A financial recommendation system and method based on large model multi-agent

Abstract

The present invention relates to a financial recommendation system and method based on large model multi-agent, the method comprising the following steps: S1: collecting customer information, product information and historical conversation data, and constructing the elements required for the financial recommendation scenario according to the collected data; S2: based on actual business operations, analyzing the golden words of financial management and historical customer conversation data, summarizing the business touch path and marketing stage, and setting multiple marketing tasks; S3: establishing multiple agent roles according to the marketing tasks, using the multi-model fusion submodule to select the appropriate agent role, and using the large and small model collaboration submodule to complete the product recommendation; S4: using the optimized WEAM algorithm to optimize each agent role; S5: performing security verification on the final output of the agent role, and transmitting the information to the digital human for broadcasting and display through the text-to-speech service. The present invention improves customer experience, reduces the burden of manual customer service, and optimizes business operation efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of financial management recommendation, and in particular to a financial management recommendation system and method based on large model multi-agent. Background Art

[0002] With the rapid development of network technologies such as cloud computing and big data, network information has exploded. Although a large amount of data contains rich resources and huge potential, the corresponding "information overload" problem has begun to trouble everyone. Therefore, how to obtain the most valuable information from massive data in the shortest time has become a hot topic and difficulty in academic and industrial research. For the banking industry, there are even more customers, and with them, there will also be a huge amount of data. Therefore, personalized financial recommendation methods are needed to deeply explore the interests of customers, discover projects or products that customers may be interested in, and recommend them to customers in a personalized way.

[0003] In the context of existing technologies, the main recommendation method is to identify potential marketing customers by providing services such as problem consultation and information query, and assist them in choosing suitable products, simulating the role of product marketers to promote purchase intentions and complete the purchase process. This is intended to reduce the operating costs of enterprises and increase product sales. However, traditional financial marketing methods mainly rely on customer service chat marketing, including customer service dialogue methods based on traditional robots and manual customer service, which have the following shortcomings:

[0004] (1) Traditional robots have limited multi-round conversation capabilities: Traditional robots rely on rules and FAQs (Frequently Asked Questions). Their processing capabilities are limited to simple scenarios and they cannot effectively deal with complex problems.

[0005] Insufficient understanding of intent: In multi-round conversations, the robot often misunderstands and gives irrelevant answers.

[0006] The question bank has great limitations: it is difficult to fully cover customer needs.

[0007] Difficulty in process design: Frequent deviations from established process execution.

[0008] (2) For manual customer service, the requirements for account managers are too high: manual customer service is required to have rich experience and high marketing capabilities, but due to the huge amount of knowledge and limited knowledge, it is difficult to effectively recommend financial products, resulting in low efficiency and high pressure on customer service. It also increases the difficulty of recruitment and limits business growth and scale.

[0009] Knowledge management is complex: manual customer service staff need to master a large amount of product knowledge and speech skills.

[0010] There are many products and materials: it is difficult for manual customer service to fully grasp them.

[0011] Professional knowledge and scripts are updated quickly: manual customer service staff find it difficult to learn and adapt in a timely manner.

[0012] Difficulty in inheriting business experience: The professionalism of manual customer service staff varies and the training costs are high.

[0013] (3) Product language and market changes are rapid: As time goes by and the business develops, both manual customer service and traditional robots find it difficult to learn new content quickly and at low cost, and are unable to fully integrate existing methods, resulting in an increase in the error rate, which affects performance and may increase customer complaints.

[0014] Therefore, it is necessary to provide a financial recommendation system and method based on large model multi-agent to improve customer experience and conversion rate, while reducing the burden of manual customer service and optimizing business operation efficiency. Summary of the invention

[0015] The purpose of the present invention is to provide a financial management recommendation system and method based on large model multi-agent, so as to improve customer experience and conversion rate, reduce the burden of manual customer service and optimize business operation efficiency.

[0016] In order to solve the problems existing in the prior art, the present invention provides a financial recommendation system based on a large model and multiple agents, comprising:

[0017] The agent component element preparation module is configured to collect customer information, product information, and historical conversation data, and construct the elements required for the financial recommendation scenario based on the collected data. The elements include financial management speech, financial recommendation scenario script, tool call set, and recommendation model;

[0018] The financial marketing task division module is configured to analyze the financial management gold medal speech and historical customer conversation data based on actual business operations, summarize the business contact path and marketing stage, and set multiple marketing tasks;

[0019] A large model agent module is configured to establish multiple agent roles according to the marketing task, each agent role is based on a preset decision-making thinking tree of banking business; the large model agent module includes a multi-model fusion submodule and a large and small model collaboration submodule;

[0020] The multi-model fusion submodule is configured to use multiple routing agents to process simultaneously, obtain the probability of different customer intentions, perform multi-model collaborative processing, and collaboratively output C=f(O1,O2,O3,…, O n ) where f is the cumulative weighting function, O1, O2, O3, O n is the probability of each customer's intention, C is the customer's final intention, select the appropriate agent role according to the customer's final intention, and map the selected agent role to the corresponding task goal;

[0021] The large and small model collaboration submodule is configured to use the large model and the small model as guidance and basis according to the features on the path of the decision-making thinking tree, dynamically extract multiple features as needed, fuse the acquired multiple features, match the fused features with the internal feature nodes in the decision-making thinking tree, and if the match is successful, recommend products based on the leaf product nodes under the feature nodes. If the match is unsuccessful, ask the customer through a dialogue and continue to acquire features until the match is successful;

[0022] The agent optimization module is configured to use the optimized WEAM algorithm, and its optimization method is as follows:

[0023] First, initialize the instructions and parameters of any agent role; set the sample problem set: Q={q1,q2,…,q n},q1,q2,…,q n For each question in the sample problem set, generate a standard answer set: A∗={a1∗,a2∗,…,a n ∗},a1∗,a2∗,…,a n ∗ are the standard answers corresponding to the first question, the second question, …, and the nth question, respectively, where n is the number of questions in the sample question set;

[0024] Secondly, let the agent character answer each question in the sample question set to generate the answer set A={a1,a2,…,a n},a1,a2,…,a n The difference between each answer of the agent role and the standard answer is calculated for the agent role's answer to the first question, the answer to the second question, and the answer to the nth question, and the loss is calculated using the cross entropy loss function: , They are the difference loss value between the first standard answer and the answer of the agent role to the first question, the difference loss value between the second standard answer and the answer of the agent role to the second question, and so on. The difference loss value between the nth standard answer and the answer of the agent role to the nth question; the loss between the answer of the agent role and the standard answer is recorded as: , is an arbitrary question; when calculating the loss, the customer feedback score Incorporating the loss function, the customer feedback score is calculated as follows: ;in, For the Feedback ratings from customers with questions, For customers The score of the answer to the question; m is the number of customers who participated in the scoring;

[0025] Calculate the weighted loss function and introduce adaptive weights , adjust the weight of each question as follows: ;

[0026] in, is the adaptive weight, is the tth iteration The weight of the question; is an adjustment factor used to control the magnitude of weight updates, is the loss between the agent's answer and the standard answer, For this agent role The answer to the question; For the The standard answer to the question; To control the weight of customer feedback in loss, For the Feedback ratings from customers with questions;

[0027] Again, the loss function is weighted; the weighted loss function is expressed as: ; Where L is the value of the weighted loss function;

[0028] Finally, based on the weighted loss function L, the gradient descent method is used to update the parameters of the agent role as follows: ;

[0029] in, are the parameters of the agent role after update at the t+1th iteration; is the parameter of the agent role at the tth iteration; is the learning rate; is the gradient of the loss function with respect to the parameters;

[0030] The safety output module is configured to perform safety verification on the final output of the intelligent character and transmit the information to the digital human for broadcast and display through the text-to-speech service.

[0031] Optionally, in the large model multi-agent based financial management recommendation system,

[0032] The roles of intelligent agents include greeting and chatting agent, information consultation agent, product recommendation agent, financial consulting agent, financial product purchasing agent, information verification agent and backup agent.

[0033] Optionally, in the large-model multi-agent-based financial recommendation system, each agent role is equipped with different golden words, professional knowledge and interface tools;

[0034] Each intelligent agent role has its own memory, planning, execution and observation capabilities, and each intelligent agent role is connected to LLM.

[0035] Optionally, in the large-model multi-agent-based financial recommendation system, a secondary sub-agent is provided for the financial consulting agent, which is used to extract and verify product information and provide feedback on the current collection situation;

[0036] For the financial product purchasing agent, a secondary sub-agent is set up to collect and verify information in the purchase scenario.

[0037] Optionally, in the large model multi-agent based financial management recommendation system, the multiple routing agents are a main agent and two secondary sub-agents, the main agent inputs 3 rounds of dialogue, one secondary sub-agent inputs 5 rounds of dialogue, and the other secondary sub-agent inputs 10 rounds of dialogue. LLM obtains the probability of different intentions of the customer through the dialogue between the main agent and the two secondary sub-agents.

[0038] Optionally, in the large-model multi-agent-based financial management recommendation system, the root node and child nodes of the decision-making thinking tree are feature nodes, and the leaf nodes of the decision-making thinking tree are product nodes.

[0039] Optionally, in the financial recommendation system based on large model and multiple agents, the large model is each agent role, the small model is an information extraction model, the large model generates dialogues with customers and extracts features from the dialogues, and the small model extracts features from the dialogues.

[0040] Optionally, in the large-model multi-agent-based financial management recommendation system, the financial management recommendation system further includes a session start module, including two scenarios of active service and passive service of the agent.

[0041] The present invention also provides a financial management recommendation method based on a large model and multiple agents, using the above financial management recommendation system, comprising the following steps:

[0042] S1: Collect customer information, product information, and historical conversation data, and build the elements required for financial management recommendation scenarios based on the collected data. The elements include financial management gold medal speech, financial management recommendation scenario script, tool call set, and recommendation model;

[0043] S2: Based on actual business operations, analyze the best financial management techniques and historical customer conversation data, summarize the business reach paths and marketing stages, and set multiple marketing tasks;

[0044] S3: Establish multiple intelligent agent roles according to the marketing task, and each intelligent agent role is based on a preset decision-making thinking tree of banking business;

[0045] S31: Use multiple routing agents to process simultaneously, obtain the probability of different customer intentions, perform multi-model collaborative processing, and collaboratively output C=f(O1,O2,O3,…, O n ), where f is the cumulative weighting function, O1, O2, O3, O n is the probability of each customer's intention, C is the customer's final intention, select the appropriate agent role according to the customer's final intention, and map the selected agent role to the corresponding task goal;

[0046] S32: Using the large model and the small model as guidance and basis based on the features on the path of the decision-making thinking tree, dynamically extract multiple features as needed, fuse the multiple features obtained, and match the fused features with the internal feature nodes in the decision-making thinking tree. If the match is successful, product recommendations are made based on the leaf product nodes under the feature nodes. If the match is unsuccessful, ask the customer through a dialogue and continue to obtain features until the match is successful;

[0047] S4: Use the optimized WEAM algorithm to optimize each agent role as follows:

[0048] S41: Initialize the instructions and parameters of any agent role; set the sample problem set: Q={q1,q2,…,q n},q1,q2,…,q n For each question in the sample problem set, generate a standard answer set: A∗={a1∗,a2∗,…,a n ∗},a1∗,a2∗,…,a n ∗ are the standard answers corresponding to the first question, the second question, …, and the nth question, respectively, where n is the number of questions in the sample question set;

[0049] S42: Let the agent character answer each question in the sample question set and generate an answer set A={a1,a2,…,a n},a1,a2,…,a n The difference between each answer of the agent role and the standard answer is calculated for the agent role's answer to the first question, the answer to the second question, and the answer to the nth question, and the loss is calculated using the cross entropy loss function: , ; The loss between the agent's answer and the standard answer is recorded as: , For any question;

[0050] When calculating losses, consider customer feedback scores Incorporating the loss function, the customer feedback score is calculated as follows: ;in, For the Feedback ratings from customers with questions, For customers The score of the answer to the question; m is the number of customers who participated in the scoring;

[0051] Calculate the weighted loss function and introduce adaptive weights , adjust the weight of each question as follows: ;

[0052] in, is the adaptive weight, is the tth iteration The weight of the question; is an adjustment factor used to control the magnitude of weight updates, is the loss between the agent's answer and the standard answer, For this agent role The answer to the question; For the The standard answer to the question; To control the weight of customer feedback in loss, For the Feedback ratings from customers with questions;

[0053] S43: Perform weighted calculation on the loss function; the weighted loss function is expressed as:

[0054] ; Where L is the value of the weighted loss function;

[0055] S44: Based on the weighted loss function L, the parameters of the agent role are updated using the gradient descent method as follows:

[0056] ;

[0057] in, are the parameters of the agent role after update at the t+1th iteration; is the parameter of the agent role at the tth iteration; is the learning rate; is the gradient of the loss function with respect to the parameters;

[0058] S5: Perform security verification on the final output of the intelligent character and transmit the information to the digital human for broadcast and display through the text-to-speech service.

[0059] Optionally, in the large model multi-agent based financial management recommendation method, the multiple routing agents are a main agent and two secondary sub-agents, the main agent inputs 3 rounds of dialogue, one secondary sub-agent inputs 5 rounds of dialogue, and the other secondary sub-agent inputs 10 rounds of dialogue. LLM obtains the probability of different intentions of the customer through the dialogue between the main agent and the two secondary sub-agents.

[0060] In the financial recommendation system and method based on large model multi-agent provided by the present invention, by introducing intelligent financial scenario-assisted marketing technology based on large models, financial consulting, product recommendation and purchase guidance services are provided in an intelligent manner, effectively replacing the roles of traditional customer service and robots in financial recommendation, so that the intelligent agent of the present invention can correctly understand in multiple rounds of dialogue and can fully cover customer needs; and because the intelligent agent of the present invention has mastered a large amount of product knowledge and golden words, it can also learn new content quickly and at low cost, so that it can flexibly talk with customers, recommend more suitable products to customers and guide purchasing methods, thereby improving customer experience and conversion rate (converting low-frequency customers into high-frequency customers), while reducing the burden of manual customer service, optimizing business operation efficiency, and increasing turnover. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 A module diagram of a financial management recommendation system provided by an embodiment of the present invention;

[0062] Figure 2 An example diagram of a decision-making thinking tree provided by an embodiment of the present invention;

[0063] Figure 3 A recommended flow chart of a large model agent module provided by an embodiment of the present invention;

[0064] Figure 4 A client layered architecture diagram provided for an embodiment of the present invention;

[0065] Figure 5 The Agent business flow chart of the financial product purchase provided by the embodiment of the present invention;

[0066] Figure 6 A flowchart of the active service of an intelligent agent provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0067] The specific implementation of the present invention will be described in more detail below in conjunction with the schematic diagram. The advantages and features of the present invention will become clearer based on the following description. It should be noted that the drawings are all in a very simplified form and are not in exact proportions, and are only used to facilitate and clearly assist in explaining the purpose of the embodiments of the present invention.

[0068] Hereinafter, if the method described herein includes a series of steps, the order in which the steps are presented herein is not necessarily the only order in which the steps may be performed, and some of the steps described may be omitted and / or some other steps not described herein may be added to the method.

[0069] In the context of existing technologies, the main recommendation method is to identify potential marketing customers by providing services such as problem consultation and information query, and assist them in choosing suitable products, simulating the role of product marketers to promote purchase intentions and complete the purchase process. This is intended to reduce the operating costs of enterprises and increase product sales. However, traditional financial marketing methods mainly rely on customer service chat marketing, including customer service dialogue methods based on traditional robots and based on manual customer service, but both traditional robot-based and manual customer service-based customer service dialogue methods have multiple shortcomings.

[0070] In order to solve the problems existing in the prior art, the present invention provides a financial recommendation system based on a large model and multiple agents, such as Figure 1 As shown, including:

[0071] The intelligent agent component element preparation module is configured to collect customer information, product information and historical conversation data, wherein customer information includes customer behavior data, customer basic information and prior knowledge such as customer habits, etc., and product information includes financial product information query system and financial product income query system, etc. The elements required for the financial recommendation scenario are constructed based on the collected data, and the elements include golden financial management speeches, financial recommendation scenario scripts, tool call sets and recommendation models; the above data for a single specific customer will be directly passed in by the system through specific slot variables during multiple rounds of conversations.

[0072] Furthermore, customer behavior data includes: recent customer purchase behavior and click behavior, which are mainly used as input parameters of the recommendation model to achieve more accurate personalized recommendations. Customer basic information includes: customer identity information and basic portraits, such as age, name, etc. Prior knowledge such as customer habits includes: preferences of different customers and the current relationship between customer groups and products, which are used to recommend cold start, improve intelligence, specialization and personification. Financial product information query system and financial product income query system include: professional financial information sources and query interfaces. Historical dialogue data: Contains all previous dialogue data with customers, which is used to build marketing scripts, refine speech, precipitate business paths, and conduct final effect verification and testing. Financial gold medal speech: refers to a series of efficient and professional communication skills and language expressions used in the communication process between financial advisors and customers, which are used for Agent background information and capability setting to help better understand customer needs, convey the characteristics of financial products, answer customer questions, and enhance customer trust and willingness to buy.

[0073] The financial marketing task division module is configured to analyze the golden words of financial management and historical customer dialogue data based on actual business operations, summarize the business contact path and marketing stage, and set multiple marketing tasks. For example, marketing tasks include: initiating topics and greetings, asking about subsequent funding arrangements, understanding preferences, product recommendations, and helping with purchases. Each task sets targeted business goals, divides specific business words, answer examples, and callable interface tools to help divide business boundaries and element data in the subsequent intelligent agent role construction stage.

[0074] The large model agent module introduces a rule model and a machine learning recommendation model to help the large model agent module make decisions. The large model agent module is configured to establish multiple agent roles according to the marketing task, and the agent role (ie, Agent) is constructed by LLM (Large Language Model).

[0075] Specifically, the agent roles include the chat agent, information consulting agent, product recommendation agent, financial consulting agent, financial product purchase agent, information verification agent, and backup agent. Each agent role is equipped with different golden words, professional knowledge, and interface tools (for example, the information consulting agent is equipped with product details query and product information query tools, and the query tools are connected to the financial product interface within the bank; the product recommendation agent is equipped with personalized recommendation models and rule recommendation models. These two recommendation models help customers find suitable products); each agent role itself has memory, planning, execution, and observation capabilities, and each agent role is connected to LLM.

[0076] Among them, all the above-mentioned financial management intelligent agent roles contain the following business data: financial management recommendation requirements description, financial management recommendation background description, financial management process decomposition, prior experience of financial management recommendations, the words of gold medal account managers, real dialogue data, dialogue scripts, stage classification data, customer portraits, behavioral data, transaction data, product information (document type), product data (structured), rule model interface, financial management recommendation model interface and other business function interfaces.

[0077] Furthermore, each intelligent agent role itself has memory, planning, execution and observation capabilities. Specifically, memory ability: the current state of the chat record, including dialogue scripts, business background, requirements, access paths, historical dialogues and dialogue data. Planning: mainly planning the entire task, including sub-script processes, task planning status, executed status and pending status. Tools: Tools are capabilities and functions that Agents can call, including callable interfaces and business scripts used, such as product revenue interfaces and product position details interfaces. Observation: mainly observing the results of tool execution and providing feedback, mainly including historical tool call results and current tool call results.

[0078] Furthermore, the prompt settings corresponding to each agent role mainly include the description and capabilities of all sub-agents, and allow LLM to determine the customer's intention based on the customer's conversation content and assign specific agent roles to perform corresponding tasks. For example: (1) The prompt settings corresponding to the greeting and chatting agent mainly include specific words, which are obtained based on the actual summary of the gold medal customer service, and include several key capabilities: 1) Reminder of product expiration. 2) Summarize the previous income situation; 3) Analyze the current market situation; these three key capabilities rely on calling the corresponding interfaces, which correspond to the customer portrait interface, the customer holding product interface, and the current market situation interface. (2) The prompt settings corresponding to the financial consulting agent mainly include a basic and simple description of the company's financial products, the overall overview of the company's products, and other knowledge; it includes several key capabilities: 1) Document information query for specific products; 2) Historical income query for specific products. (3) The prompt setting corresponding to the financial product purchase agent is mainly to guide customers to make purchases, help customers confirm their personal information, collect some information before purchase, verify purchase factor information, etc.; it includes several key capabilities: 1) customer information query; 2) customer purchase factor information extraction; 3) purchase interface awakening. (4) The backup agent is mainly to provide a backup for unrecognized customer intentions, or customer intentions that are not in the above functions, or unreasonable demands, to ensure the normal output of the large model recommendation system.

[0079] Preferably, the present invention sets logical secondary sub-agents for some key agent roles to improve processing efficiency and accuracy. (1) For the financial consulting agent, a secondary sub-agent is set up to extract and verify product information and feedback the current collection situation; different conclusions are returned based on the verification results of the consulting product information. When some information is incomplete or the product description that the customer may consult is unclear, the secondary sub-agent will confirm the code or name of the product being consulted by the customer again by asking questions to ensure that the product the customer is consulting is real; (2) For the financial product purchasing agent, a secondary sub-agent is set up to collect and verify information in the purchase scenario; when it is found that the information is wrong, missing or unclear, it will confirm with the customer again by asking questions. This approach is intended to ensure that the two sub-scenarios are handled more accurately. Through the above settings, it can be ensured that the information provided and used by customers when consulting and purchasing financial products is complete, accurate and reliable.

[0080] Further, such as Figure 2 As shown, each intelligent agent role presets a decision-making thinking tree based on banking business, and the root node and child nodes of the decision-making thinking tree are feature nodes (the child nodes are the internal feature nodes), and the leaf nodes of the decision-making thinking tree are the leaf product nodes. For example: the root node of the decision-making thinking tree is the customer's final intention feature, and the child nodes include risk tolerance features, investment period features, investment amount features, return expectation features, and liquidity features, etc. The leaf nodes include money market funds, principal-guaranteed financial products, treasury bonds / AAA-rated treasury bonds, flexible deposit and withdrawal financial products, time deposits, hybrid funds, bond funds, stock funds, high-interest bond funds, and balanced funds, etc. Preferably, all nodes of the decision-making thinking tree are set or adjusted according to the actual business.

[0081] like Figure 3 As shown, the large model agent module includes a multi-model fusion sub-module and a large and small model collaboration sub-module.

[0082] Continue to refer Figure 3 , the multi-model fusion submodule is configured with the following functions:

[0083] (1) First, multiple routing agents are used to process simultaneously to obtain the probabilities of different customer intentions. Preferably, the multiple routing agents are a main agent and two secondary sub-agents. The input of multiple routing agents to LLM only includes customer input, historical dialogues and set intention categories; the main agent inputs 3 rounds of dialogue, one secondary sub-agent inputs 5 rounds of dialogue, and another secondary sub-agent inputs 10 rounds of dialogue. LLM determines the probabilities of different intentions of the current customer through voting by the main agent and the two secondary sub-agents (for example, if all three agents determine that the customer has the intention to query financial management, the probability of querying financial management is 100%, and if one agent determines that the customer has the intention to purchase financial management, the probability of purchasing financial management is 33%). This design aims to avoid process judgment errors when the customer jumps to an intention and executes a task within a certain intention. By introducing this voting mechanism, customer intentions can be identified more accurately, and the processing effect and accuracy of multiple rounds of dialogue can be improved. Furthermore, each routing agent is constructed using LLM.

[0084] (2) Perform multi-model collaborative processing and collaborative output C = f(O1, O2, O3, …, O n ), where f is the cumulative weighting function, O1, O2, O3, O n is the probability of each customer's intention, C is the customer's final intention, and the appropriate agent role is selected according to the customer's final intention, and the selected agent role is mapped to the corresponding task goal.

[0085] Continue to refer Figure 3 , the large and small model collaboration submodule is configured with the following functions:

[0086] (1) Use large and small models as guidance and basis based on the features on the path of the decision-making thinking tree to dynamically extract multiple features as needed.

[0087] (2) The acquired multiple features are fused, and the fused features are matched with the internal feature nodes in the decision-making thinking tree. If the match is successful, the product recommendation is made based on the leaf product node under the feature node (i.e., the intelligent agent role is used to reply to the product to be recommended). If the match is unsuccessful, the customer is asked through dialogue and features are continuously acquired (the large model is each intelligent agent role, the small model is the information extraction model, the large model has a dialogue with the customer and extracts features from the dialogue, and the small model extracts features from the dialogue) until the match is successful. In one embodiment, when matching, if the customer specifies the product in the chat, the product node is directly matched. If the customer does not specify the product in the chat, the root node and the child node are matched in turn until the match is successful, and then the product recommendation is made based on the product node under the feature node.

[0088] The agent optimization module is configured to use the optimized WEAM algorithm, and its optimization method is as follows:

[0089] First, initialize the prompt instructions and parameters of any agent role; set the sample problem set:

[0090] Q={q1,q2,…,q n},q1,q2,…,q n For each question in the sample problem set, generate a standard answer set: A∗={a1∗,a2∗,…,a n ∗},a1∗,a2∗,…,a n ∗ are the standard answers corresponding to the first question, the second question, …, and the nth question, respectively, where n is the number of questions in the sample question set;

[0091] Secondly, let the agent character answer each question in the sample question set to generate the answer set A={a1,a2,…,a n}; a1, a2, …, a n The difference between each answer of the agent role and the standard answer is calculated for the agent role's answer to the first question, the answer to the second question, and the answer to the nth question, and the loss is calculated using the cross entropy loss function: , They are the difference loss value between the first standard answer and the answer of the agent role to the first question, the difference loss value between the second standard answer and the answer of the agent role to the second question, and so on. The difference loss value between the nth standard answer and the answer of the agent role to the nth question; the loss between the answer of the agent role and the standard answer is recorded as: , For any question;

[0092] When calculating losses, consider customer feedback scores Incorporating the loss function, the customer feedback score is calculated as follows: ;in, For the Feedback ratings from customers with questions, For customers The score of the answer to the question; m is the number of customers who participated in the scoring;

[0093] Calculate the weighted loss function and introduce adaptive weights , the weight of each question is adjusted according to the loss function value and customer feedback after each iteration to reflect the importance and difficulty of each question. The adaptive weight is calculated as follows: ;

[0094] in, is the adaptive weight, is the tth iteration The weight of the question; is an adjustment factor used to control the magnitude of weight updates, is the loss between the agent's answer and the standard answer, For this agent role The answer to the question; For the The standard answer to the question; To control the weight of customer feedback in loss, For the Feedback ratings from customers with questions;

[0095] Again, the loss function is weighted; the weighted loss function is expressed as: ; Where L is the value of the weighted loss function;

[0096] Finally, based on the weighted loss function L, the gradient descent method is used to update the parameters of the agent role. The updated parameters of the agent role are calculated as follows: ;in, are the parameters of the agent role after update at the t+1th iteration; is the parameter of the agent role at the tth iteration; is the learning rate; is the gradient of the loss function to the parameter. The optimized WEAM algorithm is used to adjust the prompt instructions and parameters of each agent role to minimize the total loss.

[0097] The agent optimization module performs dynamic prompt adjustments, combining adaptive weights, customer feedback, and maximization steps. Through this optimization method, the effects of each agent can be continuously improved, thereby improving the overall answer effect of the system and customer experience.

[0098] The safety output module is configured to perform safety verification on the final output of the intelligent character, and transmit the information to the digital human for broadcast and display through the text-to-speech service. Specifically, the safety output module mainly performs safety verification on the final output of the intelligent character. The module includes defined sensitive word verification, defined sensitive intent verification, and toxicity verification such as illegal and irregular violations. At the same time, a manual verification entrance can be set, and manual real-time approval verification can be adopted. Safety verification can also be performed through security operators. After the safety output module passes the verification, the reply will be given to the digital human, who will broadcast and display it.

[0099] The financial recommendation system also includes a session start module, including two scenarios: active service and passive service by the agent. When the agent actively provides service, the system introduces a rule model, and the virtual gold manager (non-agent role) initiates the session in combination with the big model; the relevant rules determine whether to actively recall and provide help, and then the big model initiates the session; the main rules in the rule model include: income reminder mark, customer recent browsing of financial products mark, customer account has idle money mark, customer holding financial products due reminder mark, recent hot search activity promotion mark, income reminder mark, and customer recent browsing of financial products mark. When the agent passively provides service, the customer actively initiates the session, and when the customer enters the chat page and enters the relevant content, the agent intervenes. In this way, the initiative and accuracy of the system service can be effectively improved to ensure that appropriate help and services are provided at the right time.

[0100] Finally, module connection and project implementation: Ultimately, through task orchestration and workflow development, the entire process is connected to achieve end-to-end business scenarios.

[0101] (1) The business architecture of the entire mobile banking client financial recommendation scenario is designed in layers, such as Figure 4 As shown, the bottom layer is the basic service layer, which mainly includes an integration platform and is responsible for the interface service of the financial recommendation business. The middle layer is the AI ​​module, which is mainly the large model intelligent agent module proposed by the present invention, including multiple intelligent agents, API service management and task scheduling management modules, etc. The interactive experience layer is mainly responsible for displaying the virtual account manager's reply to the customer in real time, mainly including TTS, ASR, card face interaction and other functions.

[0102] (2) Specific Agent implementation and the business process after serial connection are as follows Figure 5 As shown, Figure 5 The Agent business flow chart for purchasing financial products provided in the embodiment of the present invention shows how to guide customers to complete the entire purchase process. The entire process includes 5 layers, and their respective functions are as follows:

[0103] Customer interaction layer: responsible for customer interaction;

[0104] Mobile banking layer: responsible for handling customer actions and corresponding data processing.

[0105] Large model agent module: multiple agents and task flow planning.

[0106] LLM layer; the service where the large model is located.

[0107] Business integration system layer: business interfaces that need to be called for each financial recommendation scenario.

[0108] Taking the customer's financial purchase intention as an example, after the customer sends a request, the mobile banking passes the text to the large model intelligent body module through the ASR, security verification and other modules. The routing agent calls the backend LLM internally and identifies the customer's intention through the voting mechanism. When it is judged that the customer's intention is to purchase financial products, it is routed to the financial product purchase agent. After being routed to the financial product purchase agent, its internal secondary sub-agent will synchronously call LLM to extract the purchase element information. The purchase elements include: product code, purchase amount, customer account information, debit card account number, etc. The purchase element integrity is judged. If the information is judged to be complete, the complete information is handed over to the financial product purchase agent. The financial product purchase agent will generate relevant replies based on the set process using LLM to guide the purchase process, jump to the purchase page and fill in the information to help the customer complete the purchase, and the purchase process ends. If the information is incomplete, the information verification fails, and the information verification agent will return the result, prompting the customer that the information verification fails.

[0109] The present invention also provides a financial management recommendation method based on a large model and multiple agents, comprising the following steps:

[0110] S1: Collect customer information, product information, and historical conversation data. Customer information includes customer behavior data, basic customer information, and prior knowledge such as customer habits. Product information includes financial product information query system and financial product income query system. Based on the collected data, construct the elements required for the financial recommendation scenario, including financial gold medal speech, financial recommendation scenario script, tool call set, and recommendation model. For a single specific customer, the above data will be directly passed in by the system through specific slot variables during multiple rounds of conversations. For the specific content of the collected data and financial gold medal speech, please refer to the financial recommendation system based on large model multi-agent.

[0111] S2: Based on actual business operations, we analyze the best financial management techniques and historical customer conversation data, summarize the business access paths and marketing stages, and set multiple marketing tasks. For example, marketing tasks include: initiating topics and greetings, asking about subsequent funding arrangements, understanding preferences, product recommendations, and helping with purchases. Each task sets targeted business goals, divides specific business techniques, answer examples, and callable interface tools to help divide business boundaries and element data in the subsequent intelligent agent role construction stage.

[0112] S3: Establish multiple intelligent roles (i.e., Agents) based on the marketing tasks, including small talk agents, information consulting agents, product recommendation agents, financial consulting agents, financial product purchase agents, information verification agents, and backup agents. Each intelligent role is equipped with different golden words, professional knowledge, and interface tools (for example, the information consulting agent is equipped with product details query and product information query tools, and the query tools are connected to the financial product interface within the bank; the product recommendation agent is equipped with personalized recommendation models and rule recommendation models. These two recommendation models help customers find suitable products); each intelligent role itself has memory, planning, execution, and observation capabilities.

[0113] Further, such as Figure 2 As shown, each agent role presets a decision-making thinking tree based on banking business, the root node and child nodes of the decision-making thinking tree are feature nodes, and the leaf nodes of the decision-making thinking tree are product nodes. For example: the root node of the decision-making thinking tree is the customer's final intention feature, the child nodes include risk tolerance features, investment period features, investment amount features, return expectation features, and liquidity features, etc., and the leaf nodes include money market funds, principal-guaranteed financial products, treasury bonds / AAA-rated treasury bonds, flexible deposit and withdrawal financial products, time deposits, hybrid funds, bond funds, stock funds, high-interest bond funds, and balanced funds, etc. Preferably, all nodes of the decision-making thinking tree are set or adjusted according to actual business.

[0114] refer to Figure 3 , S31: The method of multi-model fusion is as follows:

[0115] (1) First, multiple routing agents are used to process simultaneously to obtain the probabilities of different customer intentions. Preferably, the multiple routing agents are a main agent and two secondary sub-agents. The input of multiple routing agents to LLM only includes customer input, historical dialogues and set intention categories; the main agent inputs 3 rounds of dialogue, one secondary sub-agent inputs 5 rounds of dialogue, and another secondary sub-agent inputs 10 rounds of dialogue. LLM determines the probabilities of different intentions of the current customer through voting by the main agent and the two secondary sub-agents (for example, if all three agents determine that the customer has the intention to query financial management, the probability of querying financial management is 100%, and if one agent determines that the customer has the intention to purchase financial management, the probability of purchasing financial management is 33%). This design aims to avoid process judgment errors when the customer jumps to an intention and executes a task within a certain intention. By introducing this voting mechanism, customer intentions can be identified more accurately, and the processing effect and accuracy of multiple rounds of dialogue can be improved. Furthermore, each routing agent is implemented using LLM.

[0116] (2) Perform multi-model collaborative processing and collaborative output C = f(O1, O2, O3, …, O n ), where f is the cumulative weighting function, O1, O2, O3, O n is the probability of each customer's intention, C is the customer's final intention, select the appropriate agent role according to the customer's final intention, and map the selected agent role to the corresponding task goal;

[0117] Continue to refer Figure 3 , S32: The coordination between large and small models is as follows:

[0118] (1) Use large and small models as guidance and basis based on the features on the path of the decision-making thinking tree to dynamically extract multiple features as needed.

[0119] (2) The acquired multiple features are fused, and the fused features are matched with the internal feature nodes in the decision-making thinking tree. If the match is successful, the product recommendation is made based on the leaf product node under the feature node (i.e., the intelligent agent role is used to reply to the product to be recommended). If the match is unsuccessful, the customer is asked through dialogue and features are continuously acquired (the large model is each intelligent agent role, the small model is the information extraction model, the large model has a dialogue with the customer and extracts features from the dialogue, and the small model extracts features from the dialogue) until the match is successful. In one embodiment, when matching, if the customer specifies the product in the chat, the product node is directly matched. If the customer does not specify the product in the chat, the root node and the child node are matched in turn until the match is successful, and then the product recommendation is made based on the product node under the feature node.

[0120] S4: Use the optimized WEAM algorithm to optimize each agent role as follows:

[0121] S41: Initialize prompt instructions and parameters of any agent role; set sample problem set: Q={q1,q2,…,q n},q1,q2,…,q n For each question in the sample problem set, generate a standard answer set: A∗={a1∗,a2∗,…,a n ∗},a1,a2,…,a n They are the agent's answer to the first question, the answer to the second question, and so on, the answer to the nth question, where n is the number of questions in the sample question set;

[0122] S42: Let the agent character answer each question in the sample question set and generate an answer set A={a1,a2,…,a n}; a1, a2, …, a n The difference between each answer of the agent role and the standard answer is calculated for the agent role's answer to the first question, the answer to the second question, and the answer to the nth question, and the loss is calculated using the cross entropy loss function: ,, They are the difference loss value between the first standard answer and the answer of the agent role to the first question, the difference loss value between the second standard answer and the answer of the agent role to the second question, and so on. The difference loss value between the nth standard answer and the answer of the agent role to the nth question; the loss between the answer of the agent role and the standard answer is recorded as: , For any question;

[0123] When calculating losses, consider customer feedback scores Incorporating the loss function, the customer feedback score is calculated as follows: ;in, For the Feedback ratings from customers with questions, For customers The score of the answer to the question; m is the number of customers who participated in the scoring;

[0124] Calculate the weighted loss function and introduce adaptive weights , the weight of each question is adjusted according to the loss function value and customer feedback after each iteration to reflect the importance and difficulty of each question. The adaptive weight is calculated as follows: ;in, is the adaptive weight, is the tth iteration The weight of the question; is an adjustment factor used to control the magnitude of weight updates, is the loss between the agent's answer and the standard answer, For this agent role The answer to the question; For the The standard answer to the question; To control the weight of customer feedback in loss, For the Feedback ratings from customers with questions;

[0125] S43: Perform weighted calculation on the loss function; the weighted loss function is expressed as: ; Where L is the value of the weighted loss function;

[0126] S44: Based on the weighted loss function L, the parameters of the agent role are updated using the gradient descent method. The updated parameters of the agent role are calculated as follows: ;in, are the parameters of the agent role after update at the t+1th iteration; is the parameter of the agent role at the tth iteration; is the learning rate; is the gradient of the loss function to the parameter. The optimized WEAM algorithm is used to adjust the prompt instructions and parameters of each agent role to minimize the total loss.

[0127] Dynamic prompt adjustment is carried out, combining adaptive weight, customer feedback and maximization step. Through this optimization method, the effect of each agent can be continuously improved, thereby improving the overall answer effect of the system and customer experience.

[0128] S5: Perform security verification on the final output of the intelligent character, and pass the information to the digital human for broadcast and display through the text-to-speech service. Specifically, the final output of the intelligent character is mainly security verified, and the verification content includes defined sensitive word verification, defined sensitive intent verification, and toxicity verification such as illegal and irregular behavior. At the same time, a manual verification entrance can be set, and manual real-time approval verification can be adopted. Security verification can also be performed through security operators. After the verification is passed, the response will be given to the digital human, who will broadcast and display it.

[0129] This method includes two scenarios: active service and passive service by intelligent agents. 1. When the intelligent agent actively provides services, a rule model is introduced, and a virtual gold medal manager (non-intelligent agent role) is used to initiate a conversation in combination with a large model; relevant rules determine whether to actively recall and provide assistance, and then the large model initiates a conversation. In one embodiment, Figure 6 As shown in the figure, the triggering conditions in the rule model include: income reminder mark, customer recent browsing financial products mark, customer account has idle money mark, customer holding financial products expiration reminder mark, recent hot search activity promotion mark, long-term inactivity mark and system update mark. The steps to actively initiate a conversation are as follows: after the triggering condition is determined (the triggering condition is as described above), the triggered scene is determined, and the corresponding scene is entered according to the triggered scene. Each scene is equipped with personalized elements and personalized scene words. The entered scene generates corresponding dialogue elements, and LLM actively dialogues according to the dialogue elements. Further, the priority order of the scene trigger is as follows: long-term inactivity, system update, hot activities, customer account change information and customer historical behavior detection; for example, if there is a customer who has not operated for a long time, then enter the long-term inactivity scene; if there is no customer who has not operated for a long time, but there is a system update, then enter the system update scene, and so on, triggering the corresponding scene. 2. When the intelligent agent is passively serving, the customer actively initiates the conversation. When the customer enters the chat page and enters the relevant content, the intelligent agent will intervene. In this way, the initiative and accuracy of the system service can be effectively improved to ensure that appropriate help and services are provided at the right time.

[0130] In one embodiment, the recommendation process of the financial recommendation method provided by the present invention is as follows: Step 1: Disassemble the process of the real bank financial customer service recommendation scenario, analyze the historical dialogue data of successful marketing, and combine the experience of gold medal customer service to comb the original process into an efficient marketing path. Step 2: According to the historical contact path, set agents of different roles, including agents of multiple slave nodes, and set a master node agent responsible for multiple rounds of dialogue and routing. Step 3: Analyze information such as basic attributes of customers, customer behaviors, and products held by customers, and form a multi-dimensional customer portrait. Step 4: Analyze the services that the bank itself can provide to the customer base, such as financial product inquiry, holding income inquiry, balance inquiry and other functions, and add corresponding functions to different role agents as tools that they can call. Step 5: In different business journeys, the virtual account manager master node automatically analyzes the needs according to the customer's historical dialogue, and allocates agent slave nodes with different professional capabilities and different roles to assist in marketing. Step 6: When it is determined that a certain agent role is processing, the agent role obtains the current business data and customer's words, and responds accordingly based on its own settings and capabilities. In this way, the agent role continues to provide services until it recommends the desired product to the customer or guides the purchase, completing the entire marketing activity.

[0131] In summary, in the financial recommendation system and method based on large model multi-agent provided by the present invention, by introducing the intelligent financial scenario-assisted marketing technology based on the large model, financial consulting, product recommendation and purchase guidance services are provided in an intelligent manner, which effectively replaces the role of traditional customer service and robots in financial recommendation, so that the intelligent agent of the present invention can understand correctly in multiple rounds of dialogue and can fully cover customer needs; and because the intelligent agent of the present invention has mastered a large amount of product knowledge and golden words, it can also learn new content quickly and at low cost, so that it can talk with customers flexibly, recommend more suitable products to customers and guide purchasing methods, thereby improving customer experience and conversion rate (converting low-frequency customers into high-frequency customers), while reducing the burden of manual customer service, optimizing business operation efficiency, and increasing turnover.

[0132] The above is only a preferred embodiment of the present invention and does not limit the present invention in any way. Any technician in the relevant technical field, without departing from the scope of the technical solution of the present invention, makes any form of equivalent replacement or modification to the technical solution and technical content disclosed in the present invention, which does not depart from the content of the technical solution of the present invention and still falls within the protection scope of the present invention.

Claims

1. A financial recommendation system based on large model multi-agent, characterized in that: include: The agent component element preparation module is configured to collect customer information, product information, and historical conversation data, and construct the elements required for the financial recommendation scenario based on the collected data. The elements include financial management speech, financial recommendation scenario script, tool call set, and recommendation model; The financial marketing task division module is configured to analyze the financial management gold medal speech and historical customer conversation data based on actual business operations, summarize the business contact path and marketing stage, and set multiple marketing tasks; A large model agent module is configured to establish a plurality of agent roles according to the marketing task, each agent role being based on a preset decision-making thinking tree of banking business; The large model agent module includes a multi-model fusion submodule and a large and small model collaboration submodule; The multi-model fusion submodule is configured to use multiple routing agents to process simultaneously, obtain the probability of different customer intentions, perform multi-model collaborative processing, and collaboratively output C=f(O1,O2,O3,…, O n ), where f is the cumulative weighting function, O1, O2, O3, O n is the probability of each customer's intention, C is the customer's final intention, select the appropriate agent role according to the customer's final intention, and map the selected agent role to the corresponding task goal; The large and small model collaboration submodule is configured to use the large model and the small model as guidance and basis according to the features on the path of the decision-making thinking tree, dynamically extract multiple features as needed, fuse the acquired multiple features, match the fused features with the internal feature nodes in the decision-making thinking tree, and if the match is successful, recommend products based on the leaf product nodes under the feature nodes. If the match is unsuccessful, ask the customer through a dialogue and continue to acquire features until the match is successful; The agent optimization module is configured to use the optimized WEAM algorithm, and its optimization method is as follows: First, initialize the instructions and parameters of any agent role; set the sample problem set: Q={q1,q2,…,q n },q1,q2,…,q n For each question in the sample problem set, generate a standard answer set: A∗={a1∗,a2∗,…,a n ∗},a1∗,a2∗,…,a n ∗ are the standard answers corresponding to the first question, the second question, …, and the nth question, respectively, where n is the number of questions in the sample question set; Secondly, let the agent character answer each question in the sample question set to generate the answer set A={a1,a2,…,a n },a1,a2,…,a n The difference between each answer of the agent role and the standard answer is calculated for the agent role's answer to the first question, the answer to the second question, and the answer to the nth question, and the loss is calculated using the cross entropy loss function: ; They are the difference loss value between the first standard answer and the answer of the agent role to the first question, the difference loss value between the second standard answer and the answer of the agent role to the second question, and so on. The difference loss value between the nth standard answer and the answer of the agent role to the nth question; the loss between the answer of the agent role and the standard answer is recorded as: , For any question; When calculating losses, consider customer feedback scores Incorporating the loss function, the customer feedback score is calculated as follows: ;in, For the Customer feedback ratings for questions, For customers For The score of the answer to each question; The number of customers who rated the participation; Calculate the weighted loss function and introduce adaptive weights , adjust the weight of each question as follows: ;in, is the adaptive weight, is the tth iteration The weight of the question; is an adjustment factor used to control the magnitude of weight updates, is the loss between the agent's answer and the standard answer, For this agent role The answer to the question; For the The standard answer to the question; To control the weight of customer feedback in loss, For the Feedback ratings from customers with questions; Again, the loss function is weighted; the weighted loss function is expressed as: ; Where L is the value of the weighted loss function; Finally, based on the weighted loss function L, the gradient descent method is used to update the parameters of the agent role as follows: ； in, are the parameters of the agent role after update at the t+1th iteration; is the parameter of the agent role at the tth iteration; is the learning rate; is the gradient of the loss function with respect to the parameters; The safety output module is configured to perform safety verification on the final output of the intelligent character and transmit the information to the digital human for broadcast and display through the text-to-speech service.

2. The financial management recommendation system based on large model multi-agent as claimed in claim 1, characterized in that: The roles of intelligent agents include greeting and chatting agent, information consultation agent, product recommendation agent, financial consulting agent, financial product purchasing agent, information verification agent and backup agent.

3. The financial management recommendation system based on large model multi-agent as claimed in claim 1, characterized in that: Each intelligent agent role is equipped with different golden words, professional knowledge and interface tools; Each intelligent agent role has its own memory, planning, execution and observation capabilities, and each intelligent agent role is connected to LLM.

4. The financial recommendation system based on large model multi-agent as claimed in claim 2, characterized in that: For the financial consulting agent, a secondary sub-agent is set up to extract and verify product information and provide feedback on the current collection situation; For the financial product purchasing agent, a secondary sub-agent is set up to collect and verify information in the purchase scenario.

5. The financial management recommendation system based on large model multi-agent as claimed in claim 1, characterized in that: The multiple routing agents are the main agent and two second-level sub-agents. The main agent receives 3 rounds of dialogue, one second-level sub-agent receives 5 rounds of dialogue, and the other second-level sub-agent receives 10 rounds of dialogue. LLM obtains the probability of different customer intentions through the dialogue between the main agent and the two second-level sub-agents.

6. The financial recommendation system based on large model multi-agent as claimed in claim 1, characterized in that: The root node and child nodes of the decision-making thinking tree are feature nodes, and the leaf nodes of the decision-making thinking tree are product nodes.

7. The financial management recommendation system based on large model multi-agent as claimed in claim 1, characterized in that: The large model is the role of each intelligent agent, and the small model is the information extraction model. The large model has a dialogue with the customer and extracts features from the dialogue, and the small model extracts features from the dialogue.

8. The financial recommendation system based on large model multi-agent as claimed in claim 1, characterized in that: The financial management recommendation system also includes a session start module, including two scenarios: active service and passive service of the intelligent agent.

9. A financial recommendation method based on large model multi-agent, characterized in that: The financial management recommendation system according to any one of claims 1 to 8 comprises the following steps: S1: Collect customer information, product information, and historical conversation data, and build the elements required for financial management recommendation scenarios based on the collected data. The elements include financial management gold medal speech, financial management recommendation scenario script, tool call set, and recommendation model; S2: Based on actual business operations, analyze the best financial management techniques and historical customer conversation data, summarize the business reach paths and marketing stages, and set multiple marketing tasks; S3: Establish multiple intelligent agent roles according to the marketing task, and each intelligent agent role is based on a preset decision-making thinking tree of banking business; S31: Use multiple routing agents to process simultaneously, obtain the probability of different customer intentions, perform multi-model collaborative processing, and collaboratively output C=f(O1,O2,O3,…, O n ), where f is the cumulative weighting function, O1, O2, O3, O n is the probability of each customer's intention, C is the customer's final intention, select the appropriate agent role according to the customer's final intention, and map the selected agent role to the corresponding task goal; S32: Using the large model and the small model as guidance and basis based on the features on the path of the decision-making thinking tree, dynamically extract multiple features as needed, fuse the multiple features obtained, and match the fused features with the internal feature nodes in the decision-making thinking tree. If the match is successful, product recommendations are made based on the leaf product nodes under the feature nodes. If the match is unsuccessful, ask the customer through a dialogue and continue to obtain features until the match is successful; S4: Use the optimized WEAM algorithm to optimize each agent role as follows: S41: Initialize the instructions and parameters of any agent role; set the sample problem set: Q={q1,q2,…,q n },q1,q2,…,q n For each question in the sample problem set, generate a standard answer set: A∗={a1∗,a2∗,…,a n ∗},a1∗,a2∗,…,a n ∗ are the standard answers corresponding to the first question, the second question, …, and the nth question, respectively, where n is the number of questions in the sample question set; S42: Let the agent character answer each question in the sample question set and generate an answer set A={a1,a2,…,a n },a1,a2,…,a n The difference between each answer of the agent role and the standard answer is calculated for the agent role's answer to the first question, the answer to the second question, and the answer to the nth question, and the loss is calculated using the cross entropy loss function: , They are the difference loss value between the first standard answer and the answer of the agent role to the first question, the difference loss value between the second standard answer and the answer of the agent role to the second question, ..., the difference loss value between the nth standard answer and the answer of the agent role to the nth question. The loss between the agent role answer and the standard answer is recorded as: , For any question; When calculating losses, consider customer feedback scores Incorporating the loss function, the customer feedback score is calculated as follows: ;in, For the Customer feedback ratings for questions, For customers The score of the answer to the question; m is the number of customers who participated in the scoring; Calculate the weighted loss function and introduce adaptive weights , adjust the weight of each question as follows: ;in, is the adaptive weight, is the tth iteration The weight of the question; is an adjustment factor used to control the magnitude of weight updates, is the loss between the agent's answer and the standard answer, For this agent role The answer to the question; For the The standard answer to the question; To control the weight of customer feedback in loss, For the Feedback ratings from customers with questions; S43: Perform weighted calculation on the loss function; the weighted loss function is expressed as: ; Where L is the value of the weighted loss function; S44: Based on the weighted loss function L, the parameters of the agent role are updated using the gradient descent method as follows: ;in, are the parameters of the agent role after update at the t+1th iteration; is the parameter of the agent role at the tth iteration; is the learning rate; is the gradient of the loss function with respect to the parameters; S5: Perform security verification on the final output of the intelligent character and transmit the information to the digital human for broadcast and display through the text-to-speech service.

10. The financial management recommendation method based on large model multi-agent as claimed in claim 9, characterized in that: The multiple routing agents are the main agent and two second-level sub-agents. The main agent receives 3 rounds of dialogue, one second-level sub-agent receives 5 rounds of dialogue, and the other second-level sub-agent receives 10 rounds of dialogue. LLM obtains the probability of different customer intentions through the dialogue between the main agent and the two second-level sub-agents.

Images