Text processing method and device based on large language model, equipment and storage medium

Abstract

The invention provides a text processing method and device based on a large language model, equipment and a storage medium, relates to the field of artificial intelligence, in particular to the technical fields of NLP, the large language model, deep learning and the like, and can be used for the application fields of generative search, document intelligent editing, intelligent assistants, intelligent e-commerce and the like. The specific implementation scheme is as follows: performing first classification on an input text to obtain a classification result; wherein the classification result is used for indicating a processing object of the input text; performing second classification on the input text to obtain a text category; based on the text category, calling a processing object to process the input text to obtain a response text; and replying the input text based on the response text.

Description

Technical Field

[0001] The present disclosure relates to the field of AI (Artificial Intelligence), specifically to technical fields such as NLP (Natural Language Processing), LLM (Large Language Model), deep learning, etc., and can be used in application fields such as generative search, document intelligent editing, intelligent assistants, and intelligent e-commerce, and especially to text processing methods, devices, electronic devices, and storage media based on large language models. Background Art

[0002] With the rapid development of NLP and AI technologies, computers have been able to understand and generate natural language text, which makes it possible to automatically process user input text (such as query text, user questions, etc.). For example, automatic text processing technology can be used to automatically process input text. At present, automatic text processing technology has been widely used in search engines, intelligent customer service, chatbots and other fields, and can achieve rapid response and processing of user input text. Summary of the invention

[0003] The present disclosure provides a method, apparatus, device and storage medium for text processing based on a large language model.

[0004] According to one aspect of the present disclosure, a text processing method based on a large language model is provided, comprising:

[0005] Performing a first classification on the input text to obtain a classification result; wherein the classification result is used to indicate a processing object of the input text;

[0006] Performing a second classification on the input text to obtain a text category;

[0007] Based on the text category, calling the processing object to process the input text to obtain a response text;

[0008] The input text is replied based on the response text.

[0009] According to another aspect of the present disclosure, a text processing device based on a large language model is provided, comprising:

[0010] A first classification module, used to perform a first classification on the input text to obtain a classification result; wherein the classification result is used to indicate a processing object of the input text;

[0011] A second classification module, used for performing a second classification on the input text to obtain a text category;

[0012] A calling module, used for calling the processing object to process the input text based on the text category to obtain a response text;

[0013] A reply module is used to reply to the input text based on the response text.

[0014] According to another aspect of the present disclosure, there is provided an electronic device, comprising:

[0015] at least one processor; and

[0016] a memory communicatively connected to the at least one processor; wherein,

[0017] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the text processing method based on the large language model proposed in the above aspect of the present disclosure.

[0018] According to another aspect of the present disclosure, a non-transitory computer-readable storage medium of computer instructions is provided, wherein the computer instructions are used to enable the computer to execute the text processing method based on the large language model proposed in the above aspect of the present disclosure.

[0019] According to another aspect of the present disclosure, a computer program product is provided, including a computer program, which, when executed by a processor, implements the text processing method based on a large language model proposed in the above aspect of the present disclosure.

[0020] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present disclosure, nor is it intended to limit the scope of the present disclosure. Other features of the present disclosure will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The accompanying drawings are used to better understand the present solution and do not constitute a limitation of the present disclosure.

[0022] Figure 1 A flowchart of a text processing method based on a large language model provided in the first embodiment of the present disclosure;

[0023] Figure 2 A flowchart of a text processing method based on a large language model provided in the second embodiment of the present disclosure;

[0024] Figure 3 A schematic diagram of an assembly method or a combination method of an atomic tool provided in an embodiment of the present disclosure;

[0025] Figure 4A flowchart of a text processing method based on a large language model provided in Embodiment 3 of the present disclosure;

[0026] Figure 5 A flowchart of a text processing method based on a large language model provided in Embodiment 4 of the present disclosure;

[0027] Figure 6 A flowchart of a text processing method based on a large language model provided in Embodiment 5 of the present disclosure;

[0028] Figure 7 A flowchart of a text processing method based on a large language model provided in Embodiment 6 of the present disclosure;

[0029] Figure 8 A flowchart of a text processing method based on a large language model provided in Embodiment 7 of the present disclosure;

[0030] Fig. 9 A flowchart of a text processing method based on a large language model provided in Embodiment 8 of the present disclosure;

[0031] Fig.10 A flowchart of a text processing method based on a large language model provided in Embodiment 9 of the present disclosure;

[0032] Fig.11 A schematic diagram of the structure of the low-code tool platform provided in the embodiment of the present disclosure;

[0033] Fig.12 A schematic diagram of the interaction process between a tool-type intelligent agent and a tool provided in an embodiment of the present disclosure;

[0034] Fig.13 A schematic diagram of the interaction relationship between intelligent agents provided in an embodiment of the present disclosure;

[0035] Fig.14 A schematic diagram of the interaction relationship between atomic tools provided in the embodiment of the present disclosure;

[0036] Fig.15 A schematic diagram of the structure of a text processing device based on a large language model provided in Embodiment 10 of the present disclosure;

[0037] Fig.16 A schematic block diagram of an example electronic device that can be used to implement embodiments of the present disclosure is shown. DETAILED DESCRIPTION

[0038] The following is a description of exemplary embodiments of the present disclosure in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0039] Currently, large models are booming, especially large language models (LLMs), which have shown strong capabilities. In terms of text processing, the industry also has many intelligent processing solutions combined with LLMs, but none of these processing solutions have the characteristics of "stability", "maintainability" and "low cost" at the same time.

[0040] Among them, stability: In the text processing process, each step needs to produce stable and reliable analysis results. However, a typical defect of LLM is the existence of "hallucination", which poses a challenge to the stability of the analysis results.

[0041] Maintainability: When processing an input text, different business parties (i.e. business teams) of the enterprise's distributed business system need to participate. Business parties have the core experience of their business. If these experiences cannot be accumulated by the business parties themselves, then even if the overall intelligent processing system is well implemented, it is not maintainable. This means that when the business system changes, the business parties cannot update the knowledge or experience in the intelligent processing system in a timely manner, making the intelligent processing system no longer effective.

[0042] Low cost: When a new text type of input text appears, it is only necessary to build new tools or new intelligent agents in a low-cost manner, and at the same time, no changes can be made to the original tools or intelligent agents. In other words, the new text type must make the intelligent processing system stronger without any operation on the processing ability of the existing text type.

[0043] In the related art, the following solutions are mainly used to process the input text:

[0044] The first one uses an intelligent agent platform to process input text.

[0045] However, the agent platform is responded by a single agent, and the input text is only distributed to a single agent, which cannot handle complex input text.

[0046] The second method uses multi-agent intelligent operation and maintenance to process input text.

[0047] Although this method uses the interaction mode between multiple agents to process input text, when encountering new forms of input text, the solution logic of the old form of input text may need to be modified, that is, it does not meet the maintainability requirements. At the same time, it does not mention how to enable different business parties to participate in a friendly manner, which does not reflect business friendliness.

[0048] The third method is to use multi-scenario intelligent operation and maintenance to realize the interactive mode between intelligent agents to process input text.

[0049] However, this method uses function_call when calling the tool, and uses the "reflection" mechanism to solve the instability caused by the illusion of the model. At the same time, it does not mention how to make different business parties participate in a friendly manner, which does not reflect business friendliness.

[0050] Therefore, in response to at least one of the above-mentioned problems, the present disclosure proposes a text processing method, apparatus, device and storage medium based on a large language model, so as to simultaneously meet the above-mentioned requirements of "stability", "maintainability" and "low cost" when processing various types of input texts.

[0051] The following describes the text processing method, apparatus, device and storage medium based on a large language model of the embodiment of the present disclosure with reference to the accompanying drawings. Before describing the embodiment of the present disclosure in detail, for ease of understanding, the commonly used technical terms are first introduced:

[0052] LLM is a type of natural language processing model based on deep learning. Its main features are huge model parameters and complex neural network structure, strong language understanding, context perception and language generation capabilities. It can automatically learn useful feature representations from input data and generate relevant text.

[0053] Figure 1 A flowchart of a text processing method based on a large language model provided in Embodiment 1 of the present disclosure.

[0054] The embodiment of the present disclosure takes the large language model-based text processing method as an example in which the large language model-based text processing device is configured. The text processing device can be applied to any electronic device so that the electronic device can perform a text processing function.

[0055] Among them, the electronic device can be any device with computing capabilities, such as a personal computer, a mobile terminal, a server, etc. The mobile terminal can be, for example, a mobile phone, a tablet computer, a personal digital assistant, a wearable device, and other hardware devices with various operating systems, touch screens and / or display screens.

[0056] like Figure 1As shown, the text processing method based on the large language model may include the following steps S101 to S104:

[0057] Step S101, performing a first classification on the input text to obtain a classification result; wherein the classification result is used to indicate a processing object of the input text.

[0058] The input text is input or provided by the user; the input method of the input text may include but is not limited to: touch input (such as sliding, clicking, etc.), keyboard input, voice input, etc.

[0059] The processing objects for processing the input text include but are not limited to: tools, agents, etc. The tools include but are not limited to: script commands, APIs (Application Programming Interface), functions, etc.

[0060] In the embodiment of the present disclosure, a text classification technology may be used to perform a first classification on the input text to obtain a classification result, wherein the classification result is used to indicate a processing object of the input text.

[0061] Step S102, performing a second classification on the input text to obtain a text category.

[0062] The text category may vary depending on the specific application scenario of the input text, business requirements, and characteristics of the text content.

[0063] Exemplarily, when the input text is a question asked by a user, the text categories include, but are not limited to: clear questions, ambiguous questions, complex questions, common questions, emotional questions, etc.

[0064] Exemplarily, when the input text is a query text, the text categories include but are not limited to: functional query (eg, querying the resource usage of a distributed business system), detailed query, summary query, fuzzy query, precise query, etc.

[0065] Exemplarily, in the case where the input text is a fault problem entered in response to a fault in a distributed business system, the text categories include, but are not limited to: query lossy problems (e.g., no search results for the user's input query, irrelevant search results, duplicate results (i.e., there are a large number of duplicate entries in the search results)), long-tail problems, Core problems (core system problems), etc.

[0066] In the embodiment of the present disclosure, the input text may be secondly classified based on the text classification technology to obtain the text category to which the input text belongs.

[0067] Step S103: Based on the text category, the processing object is called to process the input text to obtain a response text.

[0068] In the embodiment of the present disclosure, based on the text category, the processing object indicated by the classification result can be called to perform targeted processing on the input text to obtain a response text.

[0069] Exemplarily, when the input text is a query text, the response text may be a specific query result; when the input text is a user question, the response text may be an answer, solution steps or a solution; when the input text is a fault problem, the response text may be the root cause of the fault.

[0070] Step S104: reply to the input text based on the response text.

[0071] In the embodiments of the present disclosure, a timely reply can be made to the input text based on the response text, so that the user can obtain the response text in time, thereby reducing the waiting time of the user.

[0072] The text processing method based on a large language model in the embodiment of the present disclosure can be implemented for different input texts. Based on the text category to which the input text belongs, a processing object adapted to the input text is called to perform targeted processing on the input text, thereby improving the accuracy and rationality of text processing and improving the user experience.

[0073] It should be noted that in the technical solution of the present disclosure, the collection, storage, use, processing, transmission, provision and disclosure of user personal information are all carried out with the user's consent, and are in compliance with relevant laws and regulations and do not violate public order and good morals.

[0074] In order to clearly explain how any embodiment of the present disclosure processes the input text based on the text category and calls the processing object indicated by the classification result to obtain the response text, the present disclosure also proposes a text processing method based on a large language model.

[0075] Figure 2 A flowchart of a text processing method based on a large language model provided in Embodiment 2 of the present disclosure.

[0076] like Figure 2 As shown, the text processing method based on the large language model may include the following steps S201 to S205:

[0077] Step S201, performing a first classification on the input text to obtain a classification result; wherein the classification result is used to indicate a processing object of the input text.

[0078] Step S202, performing a second classification on the input text to obtain a text category.

[0079] For explanations of steps S201 to S202, reference may be made to the relevant descriptions in any embodiment of the present disclosure and will not be repeated here.

[0080] Step S203, in response to the processing object being a tool, obtaining a target tool adapted to the text category; wherein the target tool is assembled from at least one atomic tool.

[0081] It should be noted that for different text categories, the adapted target tools may be different to improve the pertinence and accuracy of different input text processing.

[0082] In the embodiment of the present disclosure, when the processing object of the input text is a tool, a target tool adapted to the text category can be obtained, wherein the target tool is assembled from at least one atomic tool. Exemplarily, at least one atomic tool adapted to the text category can be assembled by dragging and dropping to obtain the target tool.

[0083] In any embodiment of the present disclosure, the target tool adapted to the text category can be obtained by following steps A to D:

[0084] Step A: In response to the processing object being a tool, query whether there is an assembly tool associated with the text category to which the input text belongs. If so, execute step B; if not, execute steps C to D.

[0085] The assembly tool is obtained by assembling at least one atomic tool associated with a text category in a tool library within a historical period.

[0086] Step B: Use the assembly tool as the target tool.

[0087] In the embodiment of the present disclosure, when there is an assembly tool associated with the text category to which the input text belongs, the assembly tool can be directly used as a target tool adapted to the text category.

[0088] That is to say, the text category to which the input text belongs is an existing known text category, and the execution subject of the present disclosure has processed other texts that belong to the same text category as the input text within a historical period. When processing the other texts, the execution subject obtains at least one atomic tool associated with the same text category from the tool library, assembles the at least one atomic tool to obtain an assembly tool, and uses the assembly tool to process the other texts.

[0089] Furthermore, in order to improve the processing efficiency of subsequent texts, the executing entity may also establish an association relationship between the same text category and the assembly tool. Thus, in the present disclosure, in order to improve the processing efficiency of the input text, the above association relationship may be queried to determine the assembly tool associated with the text category to which the input text belongs, and use the assembly tool as a target tool adapted to the text category to which the input text belongs.

[0090] Step C: Obtain an atomic tool adapted to the text category to which the input text belongs; wherein the adapted atomic tool includes: an atomic tool created based on the input text in response to the first configuration instruction, and / or an atomic tool in a tool library.

[0091] As an example, in the absence of an assembly tool associated with the text category to which the input text belongs, relevant personnel (such as administrators, business personnel, etc.) may specify an atomic tool that is adapted to the text category; wherein the atomic tool that is adapted to the text category includes: manually created atomic tools (i.e., atomic tools created based on the input text in response to a first configuration instruction triggered by relevant personnel), and / or atomic tools that have been created in the tool library.

[0092] Step D: Assemble the atomic tools adapted to the text category to obtain the target tool.

[0093] As a possible implementation method, in response to configuration instructions triggered by relevant personnel, atomic tools adapted to the text category can be assembled to obtain a target tool.

[0094] As an example, in actual application scenarios, there are many types of text categories. In order to support these scenarios, the assembly or combination of tools may include: Figure 3 (a) Figure 3 (b) and Figure 3 (c) shows three types, among which, Figure 3 (a) refers to the scenario of serial assembly of multiple atomic tools. Figure 3 (b) refers to the scenario of selecting which atomic tool to trigger based on the conditions. Figure 3 (c) refers to the scenario where multiple atomic tools are assembled in parallel. In the present disclosure, relevant personnel can complete the assembly of atomic tools by "dragging and dropping" the UI (User Interface).

[0095] Among them, the "data whiteboard" can be regarded as a central data storage or message queue for transferring data between different atomic tools. The data whiteboard acts as a neutral and accessible data exchange center, allowing any atomic tool to read or write data on it.

[0096] In summary, when the text category to which the input text belongs is a known or existing text category, directly based on the association relationship between the text category and the assembly tool, the target tool associated with or adapted to the text category to which the input text belongs is determined from the assembly tool. This can not only realize rapid processing of input texts of known text categories based on existing business experience, thereby improving the processing efficiency of input texts, but also realize tool reuse without creating duplicate tools, thereby saving manpower and time costs. In the case where the text category to which the input text belongs is a newly added text category, a target tool adapted to the newly added text category is configured through configuration, so that the input text of the newly added text category can be processed based on the target tool, which can improve the accuracy and effectiveness of text processing.

[0097] Step S204, calling the target tool to process the input text to obtain a response text.

[0098] In the disclosed embodiment, a target tool may be called to process input text to obtain a response text.

[0099] Step S205: reply to the input text based on the response text.

[0100] For the explanation of step S205, reference may be made to the relevant description in any embodiment of the present disclosure, which will not be repeated here.

[0101] The text processing method based on a large language model in an embodiment of the present disclosure obtains a target tool adapted to the text category and calls the target tool to process the input text when the classification result indicates that the processing object of the input text is a tool, thereby improving the accuracy and effectiveness of text processing.

[0102] In order to clearly explain how to call the target tool to process the input text in any embodiment of the present disclosure, the present disclosure also proposes a text processing method based on a large language model.

[0103] Figure 4 This is a flowchart of a text processing method based on a large language model provided in Embodiment 3 of the present disclosure.

[0104] like Figure 4 As shown, the text processing method based on the large language model may include the following steps S401 to S408:

[0105] Step S401, performing a first classification on the input text to obtain a classification result; wherein the classification result is used to indicate a processing object of the input text.

[0106] Step S402: Perform a second classification on the input text to obtain a text category.

[0107] Step S403, in response to the processing object being a tool, obtaining a plurality of target tools adapted to the text category.

[0108] There are multiple target tools that are compatible with the text category, and each target tool is assembled from at least one atomic tool that is compatible with the text category.

[0109] The explanation of steps S401 to S403 can be found in the relevant description in any embodiment of the present disclosure, and will not be repeated here.

[0110] Step S404, determining a calling sequence among the multiple target tools, and calling the multiple target tools in sequence based on the calling sequence.

[0111] The calling order of multiple target tools adapted to the text category to which the input text belongs may be pre-specified, or may be manually specified by relevant personnel, and the embodiments of the present disclosure do not impose any limitation on this.

[0112] In the embodiment of the present disclosure, multiple target tools may be called in sequence based on the calling order among the multiple target tools.

[0113] Step S405 , in response to the currently called target tool being the first called target tool, the first target tool is called to process the input text to obtain a processing result of the first target tool.

[0114] In the embodiment of the present disclosure, when the currently called target tool is the first called target tool, the first target tool may be called to process the input text to obtain the processing result of the first target tool.

[0115] Step S406, in response to the currently called target tool being a non-first called target tool, the non-first target tool is called, and the processing result of the previously called target tool is processed to obtain the processing result of the non-first target tool.

[0116] In an embodiment of the present disclosure, when the currently called target tool is not the first called target tool, the non-first target tool can be called to process the processing result of the previously called target tool to obtain the processing result of the non-first target tool.

[0117] Step S407, determining the response text according to the processing result of the last called target tool.

[0118] In the disclosed embodiment, the response text for replying to the input text may be determined according to the processing result of the last called target tool.

[0119] Step S408: reply to the input text based on the response text.

[0120] For explanation of step S408, reference may be made to the relevant description in any embodiment of the present disclosure and will not be repeated here.

[0121] The text processing method based on a large language model of the embodiment of the present disclosure calls multiple target tools in sequence to collaboratively process the input text based on the calling order between multiple target tools. This can ensure that the processing of each target tool is based on the processing result of the previous target tool to form a complete logical chain. This sequentiality ensures the consistency and accuracy of text processing.

[0122] In order to clearly explain how any embodiment of the present disclosure processes the input text based on the text category and calls the processing object indicated by the classification result to obtain the response text, the present disclosure also proposes a text processing method based on a large language model.

[0123] Figure 5 A flowchart of a text processing method based on a large language model provided in Embodiment 4 of the present disclosure.

[0124] like Figure 5 As shown, the text processing method based on the large language model may include the following steps S501 to S506:

[0125] Step S501, performing a first classification on the input text to obtain a classification result; wherein the classification result is used to indicate a processing object of the input text.

[0126] Step S502: Perform a second classification on the input text to obtain a text category.

[0127] The explanation of steps S501 to S502 can be found in the relevant description in any embodiment of the present disclosure, and will not be repeated here.

[0128] Step S503, in response to the processing object being an agent, obtain an agent interaction topology map adapted to the text category; wherein the nodes in the agent interaction topology map are used to indicate the agents, and the edges between the nodes are used to indicate the interaction relationship between the agents, and the agents include tool-type agents and / or non-tool-type agents bound to tools.

[0129] The tools include but are not limited to: script commands, APIs, functions, etc.

[0130] In an embodiment of the present disclosure, when the classification result indicates that the processing object of the input text is an agent, an agent interaction topology map adapted to the text category to which the input text belongs can be obtained; wherein the nodes in the agent interaction topology map are used to indicate the agents, and the edges between the nodes are used to indicate the interaction relationship between the agents, wherein the agents include tool-type agents and / or non-tool-type agents bound to tools.

[0131] It should be noted that for different text categories, the adapted intelligent agent interaction topology diagram may be different to improve the pertinence and accuracy of different input text processing.

[0132] Step S504, calling the agents indicated by the nodes in the agent interaction topology diagram in sequence to process the input text.

[0133] In the disclosed embodiment, the agent indicated by each node in the agent interaction topology diagram may be called in sequence to process the input text to obtain the execution result of the agent indicated by each node.

[0134] In any embodiment of the present disclosure, the agents indicated by each node in the agent interaction topology diagram can be called in sequence. When the currently called agent is the agent indicated by the root node in the agent interaction topology diagram, it can be determined whether the agent indicated by the root node is a tool-type agent. If so, the input text can be processed by the agent indicated by the root node (i.e., the tool-type agent) and the tool bound to the agent indicated by the root node to obtain the processing result of the agent indicated by the root node; if not, the input text is directly processed by the agent indicated by the root node (i.e., the non-tool-type agent) to obtain the processing result of the agent indicated by the root node. Therefore, different processing strategies are adopted for tool-type agents and non-tool-type agents to carry out targeted processing of the input text, which can improve the pertinence, rationality and accuracy of agent processing.

[0135] When the currently called agent is the agent indicated by the non-root node in the agent interaction topology diagram, it can be determined whether the agent indicated by the non-root node is a tool-type agent. If so, the processing result of the agent indicated by the parent node of the non-root node is processed through the agent indicated by the non-root node (i.e., the tool-type agent) and the tool bound to the agent indicated by the non-root node to obtain the processing result of the agent indicated by the non-root node; if not, the processing result of the agent indicated by the parent node of the non-root node is directly processed through the agent indicated by the non-root node (i.e., the non-tool-type agent) to obtain the processing result of the agent indicated by the non-root node.

[0136] In summary, the processing results of the agent indicated by the child node in the agent interaction topology diagram are dependent on the processing results of the agent indicated by its parent node. Through clear dependencies or interactions, the logic of the text processing process can be ensured to be clear. Each agent works only after its parent node has completed processing, which reduces error processing caused by sequence errors or data inconsistencies, allowing each agent to operate in the correct context, thereby improving the reliability and accuracy of text processing. In addition, the dependencies between agents can be flexibly configured, which makes it easier for the system to add new agents or modify the dependencies between existing agents when facing new text types or processing requirements, thereby improving the flexibility and scalability of the system.

[0137] Step S505, determining the response text according to the processing result of the agent indicated by the leaf node in the agent interaction topology diagram.

[0138] In the disclosed embodiment, the processing results of the agents indicated by the leaf nodes in the agent interaction topology diagram may be integrated to determine the response text for replying to the input text.

[0139] Step S506: reply to the input text based on the response text.

[0140] For the explanation of step S506, reference may be made to the relevant description in any embodiment of the present disclosure, which will not be repeated here.

[0141] The text processing method based on a large language model in the disclosed embodiment has high accuracy and reliability because the leaf node is located at the very end of the agent interaction topology diagram, and its processing result is often obtained based on the accumulated information and processing logic of all previous nodes (including its parent node). Therefore, the response information used to reply to the input text is determined according to the processing result of the agent indicated by the leaf node, which can improve the accuracy and reliability of the text reply.

[0142] In order to clearly explain how the agents indicated by the nodes in the agent interaction topology diagram are called in sequence to process the input text in the above embodiment, the present disclosure also proposes a text processing method based on a large language model.

[0143] Figure 6 This is a flowchart of a text processing method based on a large language model provided in Embodiment 5 of the present disclosure.

[0144] like Figure 6 As shown, the text processing method based on the large language model may include the following steps S601 to S612:

[0145] Step S601, performing a first classification on the input text to obtain a classification result; wherein the classification result is used to indicate a processing object of the input text.

[0146] Step S602: Perform a second classification on the input text to obtain a text category.

[0147] Step S603, in response to the processing object being an agent, obtaining an agent interaction topology map adapted to the text category.

[0148] Among them, the nodes in the agent interaction topology diagram are used to indicate the agents, and the edges between the nodes are used to indicate the interaction relationship between the agents. The agents include tool-type agents and / or non-tool-type agents bound to tools.

[0149] Step S604, calling the agents indicated by the nodes in the agent interaction topology diagram in sequence.

[0150] For explanations of steps S601 to S604, reference may be made to the relevant descriptions in any embodiment of the present disclosure, and will not be repeated here.

[0151] Step S605, in response to the currently called agent being the agent indicated by the root node in the agent interaction topology diagram, determine whether the agent indicated by the root node is a tool-type agent, if so, execute step S606, if not, execute step S607.

[0152] It should be noted that step S606 and step S607 are two parallel implementation methods. In actual application, only one needs to be executed.

[0153] Step S606, processing the input text through the agent indicated by the root node and the tool bound to the agent indicated by the root node to obtain the processing result of the agent indicated by the root node.

[0154] In an embodiment of the present disclosure, when the currently called agent is the agent indicated by the root node in the agent interaction topology diagram, and the agent indicated by the root node is a tool-type agent, the input text can be collaboratively processed by the agent indicated by the root node and the tool bound to the agent indicated by the root node to obtain the processing result of the agent indicated by the root node.

[0155] In any embodiment of the present disclosure, the following steps A to E may be used to process the input text to obtain the processing result of the agent indicated by the root node:

[0156] Step A: Get the input parameter description and function description associated with the tool bound to the agent indicated by the root node.

[0157] The input parameter description is used to describe the input parameter description of the tool. For example, the input parameter description of the tool can be described in markdown format. For example, the input parameter description of the tool can be marked as input_param_description.

[0158] The function description is used to describe the function of the tool. For example, the function description of the tool can be marked as description.

[0159] As an example, different interfaces may be constructed in the tool, and the input parameter description and function description of the tool may be obtained by calling the above interfaces.

[0160] For example, the following interface can be constructed in the tool:

[0161] Interface name: name

[0162] Input: None.

[0163] Output: Tool name.

[0164] Interface name: description / / Function description of the tool

[0165] Input: None.

[0166] Output: Description of the tool's functionality.

[0167] Interface name: input_param_description / / tool input parameter description

[0168] Input: None.

[0169] Output: Description of tool input parameters in markdown format.

[0170] Interface name: run_with_params / / Tool execution results

[0171] Input: parameter list in json format.

[0172] Output: tool execution results.

[0173] Step B: Determine tool call parameters based on input text, input parameter description and function description.

[0174] In an embodiment of the present disclosure, the input text, input parameter description and function description may be combined to determine the calling parameters of the tool bound to the agent indicated by the root node, which are referred to as tool calling parameters in the present disclosure.

[0175] As a possible implementation method, a large language model may be used to process input text, input parameter description, and function description to obtain tool calling parameters.

[0176] As an example, first, at least one first reference example associated with the agent indicated by the root node can be obtained or queried, wherein the first reference example is used to indicate a conversion method for converting natural language information into input parameters (or parameter list) of a tool bound to the agent indicated by the root node. Thereafter, a first prompt message (Prompt) can be generated based on each first reference example, input parameter description, function description, and input text. For example, each first reference example, input parameter description, function description, and input text can be assembled to obtain the first prompt message, wherein the first prompt message is used to prompt the large language model to perform a calling parameter generation task, and then the large language model can be called to process the first prompt message to obtain the tool calling parameters.

[0177] Exemplarily, each first reference example associated with the agent indicated by the root node can be obtained by calling the agent interface.

[0178] For example, an agent interface may include:

[0179] Interface name: name

[0180] Input: None.

[0181] Output: Agent name.

[0182] Interface name: description / / Functional description of the agent

[0183] Input: None.

[0184] Output: A description of the agent’s functionality.

[0185] Interface name: binding_tool_name / / Tool name bound to the agent

[0186] Input: None.

[0187] Output: The name of the bound tool.

[0188] Interface name: set_tool_input_examples / / Set tool input parameter examples (referred to as the first reference example in this disclosure)

[0189] Input: string. A collection of reference examples showing how to convert natural language information into a list of input parameters for tools bound to an agent.

[0190] Output: None.

[0191] Interface name: tool_input_examples / / Tool input parameter examples (referred to as the first reference example in this disclosure)

[0192] Input: None.

[0193] Output: The reference examples corresponding to the set_tool_input_examples interface settings are returned.

[0194] In summary, the first prompt information can be used as prior information or task information to indicate the task information to be performed by the large language model, which can improve the prediction accuracy of the large language model. In addition, a large number of reference examples can be listed in the first prompt information to indicate how to convert natural language information (such as colloquial input text) into input parameters of the tool, which can improve the stability of the large language model generation result process.

[0195] Step C: Based on the tool calling parameters, call the tool bound to the agent indicated by the root node to obtain the execution result (or output result) of the tool, which is recorded as the tool execution result in this disclosure.

[0196] Step D: Convert the tool execution results to obtain natural language text that matches the input format required by the agent indicated by the root node.

[0197] In the disclosed embodiment, the tool execution result may be converted to obtain a natural language text (or colloquial text, colloquial result) that matches the input format required by the agent indicated by the root node.

[0198] As a possible implementation method, a large language model can be used to convert the tool execution results to obtain the natural language text required by the agent indicated by the root node.

[0199] As an example, first, the output format description associated with the tool bound to the agent indicated by the root node can be obtained, and at least one second reference example associated with the agent indicated by the root node can be obtained, wherein the second reference example is used to indicate the processing method of the output result of the tool bound to the agent indicated by the root node to the input data of the agent indicated by the root node. Afterwards, the second prompt information can be generated according to each second reference example, tool execution result, output format description and function description. For example, the second reference examples, tool execution results, output format description and function description can be assembled to obtain the second prompt information, wherein the second prompt information is used to prompt the large language model to perform a text conversion task, and then the large language model can be called to process the second prompt information to obtain the natural language text required by the agent indicated by the root node.

[0200] Exemplarily, the output format description associated with the tool can be obtained through an interface constructed in the tool. For example, the following interface can also be constructed in the tool:

[0201] Interface name: output_param_description / / Output format description of the tool

[0202] Input: None.

[0203] Output: tool output parameter description in markdown format.

[0204] Exemplarily, each second reference example associated with the agent indicated by the root node can be obtained by calling the agent interface. For example, the agent interface can also include:

[0205] Interface name: set_tool_output_examples / / Example of setting tool output format (referred to as the second reference example in this disclosure)

[0206] Input: string. A series of reference examples showing how to convert the output of the tool into the natural language text (also known as spoken text, spoken results) required by the agent.

[0207] Output: None.

[0208] Interface name: tool_output_examples / / Tool output format example (referred to as the second reference example in this disclosure)

[0209] Input: None.

[0210] Output: Corresponds to the reference examples set by the set_tool_output_examples interface, and returns these reference examples.

[0211] Interface name: query / / Input text in spoken form entered by the user

[0212] Input: None.

[0213] Output: Colloquial version of input text.

[0214] In summary, the second prompt information can be used as prior information or task information to indicate the task information to be performed by the large language model, which can improve the prediction accuracy of the large language model. In addition, a large number of reference examples can be listed in the second prompt information to indicate how to convert the output results of the tool (such as structured results in Json format) into the natural language text (such as spoken text) required by the intelligent agent, which can improve the stability of the large language model generation result process.

[0215] Step E: Use the agent indicated by the root node to process the natural language text to obtain the processing result of the agent indicated by the root node.

[0216] Step S607, the input text is processed by the agent indicated by the root node to obtain the processing result of the agent indicated by the root node.

[0217] In an embodiment of the present disclosure, when the currently called agent is the agent indicated by the root node in the agent interaction topology diagram, and the agent indicated by the root node is a non-instrumental agent, the input text can be directly processed by the agent indicated by the root node (i.e., the non-instrumental agent) to obtain the processing result of the agent indicated by the root node.

[0218] Step S608, in response to the currently called agent being the agent indicated by the non-root node in the agent interaction topology diagram, determine whether the agent indicated by the non-root node is a tool-type agent, if so, execute step S609, if not, execute step S610.

[0219] It should be noted that step S609 and step S610 are two parallel implementation methods. In actual application, only one needs to be executed.

[0220] Step S609, the processing result of the agent indicated by the non-root node is processed by the agent indicated by the non-root node and the tool bound to the agent indicated by the non-root node to obtain the processing result of the agent indicated by the non-root node.

[0221] In the disclosed embodiment, when the currently called agent is the agent indicated by the non-root node in the agent interaction topology diagram, and the agent indicated by the non-root node is a tool-type agent, the agent indicated by the non-root node and the tool bound to the agent indicated by the non-root node can be used to process the processing result of the agent indicated by the parent node of the non-root node to obtain the processing result of the agent indicated by the non-root node. The implementation principle is similar to step S606 and will not be repeated here.

[0222] As an example, the input parameter description and function description associated with the tool bound to the agent indicated by the non-root node can be obtained, and the tool calling parameters can be determined based on the processing results, input parameter description and function description of the agent indicated by the parent node of the non-root node. After that, the tool bound to the agent indicated by the non-root node can be called based on the tool calling parameters to obtain the tool execution result. Then, the tool execution result can be converted to obtain a natural language text that matches the input format required by the agent indicated by the non-root node. Finally, the natural language text can be processed by the agent indicated by the non-root node to obtain the processing result of the agent indicated by the non-root node.

[0223] Step S610, the processing result of the agent indicated by the parent node is processed by the agent indicated by the non-root node to obtain the processing result of the agent indicated by the non-root node.

[0224] In an embodiment of the present disclosure, when the currently called agent is the agent indicated by a non-root node in the agent interaction topology diagram, and the agent indicated by the non-root node is a non-instrumental agent, the processing results of the agent indicated by the parent node of the non-root node can be directly processed by the agent indicated by the non-root node (i.e., the non-instrumental agent) to obtain the processing results of the agent indicated by the non-root node.

[0225] Step S611, determining the response text according to the processing result of the agent indicated by the leaf node in the agent interaction topology diagram.

[0226] In the disclosed embodiment, the processing results of the agents indicated by the leaf nodes in the agent interaction topology diagram may be integrated to determine the response text for replying to the input text.

[0227] Step S612: reply to the input text based on the response text.

[0228] For the explanation of step S612, reference may be made to the relevant description in any embodiment of the present disclosure and will not be repeated here.

[0229] The text processing method based on a large language model in the embodiment of the present disclosure, combined with a tool-type intelligent agent and the tools bound thereto, performs collaborative processing on the input text, which can not only improve the stability of text processing, but also improve the accuracy of text processing.

[0230] In order to clearly illustrate any embodiment of the present disclosure, the present disclosure also proposes a text processing method based on a large language model.

[0231] Figure 7 This is a flowchart of a text processing method based on a large language model provided in Embodiment 6 of the present disclosure.

[0232] like Figure 7 As shown, the text processing method based on the large language model may include the following steps S701 to S708:

[0233] Step S701, performing a first classification on the input text to obtain a classification result; wherein the classification result is used to indicate a processing object of the input text.

[0234] Step S702: Perform a second classification on the input text to obtain a text category.

[0235] Step S703, in response to the processing object being an agent, obtaining a plurality of agent interaction topology graphs adapted to the text category.

[0236] Among them, the nodes in the agent interaction topology diagram are used to indicate the agents, and the edges between the nodes are used to indicate the interaction relationship between the agents. The agents include tool-type agents and / or non-tool-type agents bound to tools.

[0237] For explanations of steps S701 to S703, reference may be made to the relevant descriptions in any embodiment of the present disclosure, and will not be repeated here.

[0238] Step S704, determining the interaction order between the multiple agent interaction topology diagrams, and based on the interaction order, calling the multiple agent interaction topology diagrams in sequence.

[0239] Among them, the interaction order of multiple intelligent agent interaction topology diagrams adapted to the text category can be pre-specified, or can also be manually specified by relevant personnel, and the embodiments of the present disclosure do not limit this.

[0240] In the embodiment of the present disclosure, multiple agent interaction topology graphs can be called in sequence based on the interaction order between the multiple agent interaction topology graphs.

[0241] Step S705, in response to the currently called intelligent agent interaction topology map being the first called intelligent agent interaction topology map, the intelligent agents indicated by the nodes in the first intelligent agent interaction topology map are called in sequence to process the input text to obtain the processing result of the first intelligent agent interaction topology map.

[0242] In the embodiment of the present disclosure, when the currently called agent interaction topology diagram is the first called agent interaction topology diagram, the agents indicated by the nodes in the first agent interaction topology diagram can be called in sequence to process the input text to obtain the processing result of the first agent interaction topology diagram; wherein the processing result of the first agent interaction topology diagram can be: the processing result of the agent indicated by the leaf node in the first agent interaction topology diagram. Its implementation principle is similar to steps S604 to S610, which will not be repeated here.

[0243] Step S706, in response to the fact that the currently called intelligent agent interaction topology map is not the first called intelligent agent interaction topology map, the intelligent agents indicated by the nodes in the non-first intelligent agent interaction topology map are called in sequence, and the processing results of the previously called intelligent agent interaction topology map are processed to obtain the processing results of the non-first intelligent agent interaction topology map.

[0244] In the embodiment of the present disclosure, when the currently called agent interaction topology diagram is not the first called agent interaction topology diagram, the agents indicated by the nodes in the non-first agent interaction topology diagram can be called in sequence, and the processing result of the previously called agent interaction topology diagram can be processed to obtain the processing result of the non-first agent interaction topology diagram. Among them, the processing result of the non-first agent interaction topology diagram can be: the processing result of the agent indicated by the leaf node in the non-first agent interaction topology diagram.

[0245] Step S707, determining the response text according to the processing result of the last called intelligent agent interaction topology diagram.

[0246] In the disclosed embodiment, the processing results of the agents indicated by the leaf nodes in the last called agent interaction topology diagram can be integrated to determine the response text used to reply to the input text.

[0247] Step S708: reply to the input text based on the response text.

[0248] For the explanation of step S708, reference may be made to the relevant description in any embodiment of the present disclosure, and no further details will be given here.

[0249] The text processing method based on a large language model in the embodiment of the present disclosure, based on the interaction order between multiple agent interaction topology diagrams, sequentially calls the agents in multiple agent interaction topology diagrams to collaboratively process the input text, which can ensure that the processing of each agent is based on the processing result of the previous agent, forming a complete logical chain. This sequentiality ensures the consistency and accuracy of text processing.

[0250] In order to clearly explain how to obtain the agent interaction topology map adapted to the text category to which the input text belongs in the above embodiment, the present disclosure also proposes a text processing method based on a large language model.

[0251] Figure 8 This is a flowchart of a text processing method based on a large language model provided in Embodiment 7 of the present disclosure.

[0252] like Figure 8 As shown, the text processing method based on the large language model may include the following steps S801 to S810:

[0253] Step S801, performing a first classification on the input text to obtain a classification result; wherein the classification result is used to indicate a processing object of the input text.

[0254] Step S802: Perform a second classification on the input text to obtain a text category.

[0255] For explanations of steps S801 to S802, reference may be made to the relevant descriptions in any embodiment of the present disclosure and will not be repeated here.

[0256] Step S803, in response to the processing object being an intelligent agent, determining whether the text category is a newly added text category, if so, executing steps S806 to S808, if not, executing steps S804 to S805.

[0257] It should be noted that steps S806 to S808 and steps S804 to S805 are two parallel implementation methods. In actual application, only one needs to be executed.

[0258] Step S804: Based on the known correspondence between the text category and the business object, determine a first business object corresponding to the text category from multiple business objects.

[0259] Among them, the number of first business objects (such as business teams, business parties) can be one, or can be multiple, which is not limited in this embodiment of the present disclosure.

[0260] In an embodiment of the present disclosure, when the text category to which the input text belongs is a newly added text category, the first business object corresponding to the text category can be determined from multiple business objects based on the correspondence between known text categories and business objects (such as business teams, business parties).

[0261] Step S805, determining an agent interaction topology map adapted to the text category based on the agent interaction topology map maintained by the first business object.

[0262] Among them, the nodes in the agent interaction topology diagram are used to indicate the agents, and the edges between the nodes are used to indicate the interaction relationship between the agents. The agents include tool-type agents and / or non-tool-type agents bound to tools.

[0263] As an example, the agent interaction topology map maintained by each first business object can be used as the agent interaction topology map adapted to the text category to which the input text belongs.

[0264] Step S806: In response to the second configuration instruction, determine a second business object adapted to the text category from the multiple business objects, and establish a corresponding relationship between the text category and the second business object.

[0265] Among them, the number of second business objects (such as business teams, business parties) can be one, or can be multiple, which is not limited in the embodiments of the present disclosure.

[0266] In an embodiment of the present disclosure, when the text category to which the input text belongs is a newly added text category, a second business object that is compatible with the text category can be determined from multiple business objects in response to a configuration instruction (referred to as a second configuration instruction in the present disclosure) triggered by a relevant person (such as an administrator), that is, the relevant person manually selects the second business object that is compatible with the text category to which the input text belongs, and establishes a corresponding relationship between the text category and the second business object.

[0267] Step S807, dispatching the input text to the second business object so that the second business object updates the corresponding maintained agent interaction topology map.

[0268] In the disclosed embodiment, the input text may be dispatched to the second business object so that the second business object can update the agent interaction topology map maintained by itself based on existing business experience.

[0269] As a possible implementation method, a scheduling task can be sent to the second business object based on the input text and the text category to which it belongs, so that the second business object executes the scheduling task and updates the agent interaction topology map maintained by the second business object.

[0270] The scheduling task can be performed by following steps a to c:

[0271] Step a: Determine whether there is an agent in the agent library that is compatible with the text category to which the input text belongs. If so, execute step c; if not, execute steps b to c.

[0272] Step b: If there is no agent adapted to the text category in the agent library, then in response to the third configuration instruction triggered by the second business object, an agent adapted to the text category is created based on the input text.

[0273] As an example, the second business object can manually determine whether there is an agent in the agent library that is adapted to the text category to which the input text belongs. If not, the second business object can manually create an agent that is adapted to the text category based on the input text.

[0274] Step c: Based on the agent that is adapted to the text category to which the input text belongs, the agent interaction topology map maintained by the second business object is updated. For example, a new node is added to the agent interaction topology map maintained by the second business object, and the new node is used to indicate the agent that is adapted to the text category to which the input text belongs, or the node position, dependency relationship or interaction relationship between nodes in the agent interaction topology map maintained by the second business object can also be updated.

[0275] As an example, the agent interaction topology diagram maintained by the second business object can be updated based on the agent adapted to the text category to which the input text belongs by dragging and dropping or the like.

[0276] It should be noted that when the processing of an input text requires the participation of different business objects (or business parties), the business objects hold the most core experience of their business. If these experiences cannot be precipitated by the business objects themselves, then even if the intelligent processing system for processing input texts is implemented well, it is not maintainable. For example, when the distributed business system changes, the business objects cannot update the knowledge or experience in the intelligent processing system in a timely manner, making the intelligent processing system no longer effective. In the present disclosure, for newly added text categories, new logic or intelligent entities can be introduced in an "addition" manner to achieve business maintainability. Moreover, the new logic or intelligent entity is introduced by the business objects (or business parties) that maintain the distributed business system, which can reflect business friendliness.

[0277] As another possible implementation, the scheduling task can also be performed by following steps d to f:

[0278] Step d: Determine whether there is an atomic tool associated with an intelligent agent that is adapted to the text category to which the input text belongs among the multiple atomic tools in the tool library. If so, execute step f; if not, execute steps e to f.

[0279] Step e: If there is no atomic tool associated with the intelligent agent adapted to the text category to which the input text belongs in the tool library, then according to the fourth configuration instruction, create at least one atomic tool associated with the intelligent agent adapted to the text category, and add the at least one atomic tool to the tool library.

[0280] As an example, the second business object can manually determine whether there is an atomic tool associated with an intelligent agent that is adapted to the text category to which the input text belongs in the tool library. If not, the second business object can manually create at least one atomic tool associated with an intelligent agent that is adapted to the text category based on the input text, and add the created atomic tools to the tool library.

[0281] Step f: In response to the fifth configuration instruction, assemble at least one atomic tool associated with an agent adapted to the text category to which the input text belongs, so as to obtain a tool bound to the adapted agent.

[0282] As an example, at least one atomic tool associated with an agent adapted to the text category may be assembled by dragging and dropping or the like to obtain a tool bound to the agent adapted to the text category.

[0283] In summary, when new text categories emerge, we only need to build new tools or new intelligent agents in a low-cost manner. At the same time, we do not make any changes to the original tools or intelligent agents. This can make the intelligent processing system for processing input text more powerful without any operation on the solving capabilities of existing text categories.

[0284] Step S808, determining an agent interaction topology map adapted to the text category based on the updated agent interaction topology map maintained by the second business object.

[0285] As an example, the updated agent interaction topology map maintained by each second business object can be used as the agent interaction topology map adapted to the text category to which the input text belongs.

[0286] Step S809, calling the agents indicated by the nodes in the agent interaction topology diagram in turn to process the input text to obtain the response text.

[0287] Step S810: reply to the input text based on the response text.

[0288] The explanation of steps S809 to S810 can be found in the relevant description in any embodiment of the present disclosure and will not be repeated here.

[0289] The text processing method based on a large language model of the embodiment of the present disclosure, when the text category to which the input text belongs is an existing or known text category, directly determines the first business object corresponding to the text category to which the input text belongs from multiple business objects based on the correspondence between the known text category and the business object, and uses the agent interaction topology map maintained by the first business object as the agent interaction topology map adapted to the text category. This can realize rapid processing of input text of a known text category based on existing business experience, thereby improving the processing efficiency of the input text.

[0290] In the case where the text category to which the input text belongs is a newly added text category, a second business object that is compatible with the newly added text category is determined from multiple business objects through a visual configuration method, and the intelligent agent interaction topology map maintained by itself is updated based on the visual configuration method through the second business object, so that the input text of the newly added text category is processed based on the updated intelligent agent interaction topology map, which can improve the accuracy and effectiveness of text processing.

[0291] In order to clearly explain how the input text is first classified in the above embodiment, the present disclosure also proposes a text processing method based on a large language model.

[0292] Fig. 9 A flowchart of a text processing method based on a large language model provided in Embodiment 8 of the present disclosure.

[0293] like Fig. 9 As shown, the text processing method based on the large language model may include the following steps S901 to S906:

[0294] Step S901: obtain input text and obtain at least one third reference example; wherein the third reference example includes a first reference text and a processing object of the first reference text.

[0295] Among them, the explanation of the input text can refer to the relevant description in the above embodiment, which will not be repeated here.

[0296] The processing objects used to process the first reference text include but are not limited to: tools, agents, etc. The tools include but are not limited to: script commands, APIs, functions, etc.

[0297] Each third reference example is preset.

[0298] Step S902: Generate third prompt information according to the third reference example and the input text; wherein the third prompt information is used to prompt the large language model to perform the text classification task.

[0299] In the embodiment of the present disclosure, each third reference example can be assembled with the input text to obtain third prompt information, wherein the third prompt information is used to prompt the large language model to perform a text classification task.

[0300] Step S903: calling the large language model to process the third prompt information to obtain a classification result; wherein the classification result is used to indicate a processing object of the input text.

[0301] In the disclosed embodiment, a large language model may be called to process the third prompt information to obtain a classification result; wherein the classification result is used to indicate a processing object of the input text.

[0302] Step S904: perform a second classification on the input text to obtain a text category.

[0303] Step S905: Based on the text category, the processing object is called to process the input text to obtain a response text.

[0304] Step S906: reply to the input text based on the response text.

[0305] The explanation of steps S904 to S906 can be found in the relevant description in any embodiment of the present disclosure, and will not be repeated here.

[0306] In the text processing method based on the large language model of the embodiment of the present disclosure, the third prompt information can be used as prior information or task information to indicate the task information to be performed by the large language model, which can improve the prediction accuracy of the large language model. In addition, a large number of reference examples can be listed in the third prompt information to indicate the processing objects of different reference texts, so that the large language model can only classify within the scope of the processing objects indicated by the third prompt information, without having to classify in all processing objects, which can eliminate classification illusions, thereby improving the reliability of the classification results, that is, improving the stability of the large language model generation result process.

[0307] In order to clearly explain how the second classification of the input text is performed in the above embodiment, the present disclosure also proposes a text processing method based on a large language model.

[0308] Fig.10 A flowchart of a text processing method based on a large language model provided in Embodiment 9 of the present disclosure.

[0309] like Fig.10 As shown, the text processing method based on the large language model may include the following steps S1001 to S1007:

[0310] Step S1001, performing a first classification on the input text to obtain a classification result; wherein the classification result is used to indicate a processing object of the input text.

[0311] For the explanation of step S1001, reference may be made to the relevant description in any embodiment of the present disclosure, and no further details will be given here.

[0312] Step S1002, obtaining a plurality of fourth reference examples; wherein the fourth reference examples include the second reference text and the text category to which the second reference text belongs.

[0313] The text category to which the second reference text belongs may vary depending on the specific application scenario, business requirements, and characteristics of the text content of the second reference text.

[0314] Each fourth reference example is preset.

[0315] Step S1003: Generate fourth prompt information according to multiple fourth reference examples and the input text; wherein the fourth prompt information is used to prompt the large language model to perform a text classification task.

[0316] In an embodiment of the present disclosure, multiple fourth reference examples and input text may be assembled to obtain fourth prompt information, wherein the fourth prompt information is used to prompt the large language model to perform a text classification task.

[0317] Step S1004: calling the large language model to process the fourth prompt information to obtain the text category to which the input text belongs.

[0318] In the disclosed embodiment, a large language model may be called to process the fourth prompt information to obtain the text category to which the input text belongs.

[0319] Step S1005: Based on the text category, the processing object is called to process the input text to obtain a response text.

[0320] Step S1006: reply to the input text based on the response text.

[0321] The explanation of steps S1005 to S1006 can be found in the relevant description in any embodiment of the present disclosure, and will not be repeated here.

[0322] In the text processing method based on the large language model of the embodiment of the present disclosure, the fourth prompt information can be used as prior information or task information to indicate the task information to be performed by the large language model, which can improve the prediction accuracy of the large language model. In addition, a large number of reference examples can be listed in the fourth prompt information to indicate the existing text categories, so that the large language model can only classify within the text category range indicated by the fourth prompt information, without having to classify in all text categories, which can eliminate classification illusions, thereby improving the reliability of the classification results, that is, improving the stability of the large language model generation result process.

[0323] In any embodiment of the present disclosure, the following may be used: Fig.11 The low-code tool platform shown is used to implement the above-mentioned method embodiments, wherein the first layer (tool layer) and the second layer (scenario layer) in the low-code tool platform are the core two layers of the present disclosure.

[0324] The first layer (tool layer, also known as tool library, tool market) completes the construction of tools through low-code methods. Among them, each tool in the tool layer is also called atomic tool.

[0325] Fig.11Clickhouse in the tool layer is a completely column-based distributed database management system; the callgraph platform is a distributed service tracking system used to monitor and track the call relationships between components in a distributed business system; the tower platform is a project management tool designed for business team collaboration, which provides task management, project progress tracking, team collaboration and other functions to help business teams advance projects more efficiently; NL2SQL is a technology that converts natural language into SQL (Structured Query Language) queries (Natural Language to Structured Query Language).

[0326] The second layer (scenario layer, also called assembly layer) assembles the atomic tools of the first layer to form larger tools to process relatively complex input text. In addition, intelligent agents can be created in the second layer, and the interaction topology diagram between intelligent agents can be constructed to achieve the purpose of processing complex text through natural language.

[0327] The low-code tool platform has at least the following advantages:

[0328] Part one: stable execution capability.

[0329] 1.1. Fine-grained agent.

[0330] The simpler the responsibility of an intelligent agent, the more stable its execution is. Therefore, in the present disclosure, when constructing an intelligent agent, it is avoided to be too large, and an intelligent agent can only complete one task. If the intelligent agent has to complete too many things, the intelligent agent is split.

[0331] 1.2. Tool-type intelligent agent.

[0332] The reason why the intelligent agent has the problem of unstable execution is that it relies on LLM. However, because the tool is hard-coded logic, there is no problem of execution stability. In other words, the execution of the tool is the most stable, because in the present disclosure, the tool can also "play" the role of the intelligent agent and have the ability to interact with the outside world through natural language. This kind of intelligent agent can be called a tool-type intelligent agent.

[0333] The specific steps are as follows:

[0334] (1) Interfaces built in the tool

[0335] Interface name: name / / name of the tool

[0336] Input: None.

[0337] Output: Tool name.

[0338] Interface name: description / / Function description of the tool

[0339] Input: None.

[0340] Output: Description of the tool's functionality.

[0341] Interface name: input_param_description / / tool input parameter description

[0342] Input: None.

[0343] Output: Description of tool input parameters in markdown format.

[0344] Interface name: output_param_description / / Output format description of the tool

[0345] Input: None.

[0346] Output: tool output parameter description in markdown format.

[0347] Interface name: run_with_params / / Call the tool based on the parameters, the execution result of the tool

[0348] Input: parameter list in json format.

[0349] Output: tool execution results.

[0350] (2) Agent interface / / Agent name

[0351] Interface name: name

[0352] Input: None.

[0353] Output: Agent name.

[0354] Interface name: description / / Functional description of the agent

[0355] Input: None.

[0356] Output: A description of the agent’s functionality.

[0357] Interface name: binding_tool_name / / The name of the tool bound to the agent

[0358] Input: None.

[0359] Output: The name of the bound tool.

[0360] Interface name: set_tool_input_examples / / Set tool input parameter examples

[0361] Input: string, i.e., a list of examples showing how to convert natural language information into a list of input parameters for a tool bound to an agent, i.e., a string indicating how to convert natural language information into input parameters for a tool;

[0362] Output: None.

[0363] Interface name: tool_input_examples / / Tool input parameter examples

[0364] Input: None.

[0365] Output: Examples corresponding to the set_tool_input_examples interface setting, which are returned.

[0366] Interface name: set_tool_output_examples / / Set tool output format examples

[0367] Input: a string, i.e., a series of examples showing how to convert the output of the tool into the spoken input data of the agent, i.e., indicating the specific processing method of the output of the tool into the spoken input data;

[0368] Output: None.

[0369] Interface name: tool_output_examples / / Tool output format examples

[0370] Input: None.

[0371] Output: Examples corresponding to the set_tool_output_examples interface setting are returned.

[0372] Interface name: query / / Spoken input text entered by the user

[0373] Input: None.

[0374] Output: Colloquial version of input text.

[0375] As an example, you can use Fig.12 The interactive process shown in the figure is used to process the query input by the user, where: Fig.12 The middle gray area refers to the tool execution engine, which is responsible for coordinating function calls between tools and agents and interacting with LLM. Fig.12 The area to the left of is the tool-type agent. Fig.12 The area to the right of is the tool, and the two are bound through the agent's binding_tool_name (binding tool name) interface.

[0376] Fig.12 It consists of two parts, the upper and lower parts. Fig.12 The upper part mainly converts the input text (query) of the colloquial description into tool call parameters. Specifically, the tool execution engine calls the tool-type agent and tool-related functions (such as the tool-type agent's tool_input_examples, the tool's description, and the tool's input_param_description) to obtain the corresponding information and assemble it to obtain the prompt. Then, the tool execution engine uses the prompt to request the LLM and obtains the various parameters of the tool described in the Json format returned by the LLM (i.e., the tool call parameters). Finally, the tool execution engine can call the tool's run_with_param interface based on these parameters, execute the tool, and obtain the tool's execution result (structured result in Json format).

[0377] Fig.12 The lower part mainly completes the conversion of the tool's execution result (structured result) into the spoken result required by the tool-type agent, that is, the conversion of the tool's execution result into a spoken result that matches the input format required by the tool-type agent. Specifically, first, the tool execution engine calls the tool-type agent and tool-related functions (such as the tool's description, the tool's output_param_description, and the tool_output_examples of the tool-type agent) to obtain the corresponding information and assemble it to obtain the prompt. Then, the tool execution engine uses the prompt to request the LLM to obtain the spoken result required by the tool-type agent returned by the LLM. Finally, the tool execution engine returns the spoken result to the agent for processing to obtain the processed result output by the agent.

[0378] 1.3. Enhance stability by conveniently introducing examples.

[0379] During the execution of the above-mentioned tool-based intelligent agent, a large number of examples were listed when constructing the two prompts to further improve the stability of the LLM result generation process.

[0380] The second part is maintainability.

[0381] 2.1. Align organizational structure.

[0382] The interactive architecture of tool-based agents follows the division of business teams. First, there is a general entry agent at the top level, and under the general entry agent, there are more agents divided according to each business team. For each business team, the business team has an entry agent that connects with the general entry agent, and the entry agent of the business team can include more sub-agents for division of labor and collaboration, such as Fig.13 shown.

[0383] exist Fig.13 In the architecture shown, each business team maintains the agent within its own business scope. In addition, the granularity of the agent can be split according to the granularity that the business considers appropriate, and different agents interact with each other in natural language.

[0384] The general entry agent is responsible for understanding the original query input by the user and organizing the entry agents of each business team to complete the task.

[0385] 2.2. New text type query routing.

[0386] When a new text-type query appears, there is no need to modify the existing logic of the agent, but only to add new logic. Specifically:

[0387] (1) Add examples for queries of new text categories in the main entry agent, indicating which business team the query corresponding to the text category should be routed to, or through the collaboration of which business teams to complete it.

[0388] (2) If there is a lack of an agent necessary to complete the query within the agent maintained by the business team, a new agent is created.

[0389] The third part is low cost.

[0390] 3.1. Low-cost tool assembly.

[0391] There are many kinds of atomic tools in the tool layer, which are the basis for building scenes. In order to support the ability of "low-cost scene building", each atomic tool can be designed as follows:

[0392] (1) The input format and output format are fixed, both are a key-value map.

[0393] (2) Multiple atomic tools can be connected through Fig.14 The "data whiteboard" shown transfers data.

[0394] The "data whiteboard" can be regarded as a central data storage or message queue for transferring data between different atomic tools. The data whiteboard acts as a neutral and accessible data exchange center, so that any atomic tool can read or write data on it. Transferring data through the "data whiteboard" has at least the following advantages:

[0395] Decoupling: Atomic tools do not need to communicate directly, but communicate indirectly through the "data whiteboard", which reduces the coupling between atomic tools and improves the maintainability and scalability of the system.

[0396] Asynchronous processing: Asynchronous data processing is supported. Atomic tools can publish results to the "data whiteboard" after completing the current task without having to wait for responses from other atomic tools.

[0397] Concurrent processing: Multiple atomic tools can read and process data from the "data whiteboard" at the same time, improving overall processing efficiency.

[0398] After supporting the above two capabilities, atomic tools can be assembled through the data whiteboard to form more complex tools. In terms of implementation, the assembly of atomic tools can be completed through the "drag and drop" method of the UI (User Interface).

[0399] 3.2. Low-cost assembly of intelligent agents.

[0400] Similar to tools, agents can also be assembled by "dragging" on the UI interface. Unlike tool assembly, the assembled tools are called through deterministic APIs, while the assembled agents interact with each other through natural language.

[0401] In summary, the solution provided by the present disclosure has at least the following advantages:

[0402] 1. Stability: By maintaining the fine-grained nature of the agent, implementing a tool-type agent, and using an example-based approach (i.e., setting examples), the stability requirements of the agent execution process are ensured;

[0403] 2. Maintainability: Align the agent architecture with the organizational architecture and introduce new logic by “increasing” the processing capabilities of new text types to achieve business maintainability;

[0404] 3. Low cost: A “data whiteboard” for exchanging data between tools is implemented, and the function of supporting three assembly modes is implemented to achieve drag-and-drop assembly of tools; the assembly method of intelligent agents is similar. The logic in the intelligent agent is completely input through the natural language of the business. Through these methods, low-cost creation of tools and intelligent agents is achieved.

[0405] With the above Figures 1 to 10 Corresponding to the text processing method based on a large language model provided in the embodiment, the present disclosure also provides a text processing device based on a large language model. Figures 1 to 10 The text processing method based on a large language model provided in the embodiment corresponds to this, so the implementation of the text processing method based on a large language model is also applicable to the text processing device based on a large language model provided in the embodiment of the present disclosure, and will not be described in detail in the embodiment of the present disclosure.

[0406] Fig.15 This is a schematic diagram of the structure of a text processing device based on a large language model provided in Embodiment 10 of the present disclosure.

[0407] like Fig.15 As shown, the text processing device 1500 based on the large language model may include: a first classification module 1510 , a second classification module 1520 , a calling module 1530 and a reply module 1540 .

[0408] The first classification module 1510 is used to perform a first classification on the input text to obtain a classification result; wherein the classification result is used to indicate a processing object of the input text;

[0409] The second classification module 1520 is used to perform a second classification on the input text to obtain a text category;

[0410] A calling module 1530 is used to call a processing object to process the input text based on the text category to obtain a response text;

[0411] The reply module 1540 is used to reply to the input text based on the response text.

[0412] In a possible implementation of the embodiment of the present disclosure, the calling module 1530 is used to: in response to the processing object being a tool, obtain a target tool adapted to the text category; wherein the target tool is assembled from at least one atomic tool; and call the target tool to process the input text to obtain a response text.

[0413] In a possible implementation of the embodiment of the present disclosure, the calling module 1530 is used to: in response to the processing object being a tool, query whether there is an assembly tool associated with the text category, wherein the assembly tool is obtained by assembling at least one atomic tool associated with the text category in the tool library within a historical period; if so, the assembly tool is used as the target tool; if not, obtain the atomic tool adapted to the text category; wherein the adapted atomic tool includes: the atomic tool created based on the input text in response to the first configuration instruction, and / or the atomic tool in the tool library; assemble the adapted atomic tools to obtain the target tool.

[0414] In a possible implementation of the embodiment of the present disclosure, there are multiple target tools, and the calling module 1530 is used to: determine the calling order among the multiple target tools, and call the multiple target tools in sequence based on the calling order; in response to the target tool currently called being the first target tool called, the first target tool is called to process the input text to obtain the processing result of the first target tool; in response to the target tool currently called being not the first target tool called, the non-first target tool is called to process the processing result of the target tool previously called to obtain the processing result of the non-first target tool; and determine the response text according to the processing result of the target tool last called.

[0415] In a possible implementation of the embodiments of the present disclosure, calling module 1530 is used to: in response to the processing object being an agent, obtain an agent interaction topology map adapted to the text category; wherein the nodes in the agent interaction topology map are used to indicate the agents, and the edges between the nodes are used to indicate the interaction relationship between the agents, and the agents include tool-type agents and / or non-tool-type agents bound to tools; sequentially call the agents indicated by each node in the agent interaction topology map to process the input text; and determine the response text based on the processing results of the agents indicated by the leaf nodes in the agent interaction topology map.

[0416] In a possible implementation of the embodiment of the present disclosure, the calling module 1530 is used to: sequentially call the agents indicated by each node in the agent interaction topology diagram; in response to the currently called agent being the agent indicated by the root node in the agent interaction topology diagram, and the agent indicated by the root node being a tool-type agent, the input text is processed by the agent indicated by the root node and the tool bound to the agent indicated by the root node to obtain the processing result of the agent indicated by the root node; in response to the agent indicated by the root node being a non-tool-type agent, the input text is processed by the agent indicated by the root node to obtain the processing result of the agent indicated by the root node. the processing result of the agent indicated by the non-root node; in response to the currently called agent being the agent indicated by the non-root node in the agent interaction topology diagram, and the agent indicated by the non-root node being a tool-type agent, the processing result of the agent indicated by the parent node of the non-root node is processed through the agent indicated by the non-root node and the tool bound to the agent indicated by the non-root node to obtain the processing result of the agent indicated by the non-root node; in response to the agent indicated by the non-root node being a non-tool-type agent, the processing result of the agent indicated by the parent node is processed through the agent indicated by the non-root node to obtain the processing result of the agent indicated by the non-root node.

[0417] In a possible implementation of the embodiments of the present disclosure, the calling module 1530 is used to: determine tool calling parameters based on input text and input parameter descriptions and function descriptions associated with the tool bound to the agent indicated by the root node; based on the tool calling parameters, call the tool bound to the agent indicated by the root node to obtain the tool execution result; convert the tool execution result to obtain a natural language text that matches the input format required by the agent indicated by the root node; and process the natural language text using the agent indicated by the root node to obtain the processing result of the agent indicated by the root node.

[0418] In a possible implementation of the embodiments of the present disclosure, the calling module 1530 is used to: obtain at least one first reference example associated with the intelligent agent indicated by the root node; wherein the first reference example is used to indicate a conversion method for converting natural language information into input parameters of a tool bound to the intelligent agent indicated by the root node; generate first prompt information according to each first reference example, input parameter description, function description and input text; wherein the first prompt information is used to prompt the large language model to perform a calling parameter generation task; and use the large language model to process the first prompt information to obtain tool calling parameters.

[0419] In a possible implementation of the embodiments of the present disclosure, module 1530 is called to: obtain an output format description associated with a tool bound to an agent indicated by a root node; obtain at least one second reference example associated with the agent indicated by the root node; wherein the second reference example is used to indicate a method for processing the output result of the tool bound to the agent indicated by the root node to the input data of the agent indicated by the root node; generate a second prompt message according to each second reference example, tool execution result, output format description and function description; wherein the second prompt message is used to prompt a large language model to perform a text conversion task; and the second prompt message is processed using a large language model to obtain the natural language text required by the agent indicated by the root node.

[0420] In a possible implementation of the embodiment of the present disclosure, module 1530 is called to: determine whether the text category is a newly added text category; if not, determine a first business object corresponding to the text category from multiple business objects based on the correspondence between known text categories and business objects; and determine an agent interaction topology map adapted to the text category based on the agent interaction topology map maintained by the first business object.

[0421] In a possible implementation of the embodiment of the present disclosure, the calling module 1530 is also used to: if yes, then in response to the second configuration instruction, determine a second business object that is adapted to the text category from multiple business objects, and establish a corresponding relationship between the text category and the second business object; schedule the input text to the second business object so that the second business object updates the corresponding maintained intelligent agent interaction topology map; determine the intelligent agent interaction topology map that is adapted to the text category based on the updated intelligent agent interaction topology map maintained by the second business object.

[0422] In a possible implementation of the embodiments of the present disclosure, the first classification module 1510 is used to: obtain at least one third reference example; wherein the third reference example includes a first reference text and a processing object of the first reference text; generate third prompt information according to the third reference example and the input text; wherein the third prompt information is used to prompt the large language model to perform a text classification task; and call the large language model to process the third prompt information to obtain a classification result.

[0423] In a possible implementation of the embodiments of the present disclosure, the second classification module 1520 is used to: obtain at least one fourth reference example; wherein the fourth reference example includes a second reference text and a text category to which the second reference text belongs; generate fourth prompt information based on the fourth reference example and the input text; wherein the fourth prompt information is used to prompt the large language model to perform a text classification task; and call the large language model to process the fourth prompt information to obtain a text category.

[0424] The text processing device based on the large language model in the embodiment of the present disclosure can implement targeted processing of different input texts by calling a processing object adapted to the input text based on the text category to which the input text belongs, thereby improving the accuracy and rationality of text processing and improving the user experience.

[0425] In order to implement the above embodiments, the present disclosure also provides an electronic device, which may include at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the text processing method based on the large language model proposed in any of the above embodiments of the present disclosure.

[0426] In order to implement the above embodiments, the present disclosure further provides a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to enable a computer to execute the text processing method based on a large language model proposed in any of the above embodiments of the present disclosure.

[0427] In order to implement the above embodiments, the present disclosure further provides a computer program product, which includes a computer program. When the computer program is executed by a processor, it implements the text processing method based on a large language model proposed in any of the above embodiments of the present disclosure.

[0428] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium and a computer program product.

[0429] Fig.16 A schematic block diagram of an example electronic device that can be used to implement an embodiment of the present disclosure is shown. Among them, the electronic device may include the server and client in the above-mentioned embodiment. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workbenches, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present disclosure described and / or required herein.

[0430] like Fig.16 As shown, the electronic device 1600 includes a computing unit 1601, which can perform various appropriate actions and processes according to a computer program stored in a ROM (Read-Only Memory) 1602 or a computer program loaded from a storage unit 1607 to a RAM (Random Access Memory) 1603. In RAM 1603, various programs and data required for the operation of the device 1600 can also be stored. The computing unit 1601, ROM 1602, and RAM 1603 are connected to each other via a bus 1604. An I / O (Input / Output) interface 1605 is also connected to the bus 1604.

[0431] A number of components in the device 1600 are connected to the I / O interface 1605, including: an input unit 1606, such as a keyboard, a mouse, etc.; an output unit 1607, such as various types of displays, speakers, etc.; a storage unit 1608, such as a disk, an optical disk, etc.; and a communication unit 1609, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 1609 allows the device 1600 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0432] The computing unit 1601 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 1601 include, but are not limited to, a CPU (Central Processing Unit), a GPU (Graphic Processing Units), various dedicated AI (Artificial Intelligence) computing chips, various computing units running machine learning model algorithms, a DSP (Digital Signal Processor), and any appropriate processor, controller, microcontroller, etc. The computing unit 1601 performs the various methods and processes described above, such as the above-mentioned text processing method based on a large language model. For example, in some embodiments, the above-mentioned text processing method based on a large language model may be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as a storage unit 1608. In some embodiments, part or all of the computer program may be loaded and / or installed on the device 1600 via the ROM 1602 and / or the communication unit 1609. When the computer program is loaded into RAM 1603 and executed by the computing unit 1601, one or more steps of the text processing method based on the large language model described above may be performed. Alternatively, in other embodiments, the computing unit 1601 may be configured to perform the above-described text processing method based on the large language model in any other appropriate manner (e.g., by means of firmware).

[0433] Various embodiments of the systems and techniques described above herein may be implemented in digital electronic circuit systems, integrated circuit systems, FPGAs (Field Programmable Gate Arrays), ASICs (Application-Specific Integrated Circuits), ASSPs (Application Specific Standard Products), SOCs (System On Chips), CPLDs (Complex Programmable Logic Devices), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include: being implemented in one or more computer programs that may be executed and / or interpreted on a programmable system including at least one programmable processor that may be a dedicated or general-purpose programmable processor that may receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0434] The program code for implementing the method of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code may be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.

[0435] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or device, or any suitable combination of the foregoing. More specific examples of machine-readable storage media may include electrical connections based on one or more lines, portable computer disks, hard disks, RAM, ROM, EPROM (Electrically Programmable Read-Only-Memory) or flash memory, optical fiber, CD-ROM (Compact Disc Read-Only Memory), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0436] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (Cathode-Ray Tube) or LCD (Liquid Crystal Display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0437] The systems and techniques described herein may be implemented in a computing system that includes backend components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes frontend components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such backend components, middleware components, or frontend components. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: LAN (Local Area Network), WAN (Wide Area Network), the Internet, and blockchain networks.

[0438] A computer system may include a client and a server. The client and the server are generally remote from each other and usually interact through a communication network. The relationship between the client and the server is generated by computer programs running on the corresponding computers and having a client-server relationship with each other. The server may be a cloud server, also known as a cloud computing server or a cloud host, which is a host product in the cloud computing service system to solve the defects of difficult management and weak business scalability in traditional physical hosts and VPS services (Virtual Private Server). The server may also be a server of a distributed system, or a server combined with a blockchain.

[0439] It should be noted that artificial intelligence is a discipline that studies how computers can simulate certain human thought processes and intelligent behaviors (such as learning, reasoning, thinking, planning, etc.), and includes both hardware-level and software-level technologies. Artificial intelligence hardware technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, and big data processing; artificial intelligence software technologies mainly include computer vision technology, speech recognition technology, natural language processing technology, as well as machine learning / deep learning, big data processing technology, knowledge graph technology, and other major directions.

[0440] According to the technical solution of the embodiment of the present disclosure, it is possible to implement targeted processing of different input texts by calling a processing object adapted to the input text based on the text category to which the input text belongs, thereby improving the accuracy and rationality of text processing and improving the user experience.

[0441] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps recorded in this disclosure can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved, and this document does not limit this.

[0442] The above specific implementations do not constitute a limitation on the protection scope of the present disclosure. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A text processing method based on a large language model, comprising: Performing a first classification on the input text to obtain a classification result; wherein the classification result is used to indicate a processing object of the input text; Performing a second classification on the input text to obtain a text category; Based on the text category, calling the processing object to process the input text to obtain a response text; The input text is replied based on the response text.

2. The method according to claim 1, wherein: The step of calling the processing object to process the input text based on the text category to obtain a response text includes: In response to the processing object being a tool, obtaining a target tool adapted to the text category; wherein the target tool is assembled from at least one atomic tool; The target tool is called to process the input text to obtain the response text.

3. The method according to claim 2, wherein: In response to the processing object being a tool, obtaining a target tool adapted to the text category includes: In response to the processing object being a tool, querying whether there is an assembly tool associated with the text category, wherein the assembly tool is obtained by assembling at least one atomic tool associated with the text category in a tool library within a historical period; If yes, taking the assembly tool as the target tool; If not, then obtaining an atomic tool adapted to the text category; wherein the adapted atomic tool includes: an atomic tool created based on the input text in response to the first configuration instruction, and / or an atomic tool in the tool library; The adapted atomic tools are assembled to obtain the target tool.

4. The method according to claim 2, wherein: There are multiple target tools, and calling the target tools to process the input text to obtain the response text includes: Determine a calling sequence among a plurality of target tools, and based on the calling sequence, sequentially call the plurality of target tools; In response to the currently called target tool being the first called target tool, calling the first target tool to process the input text to obtain a processing result of the first target tool; In response to the currently called target tool being a non-first called target tool, calling the non-first target tool, processing the processing result of the previously called target tool, and obtaining the processing result of the non-first target tool; The response text is determined according to the processing result of the last called target tool.

5. The method according to claim 1, wherein: The step of calling the processing object to process the input text based on the text category to obtain a response text includes: In response to the processing object being an agent, an agent interaction topology map adapted to the text category is obtained; wherein the nodes in the agent interaction topology map are used to indicate agents, and the edges between the nodes are used to indicate the interaction relationship between agents, and the agents include tool-type agents and / or non-tool-type agents bound to tools; Invoking the agents indicated by the nodes in the agent interaction topology diagram in sequence to process the input text; The response text is determined according to the processing result of the agent indicated by the leaf node in the agent interaction topology diagram.

6. The method according to claim 5, wherein: The step of sequentially calling the agents indicated by the nodes in the agent interaction topology diagram to process the input text includes: The agents indicated by the nodes in the agent interaction topology diagram are called in sequence, and in response to the currently called agent being the agent indicated by the root node in the agent interaction topology diagram, and the agent indicated by the root node is a tool-type agent, the input text is processed by the agent indicated by the root node and the tool bound to the agent indicated by the root node to obtain the processing result of the agent indicated by the root node; In response to the agent indicated by the root node being a non-tool-type agent, the input text is processed by the agent indicated by the root node to obtain a processing result of the agent indicated by the root node; In response to the currently called agent being the agent indicated by the non-root node in the agent interaction topology diagram, and the agent indicated by the non-root node being a tool-type agent, the processing result of the agent indicated by the parent node of the non-root node is processed by the agent indicated by the non-root node and the tool bound to the agent indicated by the non-root node to obtain the processing result of the agent indicated by the non-root node; In response to the agent indicated by the non-root node being a non-tool-type agent, the processing result of the agent indicated by the parent node is processed by the agent indicated by the non-root node to obtain the processing result of the agent indicated by the non-root node.

7. The method according to claim 6, wherein: The step of processing the input text by the agent indicated by the root node and the tool bound to the agent indicated by the root node to obtain the processing result of the agent indicated by the root node includes: Determine tool call parameters according to the input text and the input parameter description and function description associated with the tool bound to the agent indicated by the root node; Based on the tool calling parameters, calling the tool bound to the agent indicated by the root node to obtain the tool execution result; Converting the tool execution result to obtain a natural language text that matches the input format required by the agent indicated by the root node; The natural language text is processed using the agent indicated by the root node to obtain a processing result of the agent indicated by the root node.

8. The method according to claim 7, wherein: The step of determining the tool calling parameters according to the input text and the input parameter description and the function description associated with the tool bound to the agent indicated by the root node comprises: Acquire at least one first reference example associated with the agent indicated by the root node; wherein the first reference example is used to indicate a conversion method for converting natural language information into input parameters of a tool bound to the agent indicated by the root node; Generate first prompt information according to each of the first reference examples, the input parameter description, the function description and the input text; wherein the first prompt information is used to prompt the large language model to execute a call parameter generation task; The first prompt information is processed using the large language model to obtain the tool calling parameter.

9. The method according to claim 7, wherein: The step of converting the tool execution result to obtain a natural language text matching the input format required by the agent indicated by the root node comprises: Obtain an output format description associated with a tool bound to the agent indicated by the root node; Acquire at least one second reference example associated with the agent indicated by the root node; wherein the second reference example is used to indicate a processing method of the output result of the tool bound to the agent indicated by the root node to the input data of the agent indicated by the root node; Generate second prompt information according to each of the second reference examples, the tool execution result, the output format description and the function description; wherein the second prompt information is used to prompt the large language model to perform a text conversion task; The second prompt information is processed using the large language model to obtain the natural language text required by the agent indicated by the root node.

10. The method according to claim 5, wherein: The step of obtaining an agent interaction topology map adapted to the text category includes: Determining whether the text category is a newly added text category; If not, determining a first business object corresponding to the text category from a plurality of business objects based on a known correspondence between the text category and the business object; According to the agent interaction topology map maintained by the first business object, determine the agent interaction topology map adapted to the text category.

11. The method according to claim 10, wherein: The step of obtaining an agent interaction topology diagram adapted to the text category further includes: If so, in response to a second configuration instruction, determining a second business object adapted to the text category from the multiple business objects, and establishing a corresponding relationship between the text category and the second business object; Dispatching the input text to the second business object so that the second business object updates the corresponding maintained agent interaction topology graph; According to the updated agent interaction topology map maintained by the second business object, an agent interaction topology map adapted to the text category is determined.

12. The method according to any one of claims 1 to 11, wherein: The first classification of the input text to obtain the classification result includes: Acquire at least one third reference example; wherein the third reference example includes a first reference text and a processing object of the first reference text; Generate third prompt information according to the third reference example and the input text; wherein the third prompt information is used to prompt the large language model to perform a text classification task; The large language model is called to process the third prompt information to obtain the classification result.

13. The method according to any one of claims 1 to 11, wherein: The second classification of the input text to obtain a text category includes: Acquire at least one fourth reference example; wherein the fourth reference example includes a second reference text and a text category to which the second reference text belongs; Generate fourth prompt information according to the fourth reference example and the input text; wherein the fourth prompt information is used to prompt the large language model to perform a text classification task; The large language model is called to process the fourth prompt information to obtain the text category.

14. A text processing device based on a large language model, comprising: A first classification module, used to perform a first classification on the input text to obtain a classification result; wherein the classification result is used to indicate a processing object of the input text; A second classification module, used for performing a second classification on the input text to obtain a text category; A calling module, used for calling the processing object to process the input text based on the text category to obtain a response text; A reply module is used to reply to the input text based on the response text.

15. The device according to claim 14, wherein: The calling module is used to: In response to the processing object being a tool, obtaining a target tool adapted to the text category; wherein the target tool is assembled from at least one atomic tool; The target tool is called to process the input text to obtain the response text.

16. The device according to claim 15, wherein: The calling module is used to: In response to the processing object being a tool, querying whether there is an assembly tool associated with the text category, wherein the assembly tool is obtained by assembling at least one atomic tool associated with the text category in a tool library within a historical period; If yes, taking the assembly tool as the target tool; If not, then obtaining an atomic tool adapted to the text category; wherein the adapted atomic tool includes: an atomic tool created based on the input text in response to the first configuration instruction, and / or an atomic tool in the tool library; The adapted atomic tools are assembled to obtain the target tool.

17. The device according to claim 15, wherein: There are multiple target tools, and the calling module is used to: Determine a calling sequence among a plurality of target tools, and based on the calling sequence, sequentially call the plurality of target tools; In response to the currently called target tool being the first called target tool, calling the first target tool to process the input text to obtain a processing result of the first target tool; In response to the currently called target tool being a non-first called target tool, calling the non-first target tool, processing the processing result of the previously called target tool, and obtaining the processing result of the non-first target tool; The response text is determined according to the processing result of the last called target tool.

18. The device according to claim 14, wherein: The calling module is used to: In response to the processing object being an agent, an agent interaction topology map adapted to the text category is obtained; wherein the nodes in the agent interaction topology map are used to indicate agents, and the edges between the nodes are used to indicate the interaction relationship between agents, and the agents include tool-type agents and / or non-tool-type agents bound to tools; Invoking the agents indicated by the nodes in the agent interaction topology diagram in sequence to process the input text; The response text is determined according to the processing result of the agent indicated by the leaf node in the agent interaction topology diagram.

19. The device according to claim 18, wherein: The calling module is used to: The agents indicated by the nodes in the agent interaction topology diagram are called in sequence, and in response to the currently called agent being the agent indicated by the root node in the agent interaction topology diagram, and the agent indicated by the root node is a tool-type agent, the input text is processed by the agent indicated by the root node and the tool bound to the agent indicated by the root node to obtain the processing result of the agent indicated by the root node; In response to the agent indicated by the root node being a non-tool-type agent, the input text is processed by the agent indicated by the root node to obtain a processing result of the agent indicated by the root node; In response to the currently called agent being the agent indicated by the non-root node in the agent interaction topology diagram, and the agent indicated by the non-root node being a tool-type agent, the processing result of the agent indicated by the parent node of the non-root node is processed by the agent indicated by the non-root node and the tool bound to the agent indicated by the non-root node to obtain the processing result of the agent indicated by the non-root node; In response to the agent indicated by the non-root node being a non-tool-type agent, the processing result of the agent indicated by the parent node is processed by the agent indicated by the non-root node to obtain the processing result of the agent indicated by the non-root node.

20. The device according to claim 19, wherein The calling module is used to: Determine tool call parameters according to the input text and the input parameter description and function description associated with the tool bound to the agent indicated by the root node; Based on the tool calling parameters, calling the tool bound to the agent indicated by the root node to obtain the tool execution result; Converting the tool execution result to obtain a natural language text that matches the input format required by the agent indicated by the root node; The natural language text is processed using the agent indicated by the root node to obtain a processing result of the agent indicated by the root node.

21. The device according to claim 20, wherein: The calling module is used to: Acquire at least one first reference example associated with the agent indicated by the root node; wherein the first reference example is used to indicate a conversion method for converting natural language information into input parameters of a tool bound to the agent indicated by the root node; Generate first prompt information according to each of the first reference examples, the input parameter description, the function description and the input text; wherein the first prompt information is used to prompt the large language model to execute a call parameter generation task; The first prompt information is processed using the large language model to obtain the tool calling parameter.

22. The device according to claim 20, wherein: The calling module is used to: Obtain an output format description associated with a tool bound to the agent indicated by the root node; Acquire at least one second reference example associated with the agent indicated by the root node; wherein the second reference example is used to indicate a processing method of the output result of the tool bound to the agent indicated by the root node to the input data of the agent indicated by the root node; Generate second prompt information according to each of the second reference examples, the tool execution result, the output format description and the function description; wherein the second prompt information is used to prompt the large language model to perform a text conversion task; The second prompt information is processed using the large language model to obtain the natural language text required by the agent indicated by the root node.

23. The device according to claim 18, wherein The calling module is used to: Determining whether the text category is a newly added text category; If not, determining a first business object corresponding to the text category from a plurality of business objects based on a known correspondence between the text category and the business object; According to the agent interaction topology map maintained by the first business object, determine the agent interaction topology map adapted to the text category.

24. The device according to claim 23, wherein: The calling module is also used for: If so, in response to a second configuration instruction, determining a second business object adapted to the text category from the multiple business objects, and establishing a corresponding relationship between the text category and the second business object; Dispatching the input text to the second business object so that the second business object updates the corresponding maintained agent interaction topology graph; According to the updated agent interaction topology map maintained by the second business object, an agent interaction topology map adapted to the text category is determined.

25. The device according to any one of claims 14 to 24, wherein: The first classification module is used to: Acquire at least one third reference example; wherein the third reference example includes a first reference text and a processing object of the first reference text; Generate third prompt information according to the third reference example and the input text; wherein the third prompt information is used to prompt the large language model to perform a text classification task; The large language model is called to process the third prompt information to obtain the classification result.

26. The device according to any one of claims 14 to 24, wherein: The second classification module is used to: Acquire at least one fourth reference example; wherein the fourth reference example includes a second reference text and a text category to which the second reference text belongs; Generate fourth prompt information according to the fourth reference example and the input text; wherein the fourth prompt information is used to prompt the large language model to perform a text classification task; The large language model is called to process the fourth prompt information to obtain the text category.

27. An electronic device comprising: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the text processing method based on a large language model according to any one of claims 1 to 13.

28. A non-transitory computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to enable the computer to execute the text processing method based on a large language model according to any one of claims 1-13.

29. A computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the computer program implements the steps of the text processing method based on a large language model according to any one of claims 1 to 13.

Images