system used to retrieve product-related information by calling external tools or internal databases based on standard slot information. The product query tool represents an external interface or internal service capable of querying real-time inventory, price, specifications, and other product data based on provided product attributes. Product information represents detailed data about a specific product obtained from the product query tool, such as inventory quantity, current price, and available specifications. Product transaction information represents transaction-related data determined based on the product information and available for feedback to the user, such as final price and inventory status.

[0024] Accordingly, the information feedback module is configured as a component of the system to generate and send inquiry feedback information to the target user based on the processing results. The initial inquiry feedback information represents the preliminary, unprocessed feedback content generated by the system based on the inquiry intent, standard slot information, and commodity transaction information. The large language model is configured as a deep learning-based natural language processing model capable of understanding, generating, and converting human language to refine and optimize the initial feedback information, making it more natural and conversational. The target inquiry feedback information represents the final, fluent, and accurate inquiry feedback content sent to the target user after processing by the large language model.

[0025] Based on this, the information processing module is configured to respond to price inquiry requests submitted by target users, determine price inquiry information and price inquiry memory information, and write these two pieces of information into a state object. Specifically, the information processing module is activated when the system receives a price inquiry request submitted by a user. This module can determine price inquiry information in various ways. For example, it can identify words related to product category, specifications, etc., from the user's input text using preset keyword matching rules. Simultaneously, price inquiry memory information can be determined as static user tags associated with the target user or a simple summary of historical interaction records. Subsequently, this determined price inquiry information and price inquiry memory information are integrated and written into a unified state object for processing by subsequent modules. For example, when a user inputs "need steel plate," the information processing module can identify "steel plate" as price inquiry information and combine it with the user's basic registration information in the system as price inquiry memory information, storing both together in the state object.

[0026] The slot extraction module is configured to read the target inquiry information from the status object when the target user's intent is determined to be an inquiry, extract slots from the target inquiry information, and determine standard slot information based on the extraction results. After the information processing module writes data to the status object, the slot extraction module checks the information in the status object to determine whether the user's intent is an inquiry. If it is determined to be an inquiry intent, the module will retrieve the target inquiry information from the status object. The slot extraction process can employ rule-based pattern matching methods, such as using predefined regular expressions to identify combinations of numbers and units in the text to extract specification information. In this way, unstructured target inquiry information is transformed into structured standard slot information, for example, extracting "3.5mm thickness" as "thickness: 3.5mm".

[0027] The information query module is configured to invoke a product query tool to retrieve product information corresponding to the standard slot information and determine product transaction information based on this information. After obtaining the standard slot information, the information query module sends it as a query parameter to a pre-defined product query tool. This product query tool can be an external product database interface or an internal inventory management system. The information query module receives product information returned by the product query tool, such as inventory quantity and price range. Subsequently, based on this product information and pre-defined business logic, the information query module determines the final product transaction information, such as the specific selling price and available quantity. For example, if the standard slot information is "Product Name: Patterned Steel Plate, Thickness: 3.5mm", the information query module will use this information to query and obtain product information such as "Patterned Steel Plate 3.5mm, Inventory 100 tons, Price 5000 yuan / ton", and then determine the product transaction information.

[0028] The information feedback module is configured to generate initial inquiry feedback based on the inquiry intent, standard slot information, and commodity transaction information. It then uses a large language model to update this initial feedback to a target inquiry feedback, which is sent to the target user. After the commodity transaction information is determined, the information feedback module combines the previously determined inquiry intent and standard slot information to generate a preliminary feedback text. This initial inquiry feedback may simply be a concatenation of information, such as "You inquired about 3.5mm patterned steel plate, priced at 5000 yuan / ton, available." To improve the naturalness of the feedback and user experience, this initial inquiry feedback is then processed by a large language model. The large language model refines the text and adjusts the word order to better align with human communication habits, thereby generating the target inquiry feedback. Finally, this target inquiry feedback is sent to the target user.

[0029] For example, User A sends a price inquiry to the system via instant messaging: "Does Liugang have Q235B 1.2*1250? How much?" First, the information processing module receives User A's inquiry. This module performs a preliminary analysis of the request text, identifying "Does Liugang have Q235B 1.2*1250? How much?" as a price inquiry. Simultaneously, the system extracts basic information from User A's historical interaction records based on User A's identity, such as previous inquiries about steel products. This price inquiry information and the inquiry memory are then written into a state object to construct the context of the current session. The slot extraction module reads the contents of the state object. By performing intent recognition on the information recorded in the state object, the system determines that User A's intent is "price inquiry." After confirming the price inquiry intent, the slot extraction module extracts "Q235B 1.2*1250 Liugang" from the state object as the target price inquiry. This module extracts slot information from target price inquiries. For example, using preset rules and a dictionary, it identifies "Q235B" as the material slot, "1.2*1250" as the specification slot, and "Liugang" as the steel mill slot. After extraction, standard slot information is generated, such as: {Material: Q235B, Specification: 1.2*1250, Steel Mill: Liugang}. The information query module receives this standard slot information. This module uses this structured slot information as parameters to call a preset product query tool. The product query tool performs a real-time query in the product database based on these parameters and returns the corresponding product information, such as: 100 tons of Q235B Liugang product of this specification are in stock, the current selling price is 4500 yuan / ton, and the update time was this morning. Based on this product information, the information query module determines the product transaction information, namely "Q235B 1.2*1250 Liugang, in stock, price 4500 yuan / ton". The information feedback module generates initial inquiry feedback based on the inquiry intent, standard slot information, and commodity transaction information, such as: "Your inquiry about Q235B 1.2*1250 Liugang has stock, priced at 4500 yuan / ton." To make the feedback more approachable, this initial inquiry feedback is input into a large language model. The large language model processes and polishes the text to make it more conversational, updating it to the target inquiry feedback, such as: "Hello, your inquiry about Q235B 1.2*1250 Liugang has stock, currently priced at 4500 yuan / ton. How much do you need?" Finally, this target inquiry feedback is sent to user A.

[0030] In summary, by integrating inquiry information and memory data through the information processing module, achieving accurate slot identification and completion through the slot extraction module, ensuring the real-time and accuracy of commodity transaction information through the information query module, and generating reliable feedback by combining a large language model through the information feedback module, this approach effectively solves the problems of difficult processing of fragmented inquiries, missing session context, and unreliable feedback information in trade. It is capable of efficiently processing fragmented inquiry requests in trade, accurately extracting key slot information, maintaining session context to avoid repeated inquiries, and generating reliable feedback information based on real-time queries, thereby improving the automation level of the inquiry process and the reliability of transactions.

[0031] Furthermore, when handling complex businesses such as steel trading, a single inquiry request often lacks sufficient contextual support, causing the system to fail to identify the customer's personalized needs, historical preferences, and current business status. This results in a lack of depth information for subsequent slot extraction and intent recognition, making it difficult to achieve accurate personalized services. To address this, in this embodiment, the information processing module is also used to receive inquiry requests submitted by the target user, parse the inquiry request to obtain inquiry information, and write the inquiry information into a status object; load the user profile, inquiry status, and memory information corresponding to the target user into a memory cache, compose inquiry memory information based on the user profile, the inquiry status, and the memory information, and write the inquiry memory information into the status object.

[0032] Specifically, receiving a price inquiry request submitted by a target user refers to the process by which the system obtains the user's inquiry intention or specific inquiry content. This serves as the starting point of the entire price inquiry processing flow, capturing the user's original needs. This process can be achieved by receiving text messages input by the user through instant messaging interfaces (e.g., WeChat, WeChat Work, DingTalk, etc.), or by receiving structured or unstructured data submitted by the user through web forms or mobile application interfaces. Parsing the price inquiry request to obtain inquiry information refers to performing structured or semi-structured processing on the received original price inquiry request, identifying and extracting information directly related to the current inquiry. Its function is to transform the user's free-form expression into a data format that the system can understand and process. This can be achieved by using Natural Language Processing (NLP) technology to segment the text, perform part-of-speech tagging, and name entity recognition, thereby extracting key information such as product name, specifications, quantity, and material; or by using preset regular expressions or template matching to extract corresponding fields from requests in specific formats. Writing the inquiry information into the state object means storing the current inquiry information parsed from the inquiry request into a unified state management entity. Its purpose is to provide the latest, standardized input data for subsequent stages such as intent recognition and slot extraction, and to ensure the continuity of the session context. This can be achieved by storing the inquiry information as key-value pairs in a hash table or dictionary structure in memory; or by serializing it into a JSON or XML string and storing it in a session state object maintained by the session management module.

[0033] Loading the user profile, inquiry status, and memory information corresponding to the target user into the memory cache refers to the system retrieving and obtaining historical data related to the current target user from the pre-stored memory cache. Its function is to provide rich contextual information for the current inquiry request, including the user's long-term preferences, historical transaction records, and the status of currently incomplete inquiry sessions. This can be achieved by using the user ID as an index to query and load the user profile (e.g., frequently used products, preferred brands), historical inquiry status (e.g., products in the last inquiry, incomplete orders), and long-term memory (e.g., user's habitual expressions, special requirements) from a distributed caching system (e.g., Redis, Memcached); or by reading this information from persistent storage (e.g., a database) and temporarily storing it in a fast-access area in memory. Composing inquiry memory information based on the user profile, inquiry status, and memory information refers to integrating multi-dimensional historical data loaded from the memory cache to form a comprehensive memory set reflecting the user's history and current business context. Its function is to provide a dynamic and personalized context for the current inquiry, enabling the system to more accurately understand the user's intent and needs. This can be achieved by structurally merging user profiles, query status, and memory information. For example, by organizing them into a unified JSON object or data structure that includes the user's basic attributes, historical interaction data, and the progress of the current session; or by using a specific algorithm to weightedly fuse this information to generate a comprehensive user memory vector. Writing the query memory information into the state object means storing the integrated query memory information along with the current query information in the state object. Its purpose is to ensure that the state object not only contains details of the current request but also rich historical context, providing a comprehensive basis for subsequent decisions and processing. This can be achieved by updating the composed query memory information as a specific field or nested structure of the state object; or by merging the query memory information with existing query information to form a more complete and context-aware state representation, overwriting the original state object content.

[0034] Based on this, a memory caching mechanism was introduced to achieve dynamic integration of comprehensive user information, thereby solving the problem of insufficient information dimensions in a single inquiry request. The information processing module first receives the inquiry request submitted by the target user, which is the starting point of the entire inquiry process. After receiving the request, the system parses it and extracts the core content of this inquiry, namely the inquiry information. This inquiry information is then written into a state object as the basic data for the current session. To overcome the problem of insufficient information dimensions in a single request, the information processing module actively loads user profiles, inquiry statuses, and memory information related to the target user into the memory cache. The user profile provides the user's long-term preferences and background information, the inquiry status reflects the user's current unfinished business progress, and the memory information contains the user's past communication habits and historical interactions. This multi-dimensional data is integrated to form comprehensive inquiry memory information. Finally, this composition of inquiry memory information is also written into the same state object. In this way, the state object no longer only contains the user's current inquiry content, but also integrates the user's historical background, behavioral preferences, and real-time business status. This dynamic integration transforms the state object into a carrier rich in contextual information, providing more accurate and comprehensive input for subsequent slot extraction, information query, and information feedback modules. For example, when processing target inquiry information, the slot extraction module can combine the user profile in the state object to more accurately identify industry abbreviations or specific terms; when querying product information, the information query module can refer to the inquiry status to avoid duplicate queries or providing irrelevant results; and when generating feedback information, the information feedback module can utilize memory information and adopt more colloquial expressions that align with user habits. This mechanism ensures that the system can process inquiries based on a comprehensive understanding of the user, thereby significantly improving its ability to understand and respond to complex and fragmented inquiry needs.

[0035] For example, suppose a steel trading customer sends a price inquiry message to the system via WeChat: "Pattern 3.5*1500 4.5*1500, please check the price." First, the information processing module receives this inquiry request. The system will launch a parser, such as a Natural Language Understanding (NLU) component based on rules and machine learning models, to analyze "Pattern 3.5*1500 4.5*1500, please check the price." The parsing result might identify "pattern" as the product category, "3.5*1500" and "4.5*1500" as specifications, and determine that the user's intent is "price inquiry." This parsed information, including product category, specifications, and intent, constitutes the price inquiry information. These inquiry information are then written to the AgentState's in-memory state object, for example, stored in JSON format: {"current_request": {"category": "pattern", "specs": ["3.5*1500", "4.5*1500"], "intent": "inquiry"}}. Simultaneously, the system retrieves the customer's historical data from the memory cache based on the customer's WeChat ID. The memory cache can be a distributed key-value store system, such as Redis. The system loads the customer's user profile (e.g., the customer frequently purchases Q235B steel and prefers the Liuzhou Steel brand), their current inquiry status (e.g., the last inquiry for "1.2*1250 Liuzhou Steel" is not yet complete and is in the material supplementation stage), and memory information (e.g., the customer habitually uses the abbreviation "Pu" to represent ordinary carbon steel). The loaded information is integrated to form inquiry memory information, such as: {"user_profile": {"material_preference": "Q235B", "brand_preference": "Liugang"}, "inquiry_status": {"last_inquiry_specs": "1.2*1250", "status": "waiting_material_supplement"}, "memory": {"abbreviation_map": {"Pu": "Pu carbon steel"}}}. Finally, this assembled inquiry memory information is also written into the AgentState state object mentioned above, and merged with the current inquiry information.At this point, AgentState might be updated to: {"current_request": {"category": "pattern", "specs": ["3.5*1500", "4.5*1500"], "intent": "inquiry"}, "user_profile": {"material_preference": "Q235B", "brand_preference": "Liugang"}, "inquiry_status":{"last_inquiry_specs": "1.2*1250", "status": "waiting_material_supplement"},"memory": {"abbreviation_map": {"plain": "plain carbon steel"}}}. In this way, the subsequent slot extraction module, when processing "pattern," can combine user profile and memory information to more accurately understand its specific meaning and provide richer context for subsequent product queries and feedback.

[0036] In summary, by loading and integrating the target user's profile, inquiry status, and memory information into the memory cache, the system can deeply correlate the current request with the user's historical behavior, long-term preferences, and current business progress. This dynamic integration allows the state object to contain a complete contextual environment of the current demand and historical background, thereby significantly improving the system's ability to understand and respond to complex and fragmented inquiry requests.

[0037] Furthermore, the system often lacks a pre-judgment mechanism for user input intent, making it unable to distinguish whether the user has a clear inquiry intent or is in a non-inquiry communication stage. This results in a lack of flexible response strategies when faced with non-inquiry intent input, easily leading to inaccurate responses or rigid processes, failing to meet the complex and ever-changing communication needs of steel trade. To address this, in this embodiment, the information generation system also includes an intent recognition module. This module is used to identify the intent of the information recorded in the state object according to preset intent recognition rules and intent classification models, and determine the intent tag based on the intent recognition result. If the intent tag is an inquiry intent tag, the target user is determined to have an inquiry intent. If the intent tag is not an inquiry intent tag, inquiry communication information is sent to the target user according to a preset response strategy.

[0038] Specifically, the intent recognition module is a functional unit within the information generation system. Its primary responsibility is to analyze user input and determine the user's core intent, providing a basis for decision-making in subsequent business processes. This module can be an independent software service that interacts with the information processing module via an API interface, receiving information from state objects and returning intent tags; alternatively, it can be a sub-component integrated within the information processing module, directly processing user input data received by the module. Predefined intent recognition rules refer to a series of conditions, patterns, or logics predefined during system design based on business scenarios and domain knowledge to determine user intent. These rules can be a set of rules based on regular expressions, keyword matching, or semantic templates; for example, statements containing keywords such as "price," "how much," or "quote" may be identified as an inquiry intent. Alternatively, they can be decision trees or rule engines built based on expert experience.

[0039] Correspondingly, the intent classification model is a machine learning model that, after training, can classify user-input text into predefined intent categories. It can be a deep learning-based text classification model, such as fine-tuning pre-trained language models like BERT or RoBERTa to adapt to intent recognition tasks in the steel trading industry; or it can be a model built based on traditional machine learning algorithms (such as Support Vector Machines (SVM) and Naive Bayes). Intent recognition of information recorded in the state object aims to extract intent from the current context of user interaction with the system. The intent recognition module can read the most recently submitted inquiry request text and historical dialogue records stored in the state object, and comprehensively analyze them to determine the user's intent; alternatively, the intent recognition module can analyze only the most recently written inquiry information in the state object to quickly determine the intent of the current round. Determining intent labels based on the intent recognition results involves outputting the analysis results of the intent recognition module in a standardized label format for subsequent processing. Intent labels can be an enumerated value, such as "inquiry," "order placement," "negotiation," or "casual conversation"; or they can be structured data containing intent type and confidence score. When the intent tag is "inquiry intent," identifying the target user as having an inquiry intent is a crucial branch point in the system process. It confirms whether the user genuinely has an inquiry need. The system can directly classify intent tags of "inquiry" or "purchase request" as inquiry intent and trigger subsequent inquiry processing. Alternatively, in some high-risk scenarios, the system can further confirm this by combining user profiles and historical behavior information after identifying the inquiry intent tag. When the intent tag is not "inquiry intent," an inquiry communication message is sent to the target user according to a preset response strategy. This aims to provide a flexible handling mechanism for non-inquiry intents, avoiding rigid system responses. The preset response strategy can be a set of templated responses for different non-inquiry intents; alternatively, it can combine a large language model to generate more personalized and guiding communication messages based on the non-inquiry intent and conversation context.

[0040] Based on this, by introducing an intent recognition module, a crucial logical filtering layer is added to the information processing flow, enabling accurate classification and differentiated processing of user input intent. When the information processing module responds to a price inquiry request submitted by a target user, determines the price inquiry information and its memory, and writes them into the state object, the intent recognition module begins to function. Following preset intent recognition rules and intent classification models, the intent recognition module performs in-depth analysis of the current session information recorded in the state object, thereby accurately determining the user's intent tag. This dual-driven recognition mechanism effectively addresses the diverse and colloquial nature of user expressions in steel trading, ensuring the accuracy of intent judgment.

[0041] Specifically, if the intent recognition module identifies the intent tag as a price inquiry intent tag, the system will explicitly determine the target user's current need as a price inquiry intent. Based on this, the system will activate the slot extraction module, enabling it to read the target price inquiry information based on the identified price inquiry intent in the state object, extract the slots, and determine the standard slot information. Subsequently, the information query module will call the product query tool, using this standard slot information to query the corresponding product information and ultimately determine the product transaction information. Finally, the information feedback module will generate initial price inquiry feedback information based on this series of processing results—price inquiry intent, standard slot information, and product transaction information—and update it using a large language model to create a more natural and user-friendly target price inquiry feedback information, which will then be sent to the target user. Conversely, if the intent recognition module identifies the intent tag as a non-price inquiry intent tag, the system will not proceed to the subsequent slot extraction and product query processes. Instead, it will send the corresponding price inquiry communication information to the target user according to a preset response strategy. This mechanism cleverly and effectively separates price inquiry business from regular communication. By identifying intent early in the processing flow, the system avoids unnecessary resource consumption in non-inquiry scenarios, such as blindly calling product search tools. Simultaneously, by sending inquiry communication messages, the system proactively guides users, maintains communication continuity, and provides services more tailored to their current intent. This solves the problem of the system's simplistic response logic when facing different user intents, providing a reliable business prerequisite for subsequent accurate inquiries. This pre-emptive intent judgment mechanism enables the entire information generation system to demonstrate greater flexibility and intelligence when handling complex and ever-changing communication needs in steel trade.

[0042] For example, suppose a target user sends a message via instant messaging: "Do you have Q235B rebar, 12mm, how much?". Upon receiving this inquiry, the information processing module writes the inquiry information along with inquiry memory information (including user profile and inquiry status loaded from the memory cache) into a status object. At this point, the intent recognition module begins its work. This module can be configured with an intent classification model based on the Transformer architecture. This model, trained on extensive dialogue data from the steel trading sector, can identify various intents such as "inquiry," "order placement," "negotiation," and "casual conversation." Simultaneously, the intent recognition module incorporates a set of preset intent recognition rules. For instance, if the message contains keywords such as "how much," "quote," or "price," the rules tend to identify it as an inquiry intent. For the message "Do you have Q235B rebar, 12mm, how much?", the intent recognition module combines the model's prediction results with rule matching to ultimately determine its intent label as "inquiry intent label." Since the intent label is an inquiry intent label, the system determines that the target user currently has an inquiry intent. Subsequently, the system will invoke the slot extraction module to read this message from the status object as the target inquiry information, extract its slot, identify information such as "Material: Q235B", "Category: Rebar", and "Specification: 12mm", and standardize it into standard slot information. Next, the information query module will use this standard slot information to call the internal product query tool to check the corresponding product inventory and real-time price, and determine the product transaction information. Finally, the information feedback module will generate initial feedback based on this information and process it into conversational language using a large language model, for example, replying, "Hello! Q235B rebar 12mm is currently in stock, the real-time price is XXXX yuan / ton, how many tons do you need?" and send it to the user. Alternatively, if the target user sends the message "The weather is so nice today," the information processing module will also write it into the status object. When processing this message, the intent recognition module will identify it as "casual conversation," a non-inquiry intent tag, based on its content, combined with the intent classification model and preset rules. In this situation, the system will determine that the target user is not intending to inquire about prices, and will send an inquiry communication message to the target user according to the preset response strategy, such as "Hello, how can I help you inquire about prices?" or "It's a pleasure to serve you. What steel information do you need to inquire about?", thereby avoiding unnecessary slot extraction and product query operations, and guiding the user into the inquiry process.

[0043] In summary, by introducing an intent recognition module, the problem of traditional systems lacking a pre-judgment mechanism for user intent when processing user input is effectively solved, making it impossible to distinguish whether the user is in a clear inquiry or in a non-inquiry communication phase. By using pre-defined intent recognition rules and intent classification models to identify intent in the information recorded in the state object at an early stage of the information processing flow, the user's intent label can be accurately determined.

[0044] Furthermore, customer-entered inquiry information often contains industry abbreviations, non-standard formats, or incomplete descriptions. Directly extracting slot information can easily lead to inaccurate results or missing information, making it difficult to meet the need for accurate subsequent queries of product transaction information. Therefore, in this embodiment, the slot extraction module is further used to extract slots from the target inquiry information to obtain initial slot information; normalize the initial slot information to obtain intermediate slot information; and supplement the intermediate slot information to obtain standard slot information.

[0045] Specifically, slot extraction is performed on the target inquiry information to obtain initial slot information, aiming to identify preliminary key information elements from unstructured user inquiry text. This step can be implemented using various techniques. For example, pattern matching can be performed based on a pre-set keyword dictionary and regular expression rules to identify information such as product name, material, specifications, and quantity; alternatively, named entity recognition (NER) models from the field of natural language processing can be used, such as those based on Conditional Random Fields (CRF) or deep learning models (such as Bi-LSTM-CRF, BERT, etc.) for training, to automatically identify and extract entity information from the text. Through this process, the system can transform raw, free-form inquiry text into structured initial slot information, laying the foundation for subsequent refined processing.

[0046] Correspondingly, normalizing the initial slot information to obtain intermediate slot information addresses the inconsistency caused by the diversity of user expressions. In steel trading, users often use abbreviations, colloquialisms, or non-standard formats. This processing step can establish an industry terminology dictionary and thesaurus, mapping non-standard expressions identified in the initial slot information (such as "ordinary" and "manganese") to standardized industry terms (such as "ordinary coil" and "manganese plate"). Furthermore, semantic similarity calculation technology can be used to unify information with the same meaning but different expressions by comparing the semantic similarity between user input and standard terms. For example, specifications like "1.2*1500" can be standardized into the system's internal standard specification format. Through normalization, it is ensured that information from different users and with different expressions can be uniformly understood and processed by the system.

[0047] Correspondingly, the intermediate slot information is supplemented to obtain standard slot information, aiming to solve the problem of incomplete user inquiry information and fill in missing key attributes. For example, a user may only mention "ordinary" without specifying the material. This step can be based on a preset industry knowledge graph or business rule base for reasoning. For example, when the product name is "ordinary roll," the system can automatically supplement its default material as "Q235B"; when "manganese" is mentioned, it can supplement it as "manganese plate" or "low alloy" based on the context or common combinations, and further supplement the material as "Q355B." In addition, historical inquiry data and machine learning models can be combined to predict and supplement key information that users may have missed. Through slot supplementation, it is ensured that the final generated standard slot information is complete and can be used for subsequent product queries.

[0048] Based on this, by decomposing the complex slot extraction process into three stages—"extraction-normalization-supplementation"—the system can more effectively handle unstructured, incomplete, and diverse inquiry information in steel trading. First, the information processing module responds to the inquiry request submitted by the target user, determines the inquiry information and inquiry memory information, and writes them into the state object. Then, the slot extraction module, upon confirming the target user's intent to inquire, reads the target inquiry information from the state object. At this point, the slot extraction module no longer performs a simple one-time slot extraction but first performs preliminary slot extraction on the target inquiry information to obtain initial slot information. This step quickly captures explicit elements in the user's expression. Next, these initially extracted slot information, which may contain non-standard expressions, are normalized, converting them into unified intermediate slot information within the system, thereby eliminating ambiguity caused by differences in user expression habits. Finally, for any intermediate slot information that may still be missing after normalization, intelligent slot supplementation is performed. For example, material information is automatically supplemented based on product category, and thickness range is supplemented based on specifications, ultimately forming complete and accurate standard slot information. This multi-stage, progressive processing mechanism significantly improves the accuracy and completeness of parsing complex inquiry information. It enables the information query module to more accurately call the product query tool to query corresponding product information based on high-quality standard slot information, thereby determining product transaction information. This effectively solves the technical problem of traditional solutions failing to form a stable closed loop in steel trading scenarios due to inaccurate or missing information.

[0049] For example, suppose a target user submits an inquiry message as "pattern 3.5*1500 4.5*1500, please check the price". First, the slot extraction module extracts slot information from this inquiry message to obtain initial slot information, such as identifying "pattern" as the product name and "3.5*1500" and "4.5*1500" as specifications. Next, this initial slot information is normalized, for example, "pattern" is normalized to "patterned plate", and "3.5*1500" and "4.5*1500" are normalized to standard specification formats. Based on this, slots are supplemented. For example, based on the product name "patterned plate", the system can automatically supplement its commonly used material as "Q235B" or "Q345B" according to preset rules or a knowledge base, and supplement the unit as "ton" or "sheet". Ultimately, the system obtains standard slot information containing complete information such as product name, material, specifications, quantity, and unit. For example: Product name: Patterned plate, Material: Q235B, Specification: 3.5*1500, Quantity: To be determined; Product name: Patterned plate, Material: Q235B, Specification: 4.5*1500, Quantity: To be determined. Another example is when a user enters "Does Liugang have 1.2*1250? How much?", the initial extraction might yield the specification "1.2*1250" and the steel mill "Liugang". After normalization, the specification format is standardized. During the slot replenishment stage, the system can supplement the product name as "hot-rolled coil" or "cold-rolled coil" based on the steel mill information "Liugang" and its main products, and supplement the material as "Q235B" or "SPCC". In this way, even short or incomplete inquiries can be parsed by the system into structured and complete standard slot information.

[0050] In summary, the above processing effectively addresses the issues of fragmented, non-standardized, and incomplete customer inquiry information in steel trading. Through multi-stage slot processing, the accuracy and completeness of identifying user intent and product attributes are significantly improved, thus providing high-quality input for subsequent information query modules and ensuring the accurate acquisition of commodity transaction information. This not only enhances the automation level of inquiry processing but also reduces the need for manual intervention, improving the overall system efficiency and user experience.

[0051] Furthermore, existing information generation systems often lack effective interaction mechanisms to guide users to supplement key information when the initial inquiry information is incomplete or does not meet feedback conditions. They also fail to effectively integrate subsequently supplemented, fragmented information with the historical session context, making it difficult for the system to accurately identify complete transaction needs and consequently affecting the accuracy of subsequent product queries and feedback. To address this, in this embodiment, the information generation system further includes an information merging module. This module is used to send an inquiry supplementation request to the target user when the target inquiry information does not meet the inquiry feedback conditions, and to receive supplementary inquiry information from the target user in response to the inquiry supplementation request. The inquiry supplementation request is determined based on the session context and the pending inquiry status. The module performs slot extraction on the supplementary inquiry information to obtain supplementary slot information. Finally, it merges the supplementary slot information with the slot extraction results corresponding to the target inquiry information to obtain standard slot information.

[0052] Specifically, the information merging module is a functional unit whose main function is to process and integrate information acquired from different sources or at different times. This module can be a standalone hardware unit, such as a dedicated processor or microcontroller configured with corresponding memory and communication interfaces; or it can be a software module, such as a set of program code running on a server or cloud computing platform, responsible for performing specific data processing and logical judgment tasks. Its core function is to receive, process, and output structured information to support subsequent business processes. When the system determines that the existing target inquiry information is insufficient to generate effective inquiry feedback, the information merging module will proactively initiate interaction with the user. Target inquiry information not meeting the inquiry feedback conditions means that the system, based on preset business rules or models, determines that the currently collected inquiry information (e.g., lacking key product specifications, materials, quantities, etc.) is insufficient for accurate product searching or generating meaningful feedback. Sending a supplementary inquiry request to the target user is the system's proactive action to guide the user to provide missing information; this request can be a text message, voice prompt, or other form of interactive prompt. Receiving supplementary inquiry information from the target user in response to the supplementary inquiry request means that the system receives the user's response to the supplementary request. The request for supplementary pricing information is determined based on the session context and the pending status. The session context refers to all relevant information from the current user's historical interactions with the system, including previous pricing inquiries, system feedback, and user questions. The pending status refers to a state variable maintained internally by the system, recording the specific information type or field the system is currently waiting for the user to supplement. Slot extraction of the supplementary pricing information involves identifying and extracting structured key information from unstructured text information. This can be done using rule-based methods, such as regular expression matching, or machine learning-based methods, such as Named Entity Recognition (NER) models or intent-slot joint models. Merging the supplementary slot information with the slot extraction results corresponding to the target pricing information to obtain standard slot information involves logically integrating the supplementary slot information with the slot results previously extracted from the target pricing information. This merging process may involve deduplication, conflict resolution, and information completion.

[0053] Based on this, an information merging module was introduced to construct a closed-loop interactive information completion mechanism, effectively solving the problem of information loss in complex inquiry scenarios. In the information generation system, the information processing module responds to the inquiry request submitted by the target user, determines the inquiry information and inquiry memory information, and writes this information into the state object. Subsequently, the slot extraction module extracts slots based on the information in the state object, attempting to determine standard slot information. However, in practical applications, the initial inquiry information submitted by the user may be incomplete. At this time, the information merging module plays a crucial role; it determines whether the current target inquiry information meets the inquiry feedback conditions. If not, the information merging module will not directly pass the incomplete information to the information query module, but will send a targeted inquiry supplementation request to the target user based on the session context and the pending inquiry status. For example, if the system has identified the product name and specifications but lacks material information, the pending inquiry status will record "material" as a supplementary item, and the information merging module will generate a supplementary request asking, "What specific material do you need?" When the target user provides supplementary inquiry information, the information merging module extracts slot information from this supplementary information again to obtain supplementary slot information. Finally, the information merging module intelligently merges this supplementary slot information with the slot results previously extracted from the target inquiry information, thereby forming a complete and accurate standard slot information. This iterative and merging mechanism ensures that even if the initial inquiry information is incomplete, the system can eventually obtain complete inquiry elements through multiple rounds of interaction. This provides reliable and comprehensive input for the subsequent information query module to call the product query tool to query product information, greatly improving the robustness and accuracy of the entire information generation system in handling complex and fragmented inquiry scenarios.

[0054] For example, suppose a target user sends a price inquiry message: "Pattern 3.5*1500, please check the price." The information processing module receives and processes the inquiry request, writing the relevant information into a status object. The slot extraction module reads the target price inquiry information from the status object and performs slot extraction, identifying "Product Name: Pattern" and "Specification: 3.5*1500". At this point, the information merging module determines that the currently extracted slot information does not meet the price inquiry feedback conditions because the crucial "material" information is missing. Based on the session context (identified product name and specification) and the pending inquiry status (missing material), the information merging module sends a supplementary price inquiry request to the target user, for example: "Boss, is this ordinary or manganese?". After receiving the request, the target user replies: "Q235B". The information merging module receives this supplementary price inquiry information and performs slot extraction, obtaining the supplementary slot information "Material: Q235B". Subsequently, the information merging module combines this supplementary slot information with the previously extracted "Product Name: Pattern" and "Specifications: 3.5*1500" to obtain the complete standard slot information, namely "Product Name: Pattern, Specifications: 3.5*1500, Material: Q235B". This complete standard slot information will then be passed to the information query module for further product information retrieval.

[0055] In summary, the above-described approach effectively addresses the challenges faced by traditional systems when processing fragmented, multi-round inquiry information. This solution, by introducing an information merging module, proactively follows up on incomplete inquiry information and integrates information from different rounds, ensuring the system can construct complete and accurate standard slot information from fragmented user expressions. This not only avoids query failures or inaccurate feedback due to missing information but also significantly improves user communication efficiency and experience, making the entire inquiry process more intelligent and closed-loop, particularly suitable for complex and volatile industry scenarios such as steel trading.

[0056] In practical implementation, when faced with complex inquiry scenarios where standard slot information corresponds to multiple product specifications, simple query methods can easily lead to chaotic or missing information feedback. Furthermore, without an effective control mechanism, when the query interface fails to return valid product information, the system may be unable to promptly provide users with accurate business status updates, resulting in a degraded user experience or misleading information. Therefore, in this embodiment, the information query module is further configured to, when the standard slot information corresponds to multiple product specifications, invoke a product query tool to query the product information corresponding to the standard slot information according to a line-by-line query strategy, and determine the product transaction information based on the product information. The information query module is configured with control rules, which are used to provide control information to the target user when product information is not obtained.

[0057] In real-world steel trading inquiry scenarios, user requests often include requirements for multiple product specifications. For example, a user might simultaneously inquire about the prices of "pattern 3.5*1500" and "4.5*1500". This situation of "multiple product specifications" means the system needs to be able to identify and process these independent specification information separately, rather than treating them as a single, homogeneous query object. The "item-by-item query strategy" means that when the information query module receives standard slot information containing multiple product specifications, it does not merge these specifications into a single query request, but instead initiates a separate query request for each independent product specification to the product query tool. One implementation approach is for the information query module to maintain a list of specifications to be queried and iterate through each specification in the list, calling the product query tool independently for each specification. Another implementation approach is for the information query module to generate an independent query task for each product specification and utilize concurrent processing mechanisms (e.g., multithreading or asynchronous calls) to execute these query tasks simultaneously, thereby improving query efficiency. Product information typically includes detailed data such as inventory levels, real-time prices, warehouse locations, and suppliers. Determining product transaction information refers to the information query module filtering, integrating, and extracting key information directly related to the target user's inquiry intent and usable for transaction decisions from the raw product data. This key information may include available quantity, final price, delivery cycle, minimum order quantity, etc. As one implementation, the information query module can process the retrieved product information according to preset business rules (e.g., combining user level, historical transaction volume, etc.) to generate personalized product transaction information. As another implementation, the information query module can compare and sort product information provided by different suppliers to determine the optimal product transaction information.

[0058] Correspondingly, control rules are a set of pre-defined business logic and processing strategies used to regulate the behavior of the information query module under specific circumstances, especially the response mechanism when the product query tool fails to return valid results. Their role is to ensure that the system provides feedback that conforms to business specifications and user expectations under any circumstances, avoiding information distortion or misleading information. As one implementation method, control rules can define the number of retries for failed queries, timeout periods, circuit breaker mechanisms, etc., to improve query robustness. As another implementation method, control rules can include special processing logic for specific product types, price ranges, or user groups; for example, triggering manual intervention or providing alternative solutions when certain sensitive products are out of stock. "No product information obtained" means that the product query tool fails to return valid product data that can be used for transactions, which may include various situations such as no product inventory, query interface abnormalities, or non-existent product information. "Control information" is feedback sent by the information query module to the target user according to pre-defined control rules when it fails to obtain valid product information, aiming to clarify the current business status and guide the user to the next step. As one approach, control messages can be pre-set standardized scripts, such as "Sorry, the product you inquired about is currently out of stock. Do you need other specifications?" or "Let me check the price again and send it to you once it's confirmed." Another approach is to guide the target user to a live customer service representative, provide alternative product recommendations, or record the user's needs for follow-up.

[0059] Based on this, the information query module was enhanced to effectively handle complex multi-specification inquiry scenarios and ensure the accuracy and reliability of feedback information. Specifically, after the information processing and slot extraction modules have processed the data, when the information query module receives standard slot information containing multiple product specifications, it no longer performs a simple one-time query but intelligently adopts a query strategy item by item. This means that for each independent product specification in the inquiry request, the information query module will call the product query tool separately to query, thereby ensuring that the product information for each specification can be accurately retrieved and processed, avoiding the confusion or omission of specification information that may be caused by batch queries. This refined query method makes the subsequent process of determining product transaction information based on product information more accurate and reliable. On this basis, in order to further improve the robustness of the system and user experience, the information query module is also configured with control rules. These control rules come into play when the product query tool fails to obtain product information (e.g., product out of stock, query interface abnormality, or timeout). In this case, the control rules will guide the information query module to generate and provide clear control information to the target user, rather than generating false or misleading responses. This mechanism effectively solves the problem of the system blindly replying or providing incorrect information when product information is unavailable, ensuring the factual boundaries of information feedback. Through the above collaborative work, the information query module can not only efficiently and accurately process multi-specification inquiries, but also provide appropriate feedback in abnormal situations, thereby ensuring the smoothness and professionalism of the entire inquiry process and significantly improving users' trust in the inquiry results.

[0060] For example, suppose a target user sends a price inquiry request via instant messaging: "Pattern 3.5*1500 4.5*1500, please check the price." The information processing module receives and parses the request, and the slot extraction module identifies two product specifications: "Pattern 3.5*1500" and "Pattern 4.5*1500," and passes them as standard slot information to the information query module. Upon receiving these two product specifications, the information query module first calls a token acquisition interface (e.g., getMerToken) to obtain and cache a merchant token. This token can use the tenant and application ID as cache keys and be set with an appropriate expiration time to avoid duplicate token acquisition in high-concurrency scenarios. Subsequently, the information query module initiates queries for these two specifications separately. Specifically, it calls a product query tool (e.g., productApiSearch) to query the inventory and price information for "Pattern 3.5*1500," and also calls productApiSearch to query for "Pattern 4.5*1500."

[0061] Furthermore, to optimize performance and resource utilization, the information query module can employ a concurrency control mechanism, such as setting a concurrent semaphore to limit the maximum number of parallel queries. During the query process, if the query for "pattern 3.5*1500" is successful and returns valid product information (such as sufficient stock and clear price), the information query module will determine the corresponding product transaction information based on this information. However, if the query for "pattern 4.5*1500" fails, for example, if the product query tool returns "out of stock," the interface times out, or other exceptions occur, the control rules configured in the information query module will be triggered. According to the control rules, the system will not directly expose the interface error information to the target user, nor will it claim that the product is in stock. Instead, it will generate a preset control message, such as "I'll double-check the price and send it to you once it's confirmed" or "Sorry, the 'pattern 4.5*1500' you inquired about is currently out of stock. Do you need other specifications?" Finally, the information query module will pass the successfully retrieved product transaction information and the control message generated for the failed query to the information feedback module, which will then integrate them and send them to the target user.

[0062] In summary, the above-mentioned methods effectively solve the problem of information confusion or omissions caused by traditional simple query methods when dealing with complex inquiry scenarios involving multiple product specifications corresponding to standard slot information. The introduction of a line-by-line query strategy ensures that each individual product specification in the inquiry request can be accurately and thoroughly queried and processed, thereby significantly improving the accuracy of determining product transaction information.

[0063] Furthermore, responses generated solely based on large language models often lack specificity and struggle to incorporate customer historical preferences, business logic, and professional terminology. This can result in feedback information that is poorly conversational, lacks rigorous business logic, or fails to reflect personalized service. To address this, in this embodiment, the information feedback module is further configured to read the memory information from the state object and determine a preset terminology knowledge base; generate initial inquiry feedback information based on the terminology knowledge base, the memory information, the inquiry intent, the standard slot information, and the product transaction information; and use a large language model to process the initial inquiry feedback information into conversational language to obtain the target inquiry feedback information.

[0064] Specifically, memory information refers to historical interaction data stored by the system for a specific user or session, including but not limited to user profiles, past inquiry statuses, transaction habits, and preferences. Its purpose is to provide personalized and context-sensitive evidence for subsequent inquiry feedback, ensuring that the feedback content remains consistent with the user's historical behavior and needs. This can be achieved by retrieving data from persistent storage (e.g., relational or NoSQL databases) using the user ID, or loading it from the session cache. The script knowledge base is a pre-defined, structured collection of industry-specific terminology, standard response templates, sales scripts, business rules, and frequently asked questions. Its purpose is to ensure that the generated feedback information conforms to industry standards, business logic, and professional expression habits. The script knowledge base can be a structured database containing standard script templates and variable placeholders for different scenarios, or it can be a rule-based system that selects corresponding script segments by matching keywords or scenario tags. Generating initial inquiry feedback information refers to constructing a preliminary, structured draft feedback containing core business facts based on multi-source information (including the script knowledge base, memory information, inquiry intent, standard slot information, and product transaction information). This can be achieved by embedding retrieved product information, user memory, and other data into preset script templates through template matching and variable imputation, or by using a rule engine to select and combine appropriate sentence fragments from a script knowledge base based on inquiry intent and slot information. Using a large language model to process initial inquiry feedback into conversational language transforms structured or somewhat stiff initial feedback into more natural and conversational expressions that align with everyday communication habits, thereby improving user experience. This can be achieved by using the initial inquiry feedback as input and instructing the large language model's text generation capabilities to rewrite it in a conversational and friendly style. Alternatively, it can be combined with specific prompts (Prompt Engineering) to guide the large language model to perform style conversion while maintaining information accuracy, such as adding greetings or emojis.

[0065] Based on this, the system operates as follows: The information feedback module first reads the memory information from the state object. This memory information is composed of the user profile corresponding to the target user, the inquiry status, and the memory information, and then written into the state object. Simultaneously, the information feedback module determines a preset dialogue knowledge base. After obtaining the inquiry intent, standard slot information, and product transaction information, the information feedback module comprehensively utilizes the memory information, the dialogue knowledge base, and the aforementioned other information to generate an initial inquiry feedback message. This initial inquiry feedback message is a fact-based, structured feedback draft that includes all necessary business data and professional terminology. Subsequently, the system uses a large language model to process this initial inquiry feedback message into conversational language. The main task of the large language model at this stage is to optimize the expression of the text, making it more approachable and in line with everyday communication habits, rather than generating new factual information. Through this step-by-step processing, this solution ensures the accuracy, professionalism, and personalization of the feedback information, while improving the fluency and experience of user communication. This mechanism effectively solves the "illusion" problem that may occur in large language models, that is, it avoids the model generating a quote when the real-time interface is not called or claiming that the product is in stock when the interface has no results, because all core factual information comes from the accurate processing and query results of the front-end module.

[0066] For example, suppose a target user submits a price inquiry request for steel. The information processing module receives and parses the request, writing the inquiry information into a status object. Simultaneously, it loads the target user's profile, inquiry status, and memory information from the memory cache, such as the user's past preference for a specific steel mill's products and frequent inquiries about delivery cycles. This information is compiled into inquiry memory information and written into the status object. The slot extraction module determines the target user's intent to inquire based on the status object and extracts standard slot information, such as "Q235B, 3.5*1500mm, Liuzhou Steel". The information query module calls a product query tool to retrieve the corresponding product transaction information, such as "Price: 4500 yuan / ton, Inventory: 100 tons, Warehouse: ABCD Warehouse, Estimated Delivery Time: Within 3 days". At this point, the information feedback module begins its work. It reads the aforementioned memory information from the status object and determines the preset steel trading industry language knowledge base. Based on this information, the feedback module generates initial inquiry feedback, such as: "Hello, according to your needs, the current price of Q235B 3.5*1500mm Liuzhou Steel is 4500 yuan / ton, with 100 tons in stock, located in warehouses ABCD, and delivery is expected within 3 days." This initial feedback is accurate and professional, but may sound slightly stiff. Subsequently, the system uses a large language model to process this initial inquiry feedback into more conversational language, transforming it into a more natural expression, such as: "Dear customer, hello! We have the Q235B 3.5*1500mm Liuzhou Steel you inquired about! The current price is 4500 yuan / ton, with ample stock, 100 tons in warehouses ABCD. We can arrange delivery for you in as little as 3 days! How does this sound?" Finally, this conversational target inquiry feedback is sent to the target user.

[0067] In summary, the above processing effectively solves the problems of traditional large language models lacking relevance, professionalism, and accuracy in generating inquiry feedback. By introducing memory information, the system can provide personalized feedback based on users' historical preferences and behavioral habits, enhancing the user experience. The use of a script knowledge base ensures that the feedback content conforms to industry standards and business logic, improving its professionalism.

[0068] Furthermore, large language models may exhibit illusions or excessive embellishment, causing the generated feedback information to deviate from the original business logic or factual accuracy, lacking a quality verification and fallback mechanism for the model's output. To address this, in this embodiment, the information feedback module is further configured to compare the target inquiry feedback information with the initial inquiry feedback information. If the comparison result determines that the target inquiry feedback information does not meet the feedback conditions, the initial inquiry feedback information is used as the target inquiry feedback information and sent to the target user.

[0069] Specifically, the target inquiry feedback information refers to the information processed into spoken language using a large language model, while the initial inquiry feedback information is the raw, unedited information generated based on the inquiry intent, standard slot information, and product transaction information. Comparing these two types of information aims to verify whether the target inquiry feedback information meets preset feedback conditions in terms of content accuracy, completeness of key information, and compliance with business logic. Specifically, the comparison operation can be implemented in several ways. For example, key entities (such as product name, specifications, price, quantity, etc.) and keywords can be extracted from the target and initial inquiry feedback information, and then the matching degree, quantity, and accuracy of these entities and keywords can be compared to check whether core information such as product specifications and prices are consistent. Alternatively, natural language processing techniques, such as word vectors, sentence vectors, or semantic similarity models, can be used to calculate the semantic similarity between the two; if the similarity is lower than a preset threshold, a deviation is considered to exist. Furthermore, if the initial inquiry feedback information is generated based on structured data, it is also possible to attempt to extract structured information from the target inquiry feedback information and compare it with the original structured information to check whether key fields have been tampered with or omitted. After comparison, the system determines whether the target inquiry feedback information fails to meet the feedback conditions based on the comparison results. Feedback conditions are preset rules or standards used to judge whether the target inquiry feedback information is qualified. For example, an accuracy threshold can be set; if the comparison finds that the proportion of key product information (such as material, specifications, and price) in the target inquiry feedback information that is inconsistent with the initial inquiry feedback information exceeds a certain percentage, then the feedback conditions are considered not met. A completeness check can also be performed to ensure that the target inquiry feedback information contains all necessary information from the initial inquiry feedback information, such as whether important product parameters or transaction terms are missing. Simultaneously, business logic compliance rules can be preset; for example, if the target inquiry feedback information contains descriptions that do not match actual inventory or prices, or contains promises that do not comply with company policy, then the feedback conditions are considered not met. When it is determined that the target inquiry feedback information does not meet the feedback conditions, the information feedback module will activate a fallback strategy, using the initial inquiry feedback information as the target inquiry feedback information and sending it to the target user. This means that when the large language model output is unqualified, the system will revert to the original information generated based on business logic and not over-processed by the model. The notification can be sent to the target user via the API interface of a pre-set instant messaging platform (such as WeChat, WeChat Work, DingTalk, etc.), or via email or SMS service in specific scenarios as an auxiliary or alternative notification method.

[0070] Based on this, a feedback quality verification mechanism is introduced into the information feedback module to ensure the reliability and accuracy of inquiry responses. Specifically, the information feedback module first generates initial inquiry feedback information based on a preset knowledge base of dialogue, memorized information, inquiry intent, standard slot information, and product transaction information. Then, a large language model is used to process this initial inquiry feedback information into conversational language to obtain the target inquiry feedback information, aiming to improve user interaction and communication efficiency. However, to address potential "illusions" or over-embellishment issues that may arise from the large language model, the information feedback module further compares the target inquiry feedback information processed by the large language model with the original initial inquiry feedback information. This comparison aims to verify whether the target inquiry feedback information meets preset feedback conditions in terms of content accuracy, completeness of key information, and compliance with business logic. If the comparison results show that the target inquiry feedback information does not meet these feedback conditions, i.e., there is a potential error or deviation, the system will activate a fallback strategy. In this scenario, the feedback module discards the potentially problematic target query feedback generated by the large language model and instead directly uses the initial, unprocessed query feedback as the final target query feedback, sending it to the target user. This mechanism ensures that even if the output of the large language model has issues, the user receives an accurate and reliable query response, effectively avoiding business risks caused by the instability of the large language model. This enhances the user experience while guaranteeing the rigor and accuracy of information transmission.

[0071] For example, suppose a target user submits a price inquiry request. The information processing module responds to this request, determines the price inquiry information and the price inquiry memory information, and writes them into a status object. Subsequently, the slot extraction module determines the target user's intent to inquire based on the status object, reads the target price inquiry information from it, extracts the slot, and finally determines the standard slot information, such as "Material: Q235B, Specification: 3.5*1500". The information query module calls the product query tool to find the product information corresponding to the standard slot information, such as "Q235B 3.5*1500 steel, Price: 5000 yuan / ton, Inventory: 100 tons". Based on this, the information feedback module generates initial inquiry feedback based on the inquiry intent, standard slot information, and commodity transaction information, combined with memory information and a knowledge base of conversational phrases. For example: "Hello, the Q235B 3.5*1500 steel you inquired about is currently priced at 5000 yuan / ton, and we have ample stock." Next, the information feedback module uses a large language model to process this initial inquiry feedback into conversational language, generating target inquiry feedback, such as: "Dear customer, the Q235B 3.5*1500 steel you inquired about is currently in stock! The price is very favorable, only 4500 yuan / ton, with large inventory, and we can ship immediately!" Subsequently, the information feedback module compares this target inquiry feedback with the initial inquiry feedback. During the comparison, the system can detect an inconsistency between the price "4500 yuan / ton" mentioned in the target inquiry feedback and "5000 yuan / ton" in the initial inquiry feedback. Based on preset feedback conditions (e.g., price information must precisely match the actual query results), the system determines that the target inquiry feedback information does not meet the feedback conditions. Therefore, the information feedback module will activate a fallback mechanism, abandoning the conversational information generated by the large language model, and instead sending the original initial inquiry feedback information, namely, "Hello, the current price of Q235B 3.5*1500 steel you inquired about is 5000 yuan / ton, and the inventory is sufficient," as the final target inquiry feedback information to the target user.

[0072] In summary, the above processing effectively solves the problems of illusion or over-embellishment that may occur when large language models generate inquiry feedback information. By comparing the target inquiry feedback information processed by the large language model with the original initial inquiry feedback information, deviations or errors in the model output can be identified in a timely manner. In practical applications, customer inquiry needs, user characteristics, and business environments are dynamically changing in trade scenarios. Relying solely on static memory information or lacking a mechanism for continuous learning and updating historical interaction data can lead to the system's inability to accurately capture long-term customer preferences and business evolution. This makes it difficult to provide personalized and accurate feedback in subsequent inquiry cycles, resulting in a contradiction between the system's memory lag and the real-time requirements of the business. To address this, in this embodiment, the information generation system further includes a memory management module. This module is used to acquire new user characteristics, new memory information, and audit information associated with the inquiry request, and to update the memory cache using these information. The updated memory cache is then used to construct the inquiry memory information for the next inquiry cycle.

[0073] Specifically, the memory management module is a core component of the information generation system. Its main function is to dynamically maintain and update the system's stored memory cache. This module can be an independent software service unit, such as a microservice, specifically handling the logic of acquiring, processing, and storing memory data; or it can be a sub-functional module of the information processing module, implementing memory management through internal function calls. Its role is to ensure that the system can continuously learn and adapt to changes in user behavior and the business environment, providing the latest and most accurate memory information for subsequent inquiry processing. This module can identify and extract new data valuable for user profiles, historical preferences, and system operating status from the currently processed inquiry requests and their related interactions. New user characteristics can include new purchasing habits, preferences (such as a preference for specific brands, specifications, or delivery methods), or changes in their identity attributes exhibited during the current interaction. This can be obtained through semantic analysis of the user's current inquiry text, the transaction options selected by the user, or mining other behavioral data of the user in the system. New memory information refers to business data generated during the current inquiry process that can be used for future reference, such as newly proposed specific needs by the user, product information successfully matched by the system, or user feedback on a quotation. This information can be extracted from the inquiry information received by the information processing module, the standard slot information determined by the slot extraction module, and the commodity transaction information obtained by the information query module. Audit information records key operation logs, interface call details, decision paths, and any abnormalities or manual interventions during the system's processing of this inquiry request. For example, it can record the original request received by the information processing module, the extraction results from the slot extraction module, the query parameters and return results from the information query module, and the feedback content generated by the information feedback module. This information can be captured in real time through the system's internal logging mechanism or event listeners. The memory management module integrates newly acquired data into the existing memory cache to maintain its timeliness and accuracy. One update method is incremental update, which only adds new data to the memory cache or overwrites the corresponding old data in the memory cache. Another update method is rule-based update, which analyzes and evaluates the new data according to preset business rules or machine learning models before deciding how to update the memory cache. The memory cache can be a structured database, such as a relational database or a NoSQL database, used to store user profiles, historical inquiry records, preference settings, etc. The updated memory cache will serve as a crucial data source for the information processing module when handling subsequent user inquiry requests. When a user initiates a new inquiry request, the information processing module can load the user profile, inquiry status, and memory information corresponding to the target user from this latest-state memory cache, thereby constructing a more comprehensive and accurate inquiry memory information.

[0074] Based on this, after the information generation system processes an inquiry request, the memory management module does not simply end the session, but proactively acquires and stores valuable new data throughout the entire lifecycle of the inquiry interaction. Specifically, it collects new user characteristics associated with the current inquiry request, such as new preferences or habits exhibited by the user during the session; simultaneously, it also acquires new memory information, which may include the user's explicitly expressed needs in the inquiry, information on products successfully matched by the system, etc. In addition, audit information during system operation, such as interface calls, decision paths, and manual interventions, is also acquired. After acquiring this new data, the memory management module uses this information to dynamically update the system's internal memory cache. This update is not a simple overwrite, but an intelligent integration, ensuring that the memory cache can reflect the latest user status and system operation in real time. In this way, the user profiles and historical inquiry records stored by the system are always up-to-date. This dynamic update mechanism, working in conjunction with the information processing module, constitutes the core of the system's continuous learning and adaptability. When a target user initiates a new inquiry request in the future, the information processing module no longer loads a static, potentially outdated memory cache. Instead, it loads the most accurate user profile, inquiry status, and memory information from this updated memory cache. Based on this latest information, the information processing module can construct more precise inquiry memory information and write it into the status object. This allows subsequent modules such as slot extraction, information query, and information feedback to process data that is closer to the user's current situation, significantly improving the accuracy and personalization of the entire inquiry processing flow. By incorporating audit information into the updates, the system can not only learn from successful experiences but also identify and correct errors, forming a self-optimizing, continuously iterating closed loop that effectively resolves the contradiction between memory lag and the real-time requirements of business operations.

[0075] For example, after the information feedback module sends the target inquiry feedback information to the target user, it sends a completion notification to the memory management module, along with the complete context data of this inquiry session. Upon receiving the notification, the memory management module initiates the data extraction process. It can identify new steel product categories mentioned by the user in this inquiry (e.g., previously only inquiring about regular coils, now inquiring about patterned steel plates), which constitutes a new user characteristic. Simultaneously, if the user explicitly expresses a preference for a particular warehouse (e.g., "send to warehouse XX") or a dislike for a particular steel mill's products (e.g., "don't want XX steel mill's"), this information will also be extracted as new memory information. Furthermore, during system processing, operations such as the information query module calling the product query tool to retrieve product information, the information feedback module using a large language model for conversational processing, and any intervention or correction by human customer service will be recorded as audit information. This includes recording the time, parameters, return results, and time taken for API calls, as well as the original and final modified text of the feedback generated by the large language model. After data extraction is complete, the memory management module uses this new data to update the memory cache. The memory cache can consist of multiple data tables, such as: a user memory table to store long-term customer preferences, such as commonly used materials (Q235B), specifications (standard rolls), delivery methods (tax included, self-pickup), and preferred warehouses; a pending inquiry status table to record slot information that is not yet fully clarified in multiple rounds of inquiries; a long-term memory table to store historical purchasing habits, error correction information, and success stories; and an audit event table to record detailed information such as the execution of each inquiry process node, interface calls, and security rollbacks, ensuring process traceability. Through these updates, the memory management module ensures that the memory cache is always up-to-date. When a target user initiates an inquiry request again, the information processing module will load the latest user profile and memory information from this updated memory cache, thereby more accurately understanding the user's intent and generating inquiry feedback that better meets the user's needs.

[0076] In summary, the above-described approach effectively solves the problem of system memory lag caused by the static nature of memory information and the lack of a continuous learning mechanism in traditional inquiry systems for dynamic business scenarios such as steel trading. The introduction of the memory management module enables the system to capture and store new user characteristics, new memory information, and system audit information displayed in each interaction in real time. This dynamic acquisition and update mechanism ensures that the memory cache maintained by the system is always up-to-date, allowing the information processing module to construct the inquiry context based on more accurate and personalized memory information in subsequent inquiry cycles.

[0077] Corresponding to the above system embodiments, this specification also provides embodiments of information generation methods. Figure 3A flowchart illustrating an information generation method according to one embodiment of this specification is shown. Figure 3 As shown, the method includes: Step S302: In response to the inquiry request submitted by the target user, determine the inquiry information and the inquiry memory information, and write the inquiry information and the inquiry memory information into the status object.

[0078] Step S304: If the target user is determined to be inquiring about prices based on the status object, the target inquiry information is read from the status object, the slot is extracted from the target inquiry information, and the standard slot information is determined based on the slot extraction result.

[0079] Step S306: Call the product query tool to query the product information corresponding to the standard slot information, and determine the product transaction information based on the product information.

[0080] Step S308: Generate initial inquiry feedback information based on the inquiry intent, the standard slot information, and the commodity transaction information; update the initial inquiry feedback information to target inquiry feedback information using a large language model; and send the target inquiry feedback information to the target user.

[0081] In one optional embodiment, a price inquiry request submitted by a target user is received, the price inquiry request is parsed to obtain price inquiry information, and the price inquiry information is written into a status object; the user profile, price inquiry status, and memory information corresponding to the target user are loaded into a memory cache, price inquiry memory information is composed based on the user profile, the price inquiry status, and the memory information, and the price inquiry memory information is written into the status object.

[0082] In one optional embodiment, the information recorded in the state object is subjected to intent recognition according to a preset intent recognition rule and intent classification model, and an intent tag is determined based on the intent recognition result; if the intent tag is an inquiry intent tag, the target user is determined to be an inquiry intent; if the intent tag is not an inquiry intent tag, an inquiry communication message is sent to the target user according to a preset response strategy.

[0083] In one optional embodiment, the target inquiry information is processed to extract slots to obtain initial slot information; the initial slot information is normalized to obtain intermediate slot information; and the intermediate slot information is supplemented to obtain standard slot information.

[0084] In an optional embodiment, if the target inquiry information does not meet the inquiry feedback conditions, a supplementary inquiry request is sent to the target user, and supplementary inquiry information is received from the target user in response to the supplementary inquiry request, wherein the supplementary inquiry request is determined based on the session context and the pending follow-up status; slot extraction is performed on the supplementary inquiry information to obtain supplementary slot information; the supplementary slot information is merged with the slot extraction result corresponding to the target inquiry information to obtain standard slot information.

[0085] In one optional embodiment, when the standard slot information corresponds to multiple product specifications, a product query tool is invoked to query the product information corresponding to the standard slot information according to a query strategy, and product transaction information is determined based on the product information; wherein, the information query module is configured with control rules, which are used to provide control information to the target user when product information is not obtained.

[0086] In one optional embodiment, the memory information is read from the state object, and a preset dialogue knowledge base is determined; initial inquiry feedback information is generated based on the dialogue knowledge base, the memory information, the inquiry intent, the standard slot information, and the commodity transaction information; the initial inquiry feedback information is processed into conversational language using a large language model to obtain target inquiry feedback information.

[0087] In one optional embodiment, the target inquiry feedback information and the initial inquiry feedback information are compared. If the target inquiry feedback information does not meet the feedback conditions based on the comparison result, the initial inquiry feedback information is used as the target inquiry feedback information and sent to the target user.

[0088] In one optional embodiment, new user characteristics, new memory information, and audit information associated with the inquiry request are obtained, and the memory cache is updated using the new user characteristics, the new memory information, and the audit information; wherein, the updated memory cache is used to construct the inquiry memory information for the next inquiry cycle.

[0089] The following is in conjunction with the appendix Figure 4 Taking the application of the information generation system provided in this specification in a building materials trading scenario as an example, the information generation system will be further explained. Figure 4 This specification illustrates a flowchart of the processing procedure of an information generation system according to an embodiment, which specifically includes the following steps.

[0090] In step S402, the information processing module receives the inquiry request submitted by the target user, parses the inquiry request to obtain the inquiry information, and writes the inquiry information into the status object; loads the user profile, inquiry status and memory information corresponding to the target user into the memory cache, assembles the inquiry memory information based on the user profile, inquiry status and memory information, and writes the inquiry memory information into the status object.

[0091] In step S404, the intent recognition module performs intent recognition on the information recorded in the state object according to the preset intent recognition rules and intent classification model, and determines the intent label based on the intent recognition result; if the intent label is an inquiry intent label, the target user is determined to be an inquiry intent; if the intent label is not an inquiry intent label, an inquiry communication message is sent to the target user according to the preset response strategy.

[0092] Step S406: The slot extraction module extracts slots from the target inquiry information to obtain initial slot information; the initial slot information is normalized to obtain intermediate slot information.

[0093] In step S408, if the target inquiry information does not meet the inquiry feedback conditions, the information merging module sends an inquiry supplement request to the target user and receives supplementary inquiry information from the target user in response to the inquiry supplement request. The inquiry supplement request is determined based on the session context and the pending inquiry status. The module extracts slots from the supplementary inquiry information to obtain supplementary slot information. The module then merges the supplementary slot information with the intermediate slot information to obtain standard slot information.

[0094] In step S410, the information query module calls the product query tool to query the product information corresponding to the standard slot information, and determines the product transaction information based on the product information.

[0095] In step S412, the information feedback module reads the memory information from the state object and determines the preset script knowledge base; it generates initial inquiry feedback information based on the script knowledge base, memory information, inquiry intent, standard slot information, and commodity transaction information; it uses a large language model to process the initial inquiry feedback information into conversational language to obtain target inquiry feedback information, and sends the target inquiry feedback information to the target user.

[0096] The information generation system provided in this embodiment responds to a price inquiry request submitted by a target user through an information processing module, determining price inquiry information and price inquiry memory information, and writing the price inquiry information and price inquiry memory information into a state object. Then, using a slot extraction module, if the target user is determined to have a price inquiry intent based on the state object, the system reads the target price inquiry information from the state object, performs slot extraction on the target price inquiry information, and determines standard slot information based on the slot extraction result. On this basis, an information query module calls a product query tool to query the product information corresponding to the standard slot information, and determines the product transaction information based on the product information. Finally, an information feedback module generates initial price inquiry feedback information based on the price inquiry intent, the standard slot information, and the product transaction information. Using a large language model, the initial price inquiry feedback information is updated to target price inquiry feedback information, and the target price inquiry feedback information is sent to the target user. This system integrates inquiry information with stored data through an information processing module, achieves accurate slot identification and completion through a slot extraction module, ensures the real-time and accuracy of commodity transaction information through an information query module, and generates reliable feedback based on a large language model through an information feedback module. This effectively solves the problems of difficult processing of fragmented inquiries, missing session context, and unreliable feedback information in trade. It can efficiently process fragmented inquiry requests in trade, accurately extract key slot information, maintain session context to avoid repeated inquiries, and generate reliable feedback information based on real-time queries, thereby improving the automation level of the inquiry process and the reliability of transactions.

[0097] Corresponding to the above method embodiments, this specification also provides embodiments of an information generation apparatus. Figure 5 A schematic diagram of an information generation apparatus according to one embodiment of this specification is shown. Figure 5 As shown, the device includes: The determining unit 502 is configured to determine inquiry information and inquiry memory information in response to an inquiry request submitted by a target user, and write the inquiry information and the inquiry memory information into a status object; Extraction unit 504 is configured to, when the target user is determined to be inquiring about pricing based on the status object, read target pricing information from the status object, extract slots from the target pricing information, and determine standard slot information based on the slot extraction result; The query unit 506 is configured to call a product query tool to query the product information corresponding to the standard slot information, and determine the product transaction information based on the product information; Feedback unit 508 is configured to generate initial inquiry feedback information based on the inquiry intent, the standard slot information, and the commodity transaction information, update the initial inquiry feedback information to target inquiry feedback information using a large language model, and send the target inquiry feedback information to the target user.

[0098] In one optional embodiment, a price inquiry request submitted by a target user is received, the price inquiry request is parsed to obtain price inquiry information, and the price inquiry information is written into a status object; the user profile, price inquiry status, and memory information corresponding to the target user are loaded into a memory cache, price inquiry memory information is composed based on the user profile, the price inquiry status, and the memory information, and the price inquiry memory information is written into the status object.

[0099] In one optional embodiment, the information recorded in the state object is subjected to intent recognition according to a preset intent recognition rule and intent classification model, and an intent tag is determined based on the intent recognition result; if the intent tag is an inquiry intent tag, the target user is determined to be an inquiry intent; if the intent tag is not an inquiry intent tag, an inquiry communication message is sent to the target user according to a preset response strategy.

[0100] In one optional embodiment, the target inquiry information is processed to extract slots to obtain initial slot information; the initial slot information is normalized to obtain intermediate slot information; and the intermediate slot information is supplemented to obtain standard slot information.

[0101] In an optional embodiment, if the target inquiry information does not meet the inquiry feedback conditions, a supplementary inquiry request is sent to the target user, and supplementary inquiry information is received from the target user in response to the supplementary inquiry request, wherein the supplementary inquiry request is determined based on the session context and the pending follow-up status; slot extraction is performed on the supplementary inquiry information to obtain supplementary slot information; the supplementary slot information is merged with the slot extraction result corresponding to the target inquiry information to obtain standard slot information.

[0102] In one optional embodiment, when the standard slot information corresponds to multiple product specifications, a product query tool is invoked to query the product information corresponding to the standard slot information according to a query strategy, and product transaction information is determined based on the product information; wherein, the information query module is configured with control rules, which are used to provide control information to the target user when product information is not obtained.

[0103] In one optional embodiment, the memory information is read from the state object, and a preset dialogue knowledge base is determined; initial inquiry feedback information is generated based on the dialogue knowledge base, the memory information, the inquiry intent, the standard slot information, and the commodity transaction information; the initial inquiry feedback information is processed into conversational language using a large language model to obtain target inquiry feedback information.

[0104] In one optional embodiment, the target inquiry feedback information and the initial inquiry feedback information are compared. If the target inquiry feedback information does not meet the feedback conditions based on the comparison result, the initial inquiry feedback information is used as the target inquiry feedback information and sent to the target user.

[0105] The above is an illustrative scheme of an information generation device according to this embodiment. It should be noted that the technical solution of this information generation device and the technical solution of the information generation method described above belong to the same concept. For details not described in detail in the technical solution of the information generation device, please refer to the description of the technical solution of the information generation method described above.

[0106] Figure 6 A structural block diagram of a computing device 600 according to one embodiment of this specification is shown. The components of the computing device 600 include, but are not limited to, a memory 610 and a processor 620. The processor 620 is connected to the memory 610 via a bus 630, and a database 650 is used to store data.

[0107] The computing device 600 also includes an access device 640, which enables the computing device 600 to communicate via one or more networks 660. Examples of these networks include Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or combinations of communication networks such as the Internet. The access device 640 may include one or more of any type of wired or wireless network interface (e.g., a network interface card (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) wireless interface, a Wi-MAX (Worldwide Interoperability for Microwave Access) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, or a Near Field Communication (NFC) interface.

[0108] In one embodiment of this specification, the above-described components of the computing device 600 and Figure 6 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 6The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0109] The computing device 600 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or personal computers (PCs). The computing device 600 can also be a mobile or stationary server.

[0110] The processor 620 is configured to execute the following computer-executable instructions, which, when executed by the processor, implement the steps of the above-described information generation method.

[0111] The above is an illustrative scheme of a computing device according to this embodiment. It should be noted that the technical solution of this computing device and the technical solution of the information generation method described above belong to the same concept. For details not described in detail in the technical solution of the computing device, please refer to the description of the technical solution of the information generation method described above.

[0112] An embodiment of this specification also provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the above-described information generation method.

[0113] The above is an illustrative scheme of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium and the technical solution of the information generation method described above belong to the same concept. For details not described in detail in the technical solution of the storage medium, please refer to the description of the technical solution of the information generation method described above.

[0114] An embodiment of this specification also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described information generation method.

[0115] The above is an illustrative scheme of a computer program product according to this embodiment. It should be noted that the technical solution of this computer program product and the technical solution of the information generation method described above belong to the same concept. For details not described in detail in the technical solution of the computer program product, please refer to the description of the technical solution of the information generation method described above.

[0116] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0117] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or certain intermediate forms. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium may be appropriately added or removed according to the requirements of patent practice. For example, in some regions, according to patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0118] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0119] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0120] The preferred embodiments disclosed above are merely illustrative of this specification. Optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described in this specification. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification.

Claims

1. An information generation system, characterized in that, include: The information processing module is used to determine the inquiry information and inquiry memory information in response to the inquiry request submitted by the target user, and write the inquiry information and the inquiry memory information into the status object; The slot extraction module is used to read target inquiry information from the status object when the target user is determined to have an inquiry intention based on the status object, extract slots from the target inquiry information, and determine standard slot information based on the slot extraction results. The information query module is used to call the product query tool to query the product information corresponding to the standard slot information, and determine the product transaction information based on the product information; The information feedback module is used to generate initial inquiry feedback information based on the inquiry intent, the standard slot information, and the commodity transaction information, update the initial inquiry feedback information to target inquiry feedback information using a large language model, and send the target inquiry feedback information to the target user.

2. The information generation system according to claim 1, characterized in that, The information processing module is further configured to receive a price inquiry request submitted by a target user, parse the price inquiry request to obtain price inquiry information, and write the price inquiry information into a status object; load the user profile, price inquiry status, and memory information corresponding to the target user into a memory cache, compose price inquiry memory information based on the user profile, the price inquiry status, and the memory information, and write the price inquiry memory information into the status object.

3. The information generation system according to claim 1, characterized in that, The information generation system further includes an intent recognition module; the intent recognition module is used to perform intent recognition on the information recorded in the state object according to preset intent recognition rules and intent classification model, and determine intent labels based on the intent recognition results; If the intent tag is an inquiry intent tag, then the target user is determined to have an inquiry intent. If the intent tag is not a price inquiry intent tag, a price inquiry communication message is sent to the target user according to a preset response strategy.

4. The information generation system according to claim 1, characterized in that, The slot extraction module is further configured to extract slots from the target inquiry information to obtain initial slot information; normalize the initial slot information to obtain intermediate slot information; and supplement the intermediate slot information to obtain standard slot information.

5. The information generation system according to claim 1 or 4, characterized in that, The information generation system further includes an information merging module; the information merging module is used to send a supplementary inquiry request to the target user when the target inquiry information does not meet the inquiry feedback conditions, and to receive supplementary inquiry information from the target user in response to the supplementary inquiry request, wherein the supplementary inquiry request is determined based on the session context and the pending follow-up status; to extract slots from the supplementary inquiry information to obtain supplementary slot information; and to merge the supplementary slot information with the slot extraction result corresponding to the target inquiry information to obtain standard slot information.

6. The information generation system according to claim 1, characterized in that, The information query module is further configured to, when the standard slot information corresponds to multiple product specifications, invoke a product query tool to query the product information corresponding to the standard slot information according to a query strategy, and determine the product transaction information based on the product information; wherein, the information query module is configured with control rules, which are used to provide control information to the target user when product information is not obtained.

7. The information generation system according to claim 2, characterized in that, The information feedback module is also used to read the memory information from the state object and determine the preset dialogue knowledge base; generate initial inquiry feedback information based on the dialogue knowledge base, the memory information, the inquiry intent, the standard slot information and the commodity transaction information; and use a large language model to process the initial inquiry feedback information into colloquial language to obtain target inquiry feedback information.

8. The information generation system according to claim 7, characterized in that, The information feedback module is further configured to compare the target inquiry feedback information with the initial inquiry feedback information, and if the target inquiry feedback information does not meet the feedback conditions based on the comparison result, the initial inquiry feedback information is used as the target inquiry feedback information and sent to the target user.

9. The information generation system according to claim 2, characterized in that, The information generation system further includes a memory management module; the memory management module is used to obtain new user characteristics, new memory information and audit information associated with the inquiry request, and update the memory cache using the new user characteristics, the new memory information and the audit information; wherein, the updated memory cache is used to construct the inquiry memory information for the next inquiry cycle.

10. An information generation method, characterized in that, include: In response to a request for quotes submitted by a target user, determine the quote information and quote memory information, and write the quote information and quote memory information into a status object; If the target user is determined to be inquiring about pricing based on the status object, the target pricing information is read from the status object, the target pricing information is extracted into slots, and the standard slot information is determined based on the slot extraction result. Use the product query tool to query the product information corresponding to the standard slot information, and determine the product transaction information based on the product information; An initial inquiry feedback is generated based on the inquiry intent, the standard slot information, and the commodity transaction information. The initial inquiry feedback is then updated to a target inquiry feedback using a large language model, and the target inquiry feedback is sent to the target user.

11. An information generation device, characterized in that, include: The determining unit is configured to determine inquiry information and inquiry memory information in response to an inquiry request submitted by a target user, and write the inquiry information and the inquiry memory information into a status object; The extraction unit is configured to, when the target user is determined to have an inquiry intent based on the status object, read the target inquiry information from the status object, extract the slot from the target inquiry information, and determine the standard slot information based on the slot extraction result; The query unit is configured to call a product query tool to query the product information corresponding to the standard slot information, and determine the product transaction information based on the product information; The feedback unit is configured to generate initial inquiry feedback information based on the inquiry intent, the standard slot information, and the commodity transaction information, update the initial inquiry feedback information to target inquiry feedback information using a large language model, and send the target inquiry feedback information to the target user.

12. A computing device, characterized in that, include: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the method of claim 10.

13. A computer-readable storage medium, characterized in that, It stores computer-executable instructions that, when executed by a processor, implement the steps of the method of claim 10.

14. A computer program product, characterized in that, Includes a computer program or instructions that, when executed by a processor, implement the steps of the method of claim 10.

Images