Semantic-driven agent capability discovery method and device for agent internet

Abstract

The invention discloses a semantic-driven intelligent agent capability discovery method and device oriented to the intelligent agent Internet, and the method comprises the steps: firstly generating an intelligent agent structured portrait which covers the three-dimensional information of skills, roles and states; semantic coding is carried out on the portrait, and the portrait is converted into a high-dimensional semantic vector; then, the vector is partitioned, a codebook is generated in each subspace in a clustering mode, and the codebook comprises representative vectors and serial numbers of the representative vectors; and quantizing the sub-vectors into numbers based on a codebook, and splicing the numbers to form discrete identification codes of the agent index. And when the index is updated, incremental maintenance is executed according to the distance between the new agent sub-vector and the vector in the codebook. Meanwhile, a generative retrieval model is trained, task query is directly mapped into discrete identification codes, historical and new task samples are mixed for continuous learning in training, and stability constraints are introduced. According to the method, through semantic portraits, a quantitative index mechanism and memory enhancement continuous learning, an end-to-end retrieval and rapid updating capability discovery scheme is constructed.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of agent Internet, and particularly relates to a semantic-driven agent capability discovery method and device for agent Internet. BACKGROUND

[0002] With the rapid development of large language models, the autonomous decision-making and collaboration capabilities of AI agents in complex environments are continuously enhanced. This has given rise to the concept of agent Internet, which aims to build an open ecosystem composed of a large number of agents to support complex task collaboration across domains and platforms. In this ecosystem, capability discovery is a key prerequisite for agent collaboration, as it involves quickly retrieving agents with corresponding functions according to task requirements, thereby supporting large-scale distributed agent collaboration.

[0003] However, the existing technology still has obvious deficiencies in the capability discovery link in agent Internet: 1) semantic heterogeneity: the format of function description and the source of information of agents are highly diversified, and existing methods lack a unified representation, making it difficult to implement cross-domain and cross-platform retrieval; 2) scalability: in large-scale agent groups, existing methods often rely on direct comparison of high-dimensional semantic embeddings, resulting in high computational overhead and low retrieval efficiency; 3) dynamic evolution: the number and functions of agents in agent Internet continue to increase or update, and traditional retrieval methods lack a memory retention mechanism, which can easily lead to forgetting phenomena during continuous learning, resulting in historical agents gradually becoming unable to be correctly retrieved. SUMMARY

[0004] To address the problems in the existing technology, the application provides a semantic-driven agent capability discovery method and device for agent Internet, which aims to build an end-to-end retrieval, fast update, and stable and reliable capability discovery solution in the long-term evolution through the cooperation of semantic portrait and representation unification, quantized and incrementally maintained index mechanism, and memory-enhanced continuous learning training.

[0005] To solve the above technical problems, the application is implemented by the following technical solutions: According to a first aspect of the application, a semantic-driven agent capability discovery method for agent Internet is provided, comprising: generating a structured portrait of an agent, the structured portrait including information in three dimensions of skills and tools, roles and expertise, and state and constraints; performing semantic encoding on the structured portrait to obtain a high-dimensional semantic vector representing the capability of the agent; The high-dimensional semantic vector is divided into multiple sub-vectors, and the sub-vectors are clustered in each subspace to generate a codebook, the codebook including representative vectors of each subspace and numbers assigned to the representative vectors; Based on the codebook, each of the sub-vectors is quantized into the number of the corresponding representative vector, and the numbers of all sub-vectors are spliced to generate a discrete identification code as an index of the agent; When the index needs to be updated, based on the distance between the sub-vector of the new agent or the updated agent and the representative vector in the codebook, incremental maintenance is performed to update the codebook and the corresponding discrete identification code; The generative retrieval model is trained to directly map the received task query to the candidate discrete identification code; In the process of training the generative retrieval model, historical task samples and new task samples are mixed for continuous learning, and stability constraints are introduced to maintain the retrieval ability for historical agents.

[0006] In a possible implementation manner of the first aspect, the structured portrait of the agent is generated, including: The original description information submitted by the agent is parsed and field-mapped to convert into a unified mode; The converted information is subjected to synonym normalization, refinement completion and consistency check by using a pre-trained language model to form the structured portrait of the three dimensions.

[0007] In a possible implementation manner of the first aspect, the structured portrait is semantically encoded, specifically: The information of the three dimensions is jointly encoded by using a pre-trained language model to obtain a high-dimensional semantic vector representing the ability of the agent.

[0008] In a possible implementation manner of the first aspect, when a new agent registration and / or a change in the ability of an existing agent occurs, the index needs to be updated; The distance between the sub-vector of the new agent or the updated agent and the representative vector in the codebook is based on incremental maintenance, including: The newly generated sub-vector is compared with the existing representative vector in the same subspace in terms of distance; If the minimum distance is below a threshold, the number of the representative vector is reused to generate the discrete identification code; If all distances exceed the threshold, a new representative vector is added in the corresponding subspace of the codebook, and a new number is assigned to the new representative vector, and the discrete identification code is generated based on the new number.

[0009] In a possible implementation manner of the first aspect, the training of the generative retrieval model includes: constructing a training sample set consisting of task queries and corresponding discrete identification codes of target agents; synonym rewriting and data augmentation on the task queries; training an end-to-end generative model with the task queries as input and the corresponding discrete identification codes as output target.

[0010] In a possible implementation manner of the first aspect, when the generative retrieval model is decoded, the output format is constrained to be consistent with the code length and number range of the discrete identification code.

[0011] In a possible implementation manner of the first aspect, the continuous learning of the hybrid historical task samples and new task samples comprises: storing representative historical task queries and their successfully matched agent discrete identification codes into a memory buffer; sampling historical samples from the memory buffer and mixing them with new samples for training during the generative retrieval model training iteration; introducing a stability constraint on the output of the historical samples in the loss function to limit the drift of the generative retrieval model parameters relative to the previous version of the generative retrieval model.

[0012] According to a second aspect of the present application, a computer device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the semantic-driven agent capability discovery method for the agent internet when executing the computer program.

[0013] According to a third aspect of the present application, a computer readable storage medium is provided, which stores a computer program, wherein the computer program is executed by a processor to implement the semantic-driven agent capability discovery method for the agent internet.

[0014] According to a fourth aspect of the present application, a computer program product is provided, which is executed by a processor to implement the semantic-driven agent capability discovery method for the agent internet.

[0015] Compared with the prior art, the present application has at least the following beneficial effects: The application provides a semantic-driven intelligent agent capability discovery method for an intelligent agent Internet.

[0016] For the scale expansion problem, the application divides a high-dimensional semantic vector into multiple sub-vectors, clusters codebooks in each subspace, and then quantizes each intelligent agent into a discrete identification code composed of numbers.

[0017] For the dynamic evolution problem, the application designs an index mechanism that can be incrementally maintained.

[0018] To solve the forgetting phenomenon in continuous learning, the application uses a mixed history task sample and new task sample to perform continuous learning when training the generative retrieval model, and introduces a stability constraint.

[0019] To make the above-mentioned purposes, features and advantages of the application more obvious and easy to understand, the following preferred embodiments are described in detail below, and the accompanying drawings are described as follows. BRIEF DESCRIPTION OF DRAWINGS

[0020] To more clearly illustrate the technical solutions in the specific embodiments of the application, the following will briefly introduce the drawings needed in the description of the specific embodiments. Obviously, the drawings in the following description are some embodiments of the application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of these drawings.

[0021] Figure 1 A flow chart of a semantic-driven agent capability discovery method for an agent Internet according to an embodiment of the present application; Figure 2 A system block diagram of a semantic-driven agent capability discovery method for an agent Internet according to an embodiment of the present application; Figure 3 and Figure 4 Simulation results of an embodiment. DETAILED DESCRIPTION

[0022] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions of the present application will be described clearly and completely below with reference to the drawings. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0023] As shown in Figure 1 and Figure 2 An embodiment of the present application provides a semantic-driven agent capability discovery method for an agent Internet, which specifically comprises the following steps: S1, generating a structured portrait of an agent, the structured portrait comprising information of three dimensions of skills and tools, roles and expertise, state and constraints. The agent is an AI agent.

[0024] In a specific implementation, first, self-description information of each source agent is received through an agent registration interface; the registration information at least comprises a basic identifier, a provider, a version number, a registration timestamp, a visibility and an access control policy.

[0025] Subsequently, the agent is guided to fill in the three core dimensions according to a unified field specification: the skills and tools dimension comprises executable operations or API interfaces disclosed by the agent, such as path planning, sentiment analysis, environment monitoring, data query or device control; the roles and expertise dimension describes the functional positioning of the agent in a collaborative task, such as a planner, an assistant, a sensor, a translator or a scheduler; and the state and constraint dimension covers running conditions and resource limitations, such as memory capacity, time delay tolerance, deployment location, current load, held data and knowledge range, privacy and compliance requirements.

[0026] Then, syntax analysis and field mapping are performed on heterogeneous inputs such as natural language, key-value pairs or hierarchical configurations (such as JSON / XML), the inputs are converted into a unified mode, and missing fields are given mandatory prompts and placeholder values, and it is checked whether the minimum available field set meets the requirements.

[0027] Then, the converted information is subjected to synonym normalization, refinement completion, and consistency check using a pre-trained language model. In detail, synonym normalization and unified capability labels (e.g., merging “anomaly detection” and “anomaly identification”) are unified; refinement completion completes general descriptions into specific tools or sensor types; and consistency check ensures logical consistency between roles, tools, and constraints.

[0028] Finally, after the above processing, the system obtains a unified pattern structured agent profile and performs versioning and solidification as the basis for semantic encoding.

[0029] S2, the structured profile is subjected to semantic encoding to obtain a high-dimensional semantic vector representing the agent's capabilities.

[0030] In specific implementation, a pre-trained language model is used to jointly encode information in the three dimensions of skills and tools, roles and expertise, and state and constraints, converting the original structured fields into dense semantic vector representations. Exemplarily, the pre-trained language model is BERT or T5. The generated semantic vectors are stored in a semantic representation repository.

[0031] To improve cross-domain adaptability, the system can fine-tune the pre-trained language model on a small-scale labeled sample set to enhance the ability to recognize and distinguish professional terms, specific domain tools, and environmental constraints.

[0032] Through this step, agents with similar functions can be mapped to adjacent positions in the semantic space even if they use different description methods, achieving cross-domain and cross-platform semantic alignment.

[0033] S3, the high-dimensional semantic vector is divided into multiple sub-vectors, and the sub-vectors are clustered in each subspace to generate a codebook, which contains representative vectors of each subspace and the numbers assigned to the representative vectors.

[0034] In specific implementation, the high-dimensional semantic vector (dimension D) is first divided into multiple subspaces, each with a dimension of D / N, where N is the number of blocks, typically taking a value of 8-32 to balance compression rate and representation accuracy. Then, in each subspace, a clustering algorithm is used to train the sub-vectors of all agents to obtain K representative vectors (K can take a value of 64-256), which are called semantic anchor points and are used to depict the main capability patterns of the subspace. Exemplarily, the clustering algorithm uses K-means clustering algorithm.

[0035] Finally, the clustering results of each subspace are constructed into a lightweight codebook, which is stored and versioned for subsequent incremental updates.

[0036] S4, based on the codebook, quantize each of the sub-vectors into the number of the corresponding representative vector, and splice the numbers of all sub-vectors to generate a discrete identification code as an index of the agent.

[0037] In a specific implementation, for each sub-vector of the agent, the semantic anchor point closest to the sub-vector is found, and the number of the semantic anchor point is recorded; then, the numbers of all sub-vectors are spliced in a fixed order to generate a discrete identification code of the agent. The discrete identification code serves as a compact index of the agent for fast retrieval. Meanwhile, the system maintains an agent-discrete identification code mapping table and synchronously updates an inverted index structure for efficient table lookup.

[0038] S5, when the index needs to be updated, based on the distance between the sub-vector of the new agent or the updated agent and the representative vector in the codebook, incremental maintenance is performed to update the codebook and the corresponding discrete identification code.

[0039] In a specific implementation, when new agent registration or change of the capability of an existing agent occurs, incremental maintenance is triggered. The semantic vector of the affected agent is regenerated, and the sub-vector is obtained in the manner of step S3. The newly generated sub-vector is compared with the existing representative vector in the same subspace in terms of distance: if the minimum distance is below a threshold τ, the number of the representative vector is reused to generate the discrete identification code; if all distances exceed the threshold τ, a new representative vector is added in the corresponding subspace of the codebook, and a new number is assigned to the new representative vector, and the discrete identification code is generated based on the new number. After the local adjustment is completed, the system updates the codebook version, the agent-identification code mapping table, and the inverted index structure for fast lookup and retrieval.

[0040] S6, a generative retrieval model is trained to directly map a received task query into a candidate discrete identification code.

[0041] In a specific implementation, a training sample set composed of task queries and the discrete identification codes of the corresponding target agents is first constructed; the task query can be a natural language description or a structured parameter input. To improve the adaptability of the model to diversified expressions of tasks, a language model is used to perform synonym rewriting and data augmentation on historical task queries, and the training set is added.

[0042] Subsequently, an end-to-end generative model is trained with the task query as input and the corresponding discrete identification code as output target; when the model decodes, the output format is constrained to be consistent with the code length and number range of the discrete identification code.

[0043] The cross-entropy loss index is used for optimization during the training process, and a difficult example negative sampling strategy is introduced to enable the model to better distinguish between similar but not identical intelligent agents. After training, the generative retrieval model can quickly output a legal discrete identification code upon receiving a task query, avoiding full library comparison in high-dimensional vector space and significantly reducing online retrieval latency and computational overhead.

[0044] S7、In the process of training the generative retrieval model, historical task samples and new task samples are mixed for continuous learning, and stability constraints are introduced to maintain the retrieval ability of historical intelligent agents.

[0045] In specific implementation, the system establishes a memory buffer to store representative historical task queries and their successfully matched discrete identification codes of intelligent agents, including frequently called tasks, domain representative tasks, and easily confused samples. During the training iteration of the generative retrieval model, historical samples are sampled from the memory buffer and mixed with new samples for training, ensuring that new and old capabilities are simultaneously strengthened in training.

[0046] Stability constraints are introduced in the loss function for the output of historical samples to limit the drift of generative retrieval model parameters relative to the previous version of the generative retrieval model; at the same time, only non-core layers such as the adaptation layer are allowed to update parameters, keeping the key representation layer stable and avoiding dramatic semantic space drift. The system monitors the retention rate and retrieval accuracy of historical intelligent agents in real time, and triggers early stopping or rollback mechanisms when performance is detected to be significantly decreased, preventing forgetting from intensifying and ensuring the sustainable operation of the system in the long run.

[0047] Preferably, after completing the core process of intelligent agent capability discovery, the system ensures that the retrieval results are practical and run stably in the long term through the following mechanisms: When the generative retrieval model outputs the intelligent agent identification code, the system queries the intelligent agent-identification code mapping table to obtain the corresponding candidate intelligent agent set and returns the top M results with the highest matching degree. This output mechanism not only ensures the relevance of the retrieval results, but also provides users with appropriate selection space.

[0048] The system establishes a multi-dimensional performance monitoring system to track the following key indicators in real time: accuracy (Accuracy), recall rate (Recall@M), normalized discounted cumulative gain (nDCG), and average reciprocal rank (MRR).

[0049] When the monitoring system detects a decrease in performance indicators, it automatically triggers a multi-level correction mechanism, including small-scale incremental training, local codebook refinement, or retrieval threshold adjustment, to ensure the long-term robustness and efficiency of the system in the intelligent agent internet environment.

[0050] Existing methods are difficult to model uniformly across different sources and formats, leading to difficulties in semantic alignment and unstable retrieval. The present application encodes structured agent profiles into a shared semantic space by introducing a pre-trained language model, allowing agents with similar functions to naturally cluster in the semantic space, even if the words used to describe them are different. This improves the ability to discover across domains and modalities, and reduces the cost of manual cleaning and rule maintenance.

[0051] Considering that using high-dimensional vectors directly as retrieval identifiers would result in high storage and high comparison overhead, the present application uses subspace clustering and codebook quantization to compress each agent into a discrete identifier code composed of a small number of subcodes. During retrieval, the model directly generates the identifier code and looks up the candidates, avoiding high-dimensional neighbor comparison of the entire library and reducing latency and resource occupation in large-scale scenarios.

[0052] The present application introduces an incremental update mechanism during index maintenance. When new agents are added or existing agent capabilities change, local refinement or expansion can be performed within the affected subspace. This mechanism can reuse existing clustering centers and add new representative vectors when necessary, ensuring that existing identifier codes are stable and effective, achieving backward compatibility of the index and supporting fast access and continuous expansion of agents.

[0053] Considering that traditional continuous training can easily forget historical agent capabilities, the present application introduces a memory enhancement mechanism during training, using selective replay of representative samples and confusing samples, and combining a stability control strategy to limit unnecessary parameter fluctuations based on the previous model. This method can maintain the performance and stability of agent capability discovery in a dynamic agent internet for a long time, effectively alleviating the forgetting problem and meeting the needs of large-scale collaboration scenarios.

[0054] To verify the actual effect of the method described in the present application, simulation verification was performed in a typical scenario of the agent internet. The simulation environment was deployed on a server platform equipped with dual-channel Intel Xeon Platinum 8378C CPU, 512GB memory, and two NVIDIA A100 GPUs (80GB video memory). The system modules included three core components: semantic modeling, index construction, and generative retrieval. Agent capability descriptions were semantically encoded using the pre-trained language model BERT, generating high-dimensional semantic vectors with a dimension of 768 as input for the present application. Two types of mainstream retrieval schemes were used for comparison: i) a sparse retrieval method based on BM25; ii) a dense retrieval method based on a double encoder. The test dataset included 4000 agent instances covering typical agent internet applications such as path planning, environment perception, and multi-language interaction.

[0055] As Figure 3In terms of the precision of agent capability discovery, the application is significantly better than the comparative method in multiple key indicators. Specifically, the Recall@5 indicator reaches 0.76, which is 23.8% higher than the best baseline; the MRR@10 and nDCG@10 indicators reach 0.66 and 0.69 respectively, which are more than 41.3% and 32.8% higher respectively; the Top-1 accuracy rate reaches 0.58, while the BM25 and dense retrieval methods are only 0.35 and 0.36 respectively. These results show that the application can more accurately capture the functional similarity between agents through the semantic-driven capability representation and indexing mechanism, effectively improving the matching quality.

[0056] As Figure 4 In terms of scalability, as the size of agents increases from 1000 to 4000, the performance of traditional retrieval methods decreases, while the application relies on the semantic quantization indexing structure, and the indicators remain relatively stable. Under the 4000 agent scale, the accuracy of the application is 0.58, which is 61% higher than the BM25 method, 60% higher than the dense retrieval method, and more than 1.6 times the traditional method, proving that the application method has very good scalability in large-scale dynamic environments.

[0057] In summary, the simulation results fully show that the application not only outperforms existing retrieval schemes in terms of discovery accuracy, but also outperforms traditional methods in terms of scalability, effectively supporting large-scale, cross-domain agent collaboration applications in the agent internet environment.

[0058] In another embodiment of the application, a computer device is provided, which includes a processor and a memory, the memory being used to store a computer program, the computer program including program instructions, and the processor being used to execute the program instructions stored in the computer storage medium. The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc., which are the computing core and control core of the terminal, and are suitable for implementing one or more instructions, and are specifically suitable for loading and executing one or more instructions in the computer storage medium to implement a corresponding method process or a corresponding function; the processor in the embodiments of the application can be used for the operation of the semantic-driven agent capability discovery method for the agent internet.

[0059] In still another embodiment of the present application, the present application also provides a storage medium, specifically a computer readable storage medium (Memory), which is a memory device in a computer device, used for storing programs and data. It can be understood that the computer readable storage medium here can include the built-in storage medium in the computer device, and of course can also include the extended storage medium supported by the computer device. The computer readable storage medium provides a storage space, which stores the operating system of the terminal. Moreover, one or more instructions suitable for being loaded and executed by the processor are also stored in the storage space, and these instructions can be one or more computer programs (including program codes). It should be noted that the computer readable storage medium here can be a high-speed RAM memory, or a non-volatile memory such as at least one disk memory. One or more instructions stored in the computer readable storage medium can be loaded and executed by the processor to implement the corresponding steps of the above-mentioned embodiment of the method for discovering semantic-driven intelligent agent capability of intelligent agent Internet.

[0060] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) containing computer-usable program code.

[0061] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to produce a machine, so that the instructions executed by the computer or other programmable data processing devices generate a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one or more flows and / or blocks Figure 1 The means for implementing the functions specified in one or more flows and / or blocks.

[0062] These computer program instructions can also be stored in a computer readable storage medium, which can guide the computer or other programmable data processing devices to work in a specific manner, so that the instructions stored in the computer readable storage medium produce a manufactured product including instruction means, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1one or more processes and / or blocks Figure 1 the function specified in the one or more blocks.

[0063] These computer program instructions can also be loaded into a computer or other programmable data processing devices, so that a series of operation steps are performed on the computer or other programmable data processing devices to generate a computer-implemented process, so that the instructions executed on the computer or other programmable data processing devices provide a process for implementing the flow Figure 1 one or more processes and / or blocks Figure 1 the steps of the function specified in the one or more blocks.

[0064] The present application also provides a computer program product, which is used for executing any one of the above-mentioned semantic-driven agent capability discovery methods for an agent internet. Since the computer program product provided by the present application and the above-mentioned semantic-driven agent capability discovery method for an agent internet belong to the same inventive concept, the computer program product provided by the present application has all the advantages of the above-mentioned semantic-driven agent capability discovery method for an agent internet, and therefore the beneficial effects of the computer program product provided by the present application will not be described one by one.

[0065] In the present application, the terms "one embodiment", "some embodiments", "example", "specific example" or "some examples" etc. mean that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. In addition, the skilled person in the art can combine and combine the different embodiments or examples described in the specification and the features of the different embodiments or examples without contradiction.

[0066] Finally, it should be noted that: the above-described embodiments are only specific embodiments of the present application, which are used to illustrate the technical solutions of the present application, but not to limit them, and the protection scope of the present application is not limited thereto. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can make modifications or easily think of changes to the technical solutions recorded in the foregoing embodiments within the technical range disclosed by the present application, or make equivalent replacements to some of the technical features; and these modifications, changes or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be covered in the protection scope of the present application.

Claims

1. A semantic-driven agent capability discovery method for an Internet of Agents, characterized in that, The method comprises the following steps: generating a structured profile of an agent, the structured profile including information in three dimensions of skills and tools, roles and expertise, state and constraints; semantically encoding the structured profile to obtain a high-dimensional semantic vector representing the capabilities of the agent; dividing the high-dimensional semantic vector into multiple sub-vectors, and clustering the sub-vectors in each subspace to generate a codebook, the codebook containing representative vectors in each subspace and numbers assigned to the representative vectors; based on the codebook, quantizing each of the sub-vectors into the number of the corresponding representative vector, and concatenating the numbers of all sub-vectors to generate a discrete identification code as an index of the agent; when the index needs to be updated, performing incremental maintenance based on the distance between the sub-vectors of the new agent or the updated agent and the representative vectors in the codebook to update the codebook and the corresponding discrete identification code; training a generative retrieval model to directly map a received task query to a candidate discrete identification code; during the training of the generative retrieval model, continuously learning by mixing historical task samples with new task samples, and introducing stability constraints to maintain the retrieval capability for historical agents. 2.The semantic-driven agent capability discovery method for agent Internet according to claim 1, wherein, The method of generating a structured profile of an agent comprises: performing syntax analysis and field mapping on the original description information submitted by the agent to convert it into a unified pattern; using a pre-trained language model to perform synonym normalization, refinement completion and consistency checking on the converted information to form the structured profile in the three dimensions. 3.The semantic-driven agent capability discovery method for agent internet of claim 1, wherein, The method of semantically encoding the structured profile comprises: using a pre-trained language model to jointly encode the information in the three dimensions to obtain a high-dimensional semantic vector representing the capabilities of the agent.

4. The semantic-driven agent capability discovery method for agent internet of things according to claim 1, characterized in that, When a new agent is registered and / or the capabilities of an existing agent change, the index needs to be updated; The incremental maintenance based on the distance between the sub-vectors of the new agent or the updated agent and the representative vectors in the codebook comprises: comparing the distance between the newly generated sub-vector and the existing representative vectors in the same subspace; if the minimum distance is below a threshold, then reuse the number of the representative vector to generate the discrete identification code; if all distances exceed the threshold, then add a new representative vector in the corresponding subspace of the codebook and assign a new number to the new representative vector, and generate the discrete identification code based on the new number.

5. The semantic-driven agent capability discovery method for agent internet of things according to claim 1, characterized in that, The method of training a generative retrieval model comprises: constructing a training sample set composed of task queries and the discrete identification codes of the target agents corresponding to the task queries; performing synonym rewriting and data augmentation on the task queries; training an end-to-end generative model with the task queries as input and the corresponding discrete identification codes as output target.

6. The semantic-driven agent capability discovery method for agent internet of things according to claim 5, characterized in that, When decoding the generative retrieval model, constrain its output format to be consistent with the code length and number range of the discrete identification code.

7. The semantic-driven agent capability discovery method for agent internet of things according to claim 1, characterized in that, The method of continuously learning by mixing historical task samples with new task samples comprises: storing representative historical task queries and their successfully matched agent discrete identification codes into a memory buffer; during the training iteration of the generative retrieval model, sampling historical samples from the memory buffer and mixing them with new samples for training; Introducing a stability constraint on the outputs of historical samples in the loss function to limit the drift of the generative retrieval model parameters with respect to the previous version of the generative retrieval model.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The computer program is executed by the processor to implement the semantic-driven agent capability discovery method for the agent Internet according to any one of claims 1 to 7.

9. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 8. The computer program is executed by the processor to implement the semantic-driven agent capability discovery method for the agent Internet according to any one of claims 1 to 7.

10. A computer program product, characterised in that, The computer program product is executed by the processor to implement the semantic-driven agent capability discovery method for the agent Internet according to any one of claims 1 to 7. The computer program product is executed by the processor to implement the semantic-driven agent capability discovery method for the agent Internet according to any one of claims 1 to 7.

Images