Systems and computer programs for consulting services

Abstract

We provide a multi-agent system suitable for consulting work. [Solution] A system comprising memory, a processor operably coupled to the memory, a communication interface for communicating with an artificial intelligence platform via the processor, and a user interface for receiving input queries from a user, wherein the artificial intelligence platform comprises an orchestrator agent and a plurality of specialized agents, each having its own profile. The orchestrator agent performs the following steps via the user interface: receiving input queries from a user, sending instructions based on the input queries to each of the plurality of specialized agents, receiving the output from each of the plurality of specialized agents based on the instructions, performing calculations based on each output and each profile of the plurality of specialized agents, and transmitting the results of the calculations and presenting them to the user via the user interface.

Description

Technical Field

[0001] The present invention relates to a system and a computer program for consulting services.

Background Art

[0002] The consulting business is a business that extracts problems faced by customers and supports the solution of those problems. Generally, the main business is to provide consultations and guidance, or support the formulation of plans and measures regarding matters such as business strategies, business improvement, and business development.

[0003] There are various areas related to the problems of customers (companies, government agencies, etc.) in the consulting business, covering a wide range such as management, strategy, IT, personnel, finance, and legal affairs. Therefore, there are also specialized fields in the consulting industry, and it is normal to conduct business specialized in these fields.

[0004] In recent years, in the consulting business as well, business efficiency has been improved by utilizing AI (Artificial Intelligence) agents based on large language models (LLMs) that learn specific datasets with specific algorithms. However, as described above, due to the wide variety of specialized fields in the consulting business, when using a single AI agent (hereinafter, also simply abbreviated as "agent"), the tendency of the output diverges, and the problem that it is difficult to obtain an output that is sufficiently useful in business has not been solved.

[0005] On the other hand, Non-Patent Document 1 discloses a multi-agent system that dynamically integrates knowledge across multiple domain-specific databases for content inference. In this prior art multi-agent system, each agent is configured to specialize in a specific domain and access different data subsets. Non-Patent Document 1 compares four types of connection structures between agents: decentralized, centralized, layered, and shared pool. It reports that the decentralized structure facilitates diverse communication between agents, the layered structure is advantageous when some agents refer to unrelated domains, and the shared pool structure provides stable performance with minimal variability.

[0006] Based on these findings, the introduction of multi-agent systems is being explored in consulting services as well.

[0007] However, the conventional multi-agent systems described above still have the following problems with regard to consulting work. As a premise, consulting work generally often targets a single organization, which is the client. Therefore, if the output regarding that client is inconsistent, that is, if the output that the user receives from each agent is inconsistent, contradictory, or inconsistent, it can cause significant disruption to operations.

[0008] In conventional distributed systems, even if the specialized fields are vastly different (for example, human resources and legal affairs), the respective specialized agents will exchange information. However, the number of communications for information exchange increases exponentially, and the number of parallel processes increases explosively. As a result, not only is the processing load (overhead) large, but it is also prone to falling into logical loops, making it difficult in practice to efficiently obtain consistent and coherent output regarding the customer.

[0009] In conventional centralized systems, the processing relies on a central agent that oversees specialized agents. This central agent is a bottleneck and prone to biasing the output. As a result, there is a problem in obtaining sufficient output based on diverse perspectives.

[0010] In the case of hierarchical structures related to conventional technologies, the preceding layers essentially function as filters, which makes it easy to lose the context and nuances of the user's input, resulting in the problem of difficulty in obtaining consistent output from the user's perspective.

[0011] In the case of conventional shared pool technology, as the number of specialized agents increases, the capacity of the shared pool becomes extremely large, which tends to negatively impact computational performance and makes it prone to significant noise interference. [Prior art documents] [Non-patent literature]

[0012] [Non-Patent Document 1] Yamamoto, Atsushi, et al. "Dynamic knowledge integration in multi-agent systems for content inference." Towards Agentic AI for Science: Hypothesis Generation, Comprehension, Quantification, and Validation. 2025, https: / / openreview.net / forum?id=5XNYu4rBe4 [Overview of the project] [Problems that the invention aims to solve]

[0013] Given the aforementioned problems with conventional technologies, there is a strong demand for a multi-agent system suitable for consulting work. [Means for solving the problem]

[0014] In view of the above-mentioned problems, the following can be provided in embodiments of the present invention.

[0015] Appearance 1. Memory and A processor operably coupled to the aforementioned memory, A communication interface configured to communicate with an artificial intelligence platform via the aforementioned processor, A user interface configured to receive input queries from the user and A system that includes, The aforementioned artificial intelligence platform, Orchestrator agent and Multiple types of specialized agents, each with their own profile It has, The aforementioned processor, The process of sending input queries from the user to the orchestrator agent via the user interface, The process of causing the orchestrator agent to send instructions based on the input query to each of the multiple types of specialized agents, The process involves causing the orchestrator agent to receive the output from each of the multiple types of specialized agents based on the instructions, and to perform calculations based on each output and each profile of the multiple types of specialized agents. The process involves having the orchestrator agent transmit the result of the calculation and presenting it to the user via the user interface. Configured to do A system characterized by the following features.

[0016] Appearance 2. The aforementioned processor further, The process of receiving the acceptable variability input by the user from the user interface, Based on the allowable diversity, causing the orchestrator agent to discard the outputs from each of the plurality of types of specialist agents that do not satisfy the allowable diversity and not include them in the calculation configured to perform The system according to aspect 1, characterized in that.

[0017] Aspect 3. The processor further syntax-analyzing at least a part of the input query from the received user; comparing at least a part of the syntax-analyzed input query with the respective profiles of the plurality of types of specialist agents and calculating respective semantic similarities; configured to perform, In the step of causing the orchestrator agent to receive the outputs from each of the plurality of types of specialist agents based on the instruction and perform calculations based on each output and each profile of the plurality of types of specialist agents, further performing weighting based on the calculated semantic similarity The system according to aspect 1 or 2, characterized in that.

[0018] Aspect 4. A computer program for communicating with an artificial intelligence platform, communicating with an artificial intelligence platform via a communication interface by a processor operably coupled to a memory, wherein the artificial intelligence platform has an orchestrator agent and a plurality of types of specialist agents each having a profile; receiving an input query from a user by the processor via a user interface; transmitting the input query from the user to the orchestrator agent by the processor via the user interface; The process of the processor causing the orchestrator agent to send instructions based on the input query to each of the multiple types of specialized agents, The process involves the processor causing the orchestrator agent to receive the output from each of the multiple types of specialized agents based on the instructions, and to perform calculations based on each output and each profile of the multiple types of specialized agents. The process involves the processor causing the orchestrator agent to transmit the result of the calculation and presenting it to the user via the user interface. A computer program configured to cause the processor to perform the above action. [Effects of the Invention]

[0019] According to the present invention, the multi-agent system can be specialized to suit consulting work, which is an advantage. [Brief explanation of the drawing]

[0020] [Figure 1] This is a block diagram showing an example configuration of the system according to the present invention. [Figure 2] This block diagram shows an example configuration of an AI platform in which systems according to the present invention can collaborate. [Figure 3] This block diagram shows another example configuration of an AI platform in which systems according to the present invention can collaborate. [Figure 4] This is a flowchart illustrating an example of the operation of the program according to the present invention. [Figure 5] This is a flowchart illustrating an example of the operation of the program according to the present invention. [Modes for carrying out the invention]

[0021] In this specification, "agent" refers to an artificial intelligence (AI) agent based on a large-scale language model (LLM). Multiple types of agents refer to a group of agents, each trained on different datasets using some kind of algorithm.

[0022] The system according to the present invention is a system that can be implemented on a computer, and may be implemented on multiple computers connected via a network, and may be implemented on the server side or on the client terminal side. Alternatively, the system can also be implemented on a standalone computer. In this specification, a computer may have a general hardware configuration, and mainly comprises a processor, storage device, communication interface, etc. The present invention can also provide a computer program configured to perform the operations of the system according to the present invention.

[0023] The system according to the present invention is configured to communicate with an AI platform (foundation) that implements the agent in software. The AI ​​platform is a computing resource infrastructure consisting of one or more hardware and / or software components. The AI ​​platform may consist of cloud-based and / or local computing resources (which may be servers, workstations, personal computers, mobile computers, or a combination thereof) on which the agent can run and which can communicate with external systems. In one embodiment, the system according to the present invention may be configured to communicate with an external AI platform (such as one provided as SaaS). In another embodiment, the system according to the present invention may be configured to communicate with an AI platform consisting of another computer on a local network. In yet another embodiment, the system according to the present invention may include an AI platform as an internal component.

[0024] Figure 1 is a block diagram showing the configuration of a system (100) according to a first embodiment of the present invention, and the system (100) can be implemented by hardware, software, or a combination thereof (the same consideration should be applied to the systems described below). The system (100) includes a processor (102), memory (104), a user interface (106), and a communication interface (108). The memory (104) may include RAM and storage, and is operably coupled with the processor (102) to carry out the processes according to the present invention. The user can input queries to the system (100) via the user interface (106) and can receive output from the system (100) (which may be text, images, audio, or a combination thereof). The system (100) can also communicate with an external (which may be cloud-based) AI platform (200) via the communication interface (108). Note that input queries are not necessarily limited to text, and may include, for example, audio or its transcript data.

[0025] In another embodiment, an external AI platform (200) may connect to the system (100) via a local network. In yet another embodiment, the system may include an AI platform.

[0026] Figure 2 is a block diagram showing an example configuration of an AI platform (200) that can collaborate with the system according to the present invention. In this embodiment, the AI ​​platform (200) is distributed from the system (100) and can be connected via a network. The AI ​​platform (200) may consist of any number of hardware components, and may, for example, be an architecture built on a network in a physically distributed manner. In another embodiment, the AI ​​platform (200) and the system (100) may constitute an integrated system.

[0027] The AI ​​platform (200) includes an orchestrator agent (202), a communication interface (208), and several types of specialized agents. The orchestrator agent (202) is configured to communicate with an external system (100) via the communication interface (208) and to communicate with each of the several types of specialized agents.

[0028] Multiple types of specialist agents are implemented based on LLMs that have learned from different datasets, and each is responsible for a different area of ​​expertise depending on the client being consulted. Each specialist agent has a profile that defines its characteristics (the learning process used to create the LLM, the area of ​​expertise the agent should be responsible for, the necessary prerequisite knowledge, the organizational position and role the agent will simulate, etc.). The profile can be entered by the creator when the specialist agent is created, or it may be automatically generated when the specialist agent is created. The profile may be static and unchangeable from the time of definition, or it may be dynamically and automatically modified. The description of the profile may be in a natural language that the LLM can parse, or it may be in the format of a computer language (markup language such as Markdown, scripting language, programming language, etc.). The number of specialist agents can be set arbitrarily. In one embodiment, the profile may include a configuration file (dot file) that conforms to the specifications set by the AI ​​platform for setting the behavior of the specialist agent.

[0029] The area of ​​expertise of a specialist agent can be set according to the client's situation and purpose for the consulting work. Such settings may be made, for example, based on information extracted by the user (consultant) through dialogue with the client, or they may be automatically set based on information extracted by a computer from data within the client's organization provided by the client.

[0030] While the area of ​​expertise of a specialist agent can be arbitrarily defined, it can be defined based on known industry classifications, standards (such as JIS), or the Nippon Decimal Classification (NDC) for books. For example, broadly speaking, it may be classified into areas of expertise related to external relations for the client organization (areas related to external planning, advertising, etc.), areas of expertise defined externally (areas related to laws, standards, industry standards, standard-essential patents (SEPs), etc.), and areas of expertise related to the internal workings of the client organization (areas related to human resources, labor relations, welfare, etc.).

[0031] For example, the first specialist agent (210) can be set to specialize in "planning," the second specialist agent (220) in "legal affairs," and the third specialist agent (230) in "internal affairs," and the profiles (212, 222, 232) of each specialist agent can be recorded to reflect these specializations. Each specialist agent can then build a dataset corresponding to their specialization through machine learning. For example, if the specialization is "planning," the training dataset may include data on past planning projects in the industry to which the client belongs. If the specialization is "internal affairs," the training dataset may include data on the working environment of employees within the client organization, the content of communication texts between employees (e.g., on a business chat platform), and related metadata (time, attributes of the employee who made the statement, etc.).

[0032] The orchestrator agent (202) communicates with each of the specialized agents and sends instructions to each specialized agent based on user input queries. When the orchestrator agent (202) receives the output from each specialized agent, it performs calculations to generate an output to present to the user based on those outputs and the profiles of each specialized agent. For example, this calculation may involve the orchestrator agent (202) using machine learning on the outputs and profiles of each specialized agent to generate a new output. The results of such calculations can be sent from the AI ​​platform (200)'s communication interface (208) to the system (100), and finally the user can receive the calculation results.

[0033] In this way, this system can provide consistent and diverse professional output in response to user input queries. With conventional technologies, when dealing with seemingly unrelated areas of expertise, such as "planning" and "internal affairs," the output from expert agents could be mutually exclusive or contradictory, and it became difficult to consistently consider the user's position within their organization. The configuration of this system avoids these problems of conventional technologies. In particular, the ability to consistently obtain such formatted output from agents is highly beneficial in consulting work related to planning and policy development.

[0034] In a preferred embodiment, user input may include data on acceptable diversity. Acceptable diversity is a parameter that defines the degree of diversity allowed for keywords included in the output data. For example, the orchestrator agent (202) can calculate the statistical relevance of all keywords included in the output (e.g., a plot on a multidimensional spatial coordinate system showing the distribution of semantic distances between words) by parsing all or part of the output from each specialized agent (e.g., the remainder after removing duplicate data and headers), and can quantitatively calculate the degree of variability. Such semantic distances between words can be defined based on the cosine similarity between a reference word and the word being measured, when using Word2Vec or BERT. Such calculations can be performed based on a bi-gram model.

[0035] The orchestrator agent (202) receives the acceptable diversity input by the user and can discard any output from each specialist agent that does not meet that acceptable diversity, excluding it from the above calculation (i.e., it can be excluded from the process of generating the final output delivered to the user). This helps to prevent the user from receiving inconsistent (unrelated) answers to their input queries.

[0036] In a preferred embodiment, the system (100) or AI platform (200) may parse at least a portion of the user's input queries, compare the parsed portion of the input queries with the profiles of each expert agent, and calculate their semantic similarity. The semantic similarity can be quantitatively calculated based on the semantic distance of linguistic word vectors using existing natural language processing techniques such as Word2Vec or BERT via a programming language such as Python (it may also be calculated by referring to an existing thesaurus dictionary). Furthermore, when the orchestrator agent (202) performs the above calculation, a weighting based on this semantic similarity may be imposed. This has the effect of making it easier to obtain a response from the orchestrator agent (202) that better reflects the intentions the user was seeking through the input query.

[0037] Figure 3 is a block diagram showing the configuration of a modified AI platform (300). The AI ​​platform (300) is almost identical to the AI ​​platform (200) described above, but differs in that it also includes a shared memory (305). The orchestrator agent (302) receives input queries from the user and sends instructions to the shared memory (305), and each specialist agent receives instructions from the shared memory (305) and returns an output. The orchestrator agent (302) receives the output stored in the shared memory (305), performs calculations based on each output and profile of the specialist agents, and returns an output to the system (100) based on the results.

[0038] In a configuration with shared memory (305), the orchestrator agent (302) and the group of specialized agents do not communicate directly, which has the effect of reducing the processing load on the orchestrator agent (302). The configuration of the AI ​​platform (300) is considered to be effective when the types of specialized domains required by the customer are few.

[0039] In another embodiment, the system (100) may include a shared memory, and interactions with multiple orchestrator agents or multiple AI platforms may occur via this shared memory.

[0040] The operation of this system, as described above, can be performed by a computer program. This program can be saved to storage, read into memory, and executed on the CPU, thereby enabling each of the above operations to be carried out.

[0041] Figure 4 is a flowchart illustrating an example of the operation of the program according to the present invention when using the AI ​​platform (300) shown in Figure 3. In this example, an acceptable variability is used.

[0042] First, in step S402, the orchestrator agent receives the user input query and acceptable diversity sent to the AI ​​platform. In step S404, the orchestrator agent records the above query and acceptable diversity in shared memory.

[0043] In step S406, each specialist agent receives the query and acceptable diversity from shared memory, processes them, and then returns the output and profile to shared memory. Preferably, this output includes pairs of one or more segments extracted by the specialist agent from the input query and feedback for each of those segments.

[0044] In step S408, the orchestrator agent retrieves the outputs and profiles from the specialized agents in shared memory. In step S410, the orchestrator agent performs calculations based on these outputs and profiles, returns the resulting output to the user, and the series of processes is completed.

[0045] Figure 5 is a flowchart illustrating a preferred process that may be included in step S410 of Figure 4. Note that this process can also be performed using an AI platform such as the AI ​​platform (200) that does not have shared memory.

[0046] First, in step S502, the orchestrator agent extracts from the output of each specialist agent that references the same segment (location) in the user's input query.

[0047] In step S504, the orchestrator agent compares each of the segments extracted in step S502 with the profiles of each specialist agent and calculates semantic similarity based on the natural language processing techniques described above. For each segment, it then selects the output (result) from the specialist agent whose profile has the highest semantic similarity.

[0048] In step S506, the orchestrator agent compares the output selected in step S504 with the output from the unselected specialist agents to calculate semantic similarity.

[0049] Then, in step S508, the orchestrator agent compares the semantic similarity calculated in step S506 with the acceptable diversity entered by the user, and discards any output from the specialist agents that falls below the threshold. The output that was not discarded is then output along with the corresponding references in the query. This completes the flow.

[0050] In another embodiment, steps S506 and S508 may be omitted, and the output of the specialist agent with the most similar profile to the reference location extracted in step S504 may be used as the output to the user.

Claims

1. Memory and A processor operably coupled to the aforementioned memory, A communication interface configured to communicate with an artificial intelligence platform via the aforementioned processor, A user interface configured to receive input queries from the user and A system that includes, The aforementioned artificial intelligence platform, Orchestrator agent and Each agent has its own profile and learns from different datasets, resulting in multiple types of specialized agents each handling different areas of expertise. It has, The aforementioned processor, The process of sending the input query from the user to the orchestrator agent via the user interface, The steps include parsing at least a portion of the input query received from the user, A step of comparing at least a portion of the parsed input query with the respective profiles of the multiple types of expert agents and calculating their respective semantic similarities, The process of causing the orchestrator agent to send instructions based on the input query to each of the multiple types of specialized agents, The process involves having the orchestrator agent receive the output from each of the multiple types of specialized agents based on the instructions, perform weighting based on the calculated semantic similarity, and perform calculations based on each output and each profile of the multiple types of specialized agents. The process involves having the orchestrator agent transmit the result of the calculation and presenting it to the user via the user interface. Configured to do A system characterized by the following features.

2. The aforementioned processor further, The process of receiving the acceptable variability input by the user from the user interface, Based on the aforementioned tolerance variability, the process involves instructing the orchestrator agent to discard outputs from each of the multiple types of specialized agents that do not satisfy the aforementioned tolerance variability, and to exclude them from the calculation. Configured to do The system according to claim 1, characterized in that it is the system described in claim 1.

3. A computer program for communicating with an artificial intelligence platform, A processor coupled to memory and operablely communicates with an artificial intelligence platform via a communication interface, wherein the artificial intelligence platform comprises an orchestrator agent and multiple types of specialized agents, each having a profile and each learning different datasets to handle different areas of expertise. The aforementioned processor performs the process of receiving input queries from the user via the user interface, The processor performs the steps of parsing at least a portion of the input query received from the user, The processor performs the steps of comparing at least a portion of the parsed input query with the profiles of each of the multiple types of expert agents and calculating the semantic similarity of each, The process of the processor sending input queries from the user to the orchestrator agent via the user interface, The process of the processor causing the orchestrator agent to send instructions based on the input query to each of the multiple types of specialized agents, The process involves the processor causing the orchestrator agent to receive the output from each of the multiple types of specialized agents based on the instructions, to perform weighting based on the calculated semantic similarity, and to perform calculations based on each output and each profile of the multiple types of specialized agents. The process involves the processor causing the orchestrator agent to transmit the result of the calculation and presenting it to the user via the user interface. A computer program configured to cause the processor to perform the above action.

Images