Automated System for Intellectual Property Valuation and Customized Management Consulting Using Multi-AI Agents Based on Hybrid Vector-Knowledge Graph Retrieval-Augmented Generation (RAG) and Structured A2A Communication Protocol

Abstract

The present invention relates to an automated intellectual property valuation system using hybrid search augmentation generative (RAG) and a multi-agent architecture. The present invention includes a hybrid data repository combining a vector database and a knowledge graph, and a plurality of specialized artificial intelligence agents (user proxy, RAG, valuation, strategy synthesis) that perform mutually independent roles. To solve the problems of computational overload and inference pipeline collapse experienced by a single AI model of the prior art, the present invention enables each agent to perform divided tasks. In particular, by strictly restricting communication between agents to a structured JSON artifact-based publish-subscribe (Pub / Sub) protocol, it effectively prevents control errors such as infinite loops or deadlocks at the source and maximizes the stability and processing speed of the system.

Description

Technology Field

[0001] The present invention relates to an automated architecture for intellectual property (IP) data processing and technology valuation utilizing artificial intelligence technology. More specifically, it relates to a multi-agent-based computing system and a control method thereof in which, under a search environment combining a knowledge graph and a vector database in a hybrid manner, a plurality of distributed specialized AI agents cooperate with each other based on a structured data protocol to calculate the technical and rights value of patents and automatically synthesize customized IP collateral loans and commercialization strategies. Background Technology

[0002] In today's knowledge-based economy, intellectual property rights, such as patents, function as core corporate assets, and loans secured by such rights have become a major financing method for technology innovation companies. For financial institutions to execute such loans, a valuation process is essential to quantitatively determine the technological innovation, the robustness of the scope of rights, and the market applicability of the patent under evaluation.

[0003] Recently, Large Language Models (LLM) and Search Augmented Generative (RAG) technologies based on text similarity calculations have been introduced to automate the analysis of such vast patent literature. However, a simple semantic search method (Technology 1) that relies solely on conventional vector embeddings has limitations in that it not only fails to capture complex 'structural causal relationships,' such as physical coupling relationships between mechanical components or sequential control relationships of software algorithms described in patent specifications, but also causes hallucinations due to information disconnection in situations requiring multi-hop inference, thereby seriously undermining the objectivity of the evaluation.

[0004] To overcome the limitations of such simple vector-based search, a hybrid RAG system (Technology 2) has been proposed that transforms the relationships between entities within a document into an explicit knowledge graph and searches in parallel with a vector repository. However, when a single AI model (Single-agent) receives a vast amount of prior art data and a multi-hop knowledge network derived from such an advanced hybrid RAG environment and simultaneously performs valuation operations and the generation of complex business reports, serious computational engineering problems arise. These include exceeding the limits of the model's processable context window, causing the omission of key information, or the collapse of the entire system's inference pipeline due to conflicts between heterogeneous computational tasks. The necessity of solving these architectural problems can be confirmed by the case where a technology that separates the service provision and learning functions of an AI system to evolve the system without service interruption has been recognized for its advancement. The problem to be solved

[0005] The present invention has been devised to fundamentally solve the problems of complex computing control and data processing that inevitably arise when combining the aforementioned conventional technologies.

[0006] The first objective of the present invention is to provide a hybrid information retrieval architecture that links a vector database and a knowledge graph in parallel by extracting physical / logical causal relationships from patent data to overcome the limitations of structural factual relationship matching inherent in simple vector-based RAG models.

[0007] The second objective of the present invention is to provide a 'Multi-agent Orchestration System' that physically and logically divides the roles of query generation, hybrid search, valuation operation, and report synthesis to solve the computational overload and forgetting phenomenon of a single artificial intelligence model caused by the vast context processing of a hybrid RAG environment.

[0008] At this time, the agents are executed as independent processes in a container environment with isolated namespaces, thereby limiting the use of memory and computational resources, and all interactions between agents are performed only through an asynchronous Event Bus that provides at-most-once or at-least-once delivery guarantee options. In addition, to ensure system stability, a deadlock prevention mechanism is applied that includes a Watchdog Timer to detect response delays or failures of individual agents and a Circuit Breaker to block cascading failures, and may be configured to pass through a Semantic Caching layer to increase the response speed for repetitive queries.

[0009] In particular, the third objective of the present invention is to provide a control flow technology that forces the exchange of state data between agents only through predefined structured artifacts and publish-subscribe-based Agent-to-Agent (A2A) communication protocols, rather than natural language-based chat, in order to prevent control infinite loops or message collisions between agents that frequently occur in multi-agent environments. means of solving the problem

[0010] A multi-agent artificial intelligence system for automating intellectual property valuation and management consulting according to one embodiment of the present invention for solving the above problem is,

[0011] A hybrid data store comprising a vector database that partitions text claims of a large-scale patent document and stores them by mapping them into a high-dimensional embedding vector space, and a knowledge graph that stores explicit engineering design knowledge in the head entity :: relationship :: tail entity format extracted through token classification of a finely tuned Transformer-based language model from the patent document corpus as directed nodes and edges; and

[0012] It includes at least one main processor connected via a network bus,

[0013] The main processor executes multiple specialized AI agents logically partitioned to have independent workspaces; specifically, the A2A protocol includes a topic naming convention in the 'Agt.Evt' format, rules for enforcing the consistency and validity of message payloads based on JSON schema definition files (XSD), and a role-based access control (RBAC) authentication mechanism that controls the topic-specific publish / subscribe permissions of each agent. Additionally, all message exchanges are recorded in an immutable audit log to ensure the transparency and traceability of the system's behavior, and the subscriber client can minimize the reception of unnecessary data through subscription filtering logic based on specific attribute values ​​of message headers or payloads. The data transmission guarantee level is selectively applied between 'At-Most-Once' semantics, which does not allow retries depending on the importance of the message, and 'At-Least-Once' semantics, which guarantees idempotency, and these structural constraints function as a core enforcement mechanism that fundamentally excludes the possibility of unstructured natural language interactions between agents at the system architecture level.

[0014] A user-proxy agent that receives customer context data from a client terminal, including financial status information of a target company, R&D target technology classification, and identification number of patents subject to evaluation, and converts it into a structured query in a standardized JSON format;

[0015] A RAG agent that receives a structured query from the above user proxy agent, transmits a hybrid parallel query to the vector database and knowledge graph of the above hybrid data store, and generates a hybrid context chunk by aggregating semantically similar prior art and multi-hop knowledge graph subnet information;

[0016] A valuation agent that receives only the hybrid context chunk generated by the above-mentioned RAG agent as input, runs an internally configured SMART3-based quantitative evaluation algorithm, and calculates an intellectual property value score based on statistical Stanine grade 9 by performing weighted matrix operations with the Analytic Hierarchy Process (AHP) technique applied to parameters of technological innovation, comprehensiveness of the scope of rights, and market usability; and

[0017] It is configured to include a Strategy Synthesis Agent that combines only the array of value scores calculated by the above valuation agent and the above customer context data in a limited prompt conditional manner to synthesize a final management consulting natural language report including financing strategies utilizing the intellectual property and strategies for intellectual property collateral loans.

[0018] Here, the technical feature is that the state data transmission and control flow transfer between the divided agents are performed unidirectionally only through a JSON artifact-based Publish-Subscribe A2A protocol containing predefined essential parameter attributes, while completely excluding natural language-based conversational interfaces. Effects of the invention

[0019] The multi-agent system for automating intellectual property valuation and management consulting according to the present invention provides the following unexpected and significant advantageous effects.

[0020] First, by utilizing a hybrid repository that combines a knowledge graph and a vector database, it is possible to analyze prior art in three dimensions with high recall and multi-hop inference capabilities without missing complex chain citation relationships between heterogeneous patents and control / physical causal relationships between mechanical components that conventional simple vector search systems overlooked.

[0021] Second, the exponential expansion of data caused by such hybrid combinations and the resulting exceeding of the context window of a single large language model, computational delays, and inference pipeline collapse were completely overcome through a thoroughly divided multi-agent orchestration system. By limiting the computational processing range of tasks handled by specific agents, the waste of computing resources is prevented and the hallucination rate is drastically reduced, enabling the generation of highly reliable, precise patent valuation reports suitable for submission to financial institutions.

[0022] In particular, FIG. 3 specifically illustrates the step-by-step activation and execution flow of each agent, and FIG. 4 presents an actual example of a JSON-formatted message payload exchanged between agents within this flow. Furthermore, FIGs. 5 and 6 further clarify the operating principle of the present invention through sequence diagrams showing the interaction, message delivery, and activation / deactivation sequence of agents when performing specific tasks (e.g., prior art search, patent valuation).

[0023] The system of the present invention provides a standardized API (Application Programming Interface) for interoperability with external financial institutions or client systems, using JSON (JavaScript Object Notation) as the data exchange format. Figure 4 illustrates an example of a JSON schema for an API request payload according to the present invention.

[0024] 2.1. API Request Data Structure (JSON Schema)

[0026] Each field of the JSON payload used in API requests is defined as follows.

[0028] | Field Name | Data Type | Required | Description |

[0029] | --- | --- | --- | --- |

[0030] | query_id | String | Required | UUID that uniquely identifies each request |

[0031] | client_id | String | Required | Identifier of the client (organization) calling the API |

[0032] | patent_ids | Array | Required | List of patent numbers subject to valuation. Up to 100 can be specified. |

[0033] | K | Integer | Optional | The top K similar patents for the RAG agent to search. (Default: 10) |

[0034] | timestamp | String | Required | Requested time in ISO 8601 format (e.g., “2024-07-16T10:00:00Z”) |

[0035] | parameters | Object | Optional | Additional parameters for fine-tuning the valuation model. (e.g., technology sector weighting) |

[0037] 2.2. API Request Payload Example

[0039] (1) Example of a minimum request using only required fields

[0040] `json

[0042] “query_id”: “c3d5e7a9-2e1b-4b1e-9d8c-0f2b3a4d5e6f”,

[0043] “client_id”: “financial-institute-A”,

[0044] “patent_ids”: [“KR1020230012345”, “US20220123456A1”],

[0045] “timestamp”: “2024-07-16T10:00:00Z”

[0047] `

[0049] (2) Example of a request including optional parameters

[0050] `json

[0052] “query_id”: “f8b9e0c1-3a2d-4c1d-8b9a-1f2b3c4d5e6a”,

[0053] “client_id”: “consulting-firm-B”,

[0054] “patent_ids”: [“KR1020210054321”],

[0055] “K”: 20,

[0056] “timestamp”: “2024-07-16T11:30:00Z”,

[0057] “parameters”:

[0058] “tech_field_weight”:

[0059] “semiconductor”: 1.2,

[0060] “biotechnology”: 0.8

[0062] `

[0063] Third, by strictly restricting communication and collaboration between agents to structured data formats (JSON) and publish-subscribe protocols rather than natural language, control flow errors (infinite loops, deadlocks)—which were identified as critical flaws in multi-agent systems—are fundamentally eliminated. This maximizes the stability of the system for large-scale commercialization and the speed of business computation processing. This does not merely replace established communication methods, but rather constitutes a concrete and non-obvious solution for addressing specific technical issues. Brief explanation of the drawing

[0064] FIG. 1 is a block diagram showing the overall architecture of a multi-agent-based intellectual property valuation system according to one embodiment of the present invention. FIG. 2 is a diagram showing in detail the configuration of the hybrid data storage of the present invention. FIG. 3 is a flowchart showing the control flow of an A2A protocol-based multi-agent orchestration according to the present invention. Figure 4 is a diagram showing an example of a structured artifact (JSON) used for communication between the agents. Specific details for implementing the invention

[0065] Hereinafter, the system architecture configuration and the operating mechanism of the control flow of the present invention are clearly described based on embodiments with reference to the attached drawings, to the extent that a person skilled in the art to which the present invention pertains can easily implement it. The specification of the present invention is described in detail so that a person skilled in the art can accurately understand and reproduce the invention without excessive experimentation or the addition of specialized knowledge.

[0066] At this time, the hybrid context chunk (32a) assembly algorithm of the RAG agent (32) can be expressed by the following pseudo-code.

[0067] 1. Vector Search (Top-K Cosine Retrieval): Based on the input query embedding, the top K vector chunks whose cosine similarity scores exceed a predefined threshold (similarity_threshold) are searched first within the vector database (11).

[0068] 2. Graph Traversal (Multi-hop BFS with Weighted Path Scoring): Using the core entity of the patent to be evaluated as the starting node, a path score is calculated cumulatively based on the weight assigned to each edge, and a Multi-hop BFS is performed on the knowledge graph database (12).

[0069] 3. Context Chunk Assembly: Vector search results and graph traversal results are fused according to the merge_strategy parameter (e.g., 'vector_first', 'graph_rerank'). To ensure that the total number of tokens in the assembled context does not exceed the maximum length set by the chunk_size parameter, the content is truncated or summarized according to predefined rules (e.g., newest information first, core entity-centered) to complete the final context chunk.

[0070] 4. Caching Policy: The final generated hybrid query and the resulting context chunk (32a) are stored in the semantic cache as key-value pairs and reused for subsequent identical or similar queries to improve response speed.

[0071] FIG. 1 shows the overall architecture of a system according to one embodiment of the present invention. The system largely comprises a hybrid data storage (10) and a multi-agent orchestration module (30) running on a main processor (20). The multi-agent orchestration module (30) includes a user proxy agent (31), a RAG agent (32), a valuation agent (33), and a strategy synthesis agent (34), which communicate through an event bus (40).

[0072] 1. Offline data preprocessing and hybrid data storage (10) construction module

[0073] As illustrated in FIG. 2, the hybrid data storage (10), which is the base architecture of the present invention, refines and constructs a vast number of patent database documents in an offline state prior to real-time query processing by a client. The text splitting module divides the claims and detailed description of the invention of the collected patent specification into specified chunk sizes and converts them into multidimensional latent space vectors through an encoder-based embedding model such as BERT, and loads them into a vector database (11) (e.g., Milvus).

[0074] At the same time, the knowledge graph extraction module, which is a core component of the present invention, utilizes a finely tuned sequence-to-sequence (Seq2Seq) language model to extract noun-based technical objects within the text through token classification. For example, in the sentence “The control unit receives a signal from the sensor and decelerates the rotation of the motor,” “control unit,” “sensor,” and “motor” are set as node entities, and “signal reception” and “rotation deceleration” are set as directional edges, thereby permanently storing an explicit engineering design fact database in the form of Control Unit :: Rotation Deceleration :: Motor in a graph form in a knowledge graph database (12) (e.g., Neo4j). The content and structure of this data are a core component of the invention that achieves technical effects by organically combining with hardware.

[0075] 2. A2A protocol-based multi-agent orchestration (30) execution module

[0076] FIG. 3 illustrates the control flow according to the present invention. When a client request is received, the orchestrator of the main processor (20) operates to sequentially activate four logically isolated agents (31, 32, 33, 34). All communication is controlled via a publish-subscribe (Pub / Sub) method through a central event bus (40) rather than a natural language discussion method.

[0077] [Step 1] User proxy agent (31): Receives corporate size, financial health, M&A target technology field, and registered patent number to be evaluated through GUI input from the client. This agent refines the input noise-mixed data and publishes a standardized JSON search instruction (31a) containing keys such as , as shown in FIG. 4 to the event bus (40).

[0078] [Step 2] RAG Agent (32): The RAG agent that subscribes to the above JSON instruction (31a) dynamically generates a Cypher or GraphQL query to access the hybrid repository (10). The vector database (11) indexes the top K prior art documents with the most similar technical fields through cosine similarity, and the knowledge graph database (12) assembles a hybrid context chunk (32a) by tracing the lineage of technological development connected by multi-hop relationships with the entities of the patent under evaluation. At this stage, semantic caching logic is executed to omit repetitive operations for the same query.

[0079] [Step 3] Valuation Agent (33): The valuation agent, having received the context chunk (32a) from the previous step, is executed with weights specialized in mathematical and statistical logic processing rather than language generation. By internalizing the quantification logic of the SMART3 patent analysis system, it calculates the width of the hierarchical tree structure between independent and dependent claims to derive the 'rightness' parameter, and calculates the scarcity of the latest technology nodes within the extracted knowledge graph to determine the 'technological' parameter. The derived parameters are processed with an expert knowledge-based AHP matrix and normalized into a value score array (33a) with a final Stanine grade 9 value.

[0080] [Step 4] Strategy Synthesis Agent (34): The final strategy synthesis agent creates a final report (34a) for financial institutions by fusing the quantitative score (33a) of the 9-grade system derived above with the initial business goal of the user proxy agent. At this time, the agent is subject to strict Graph-Conditioned Generation constraints that do not allow any additional imagination or external search outside of the given context, thereby preventing data falsification or illusion. Explanation of the symbols

[0081] 10: Multi-Agent Based Intellectual Property Valuation System 11: User Interface 12: System Control Unit 20: Hybrid Data Storage 30: Multi-agent Orchestration Module 31: Task Distribution Agent 32: RAG (Retrieval-Augmented Generation) Agent 32a: Hybrid Context Chunk 33: Valuation Agent 33a: Valuation Model 34: Report Generation Agent 34a: Dynamic Report Template 40: A2A (Agent-to-Agent) Communication Protocol

Claims

Claim 1 A hybrid data repository comprising, in parallel, a vector database that parses a corpus of prior patent documents collected from an external network to convert and store segmented text claims into semantic high-dimensional spatial vectors, and a knowledge graph database that stores mechanical, chemical, or logical algorithmic actions between technical components within the prior patent documents as independent entities with technical purposes within a continuous patent specification by using a token classification technique of a Transformer-based language model fine-tuned to a large-scale corpus, and by determining and extracting the directionality of the interaction relationship between the identified independent entities through a Sequence-to-Sequence (Seq2Seq) training guidance technique, thereby storing the results as an explicit engineering design fact directional network in the format of head entity :: relationship :: tail entity. and at least one main processor coupled to the hybrid data store and event bus, wherein the main processor loads a plurality of specialized artificial intelligence agents logically partitioned to have memory-independent workspaces to execute collaborative orchestration: (a) receiving customer context data from a client terminal including the financial status of a customer company, target market variables, and a specific intellectual property identification number to be evaluated,(b) a User-Proxy Agent that parses this into a noise-removed, standardized, predefined structured query object (JSON format) and publishes it to the event bus; (b) a Retrieval-Augmented Generation Agent that subscribes to the published structured query object and performs vector similarity operation-based semantic search and multi-hop structured graph search for the explicit engineering design facts in parallel on the hybrid data store, thereby assembling and returning a hybrid context chunk that aggregates semantically similar prior art text data and interaction network data; (c) a quantitative computation specialized valuation agent that receives only the assembled hybrid context chunk data, derives raw parameters regarding the entitlement, technicality, and business utility of the intellectual property subject to evaluation based on internal morphological analysis and text mining matching results, and calculates an intellectual property value score based on statistical Stanine grade 9 by performing a dot product operation on the parameter matrix with an Analytic Hierarchy Process (AHP) algorithm matrix based on expert survey data; and (d) a strategy synthesis agent that synthesizes natural language-based management consulting result document artifacts regarding the financing and intellectual property collateral loan strategies of the client company, while the probability of generating prompt parameters is limited to be strictly conditioned only to the derived intellectual property value score and the extracted hybrid context chunk data, wherein the transmission of state data and task transfer between the divided artificial intelligence agents systematically exclude natural language-based free conversation interfaces at the source, andIt prevents control infinite loops and system bottlenecks in multi-agent environments by controlling execution to be performed sequentially through a Publish-Subscribe based Agent-to-Agent (A2A) communication protocol containing only essential field values; the agents are executed as independent processes in container environments with isolated namespaces, thereby limiting the use of memory and computational resources; all interactions between agents occur exclusively through an asynchronous Event Bus providing 'at-most-once' or 'at-least-once' delivery guarantee options; a deadlock prevention mechanism is applied that includes a Watchdog Timer to detect response delays or failures of individual agents and a Circuit Breaker to block cascading failures; and the A2A communication protocol includes a Topic Naming Convention in the 'Agt.Evt' format, rules for enforcing the consistency and validity of message payloads based on a JSON Schema Definition File (XSD), and each It includes a Role-Based Access Control (RBAC) authentication mechanism that controls the agent's topic-specific Publish / Subscribe permissions, all message exchanges are recorded in an immutable Audit Log, and the subscribing client minimizes the reception of unnecessary data through Subscription Filtering logic based on specific attribute values ​​of message headers or payloads, and the structured query object comprises a query_id field, which is a UUID uniquely identifying each request; a client_id field identifying the client; a patent_ids field, which is a list of patent numbers subject to valuation that can be specified up to 100; and a K field, which is the number of top K similar patents to be searched by the RAG agent, with a default value of 10.It includes a timestamp field representing the request time in ISO 8601 format, and a parameters field representing additional parameters for fine-tuning the valuation model, including technology field weights; the RAG agent dynamically integrates and controls join search queries between the vector database and the knowledge graph database via a GraphQL or Cypher application programming interface (API); it prioritizes searching the top K vector chunks within the vector database whose cosine similarity scores exceed a predefined threshold (similarity_threshold) based on the input query embeddings; it calculates cumulative path scores based on weights assigned to each edge with the core entity of the patent under evaluation as the starting node; it performs a Multi-hop Breadth-First Search (Multi-hop BFS) on the knowledge graph database; it fuses the vector search results and graph traversal results according to the merge_strategy parameter, but completes the final context chunk by truncating or summarizing the content according to predefined rules so that the total number of tokens in the assembled context does not exceed the maximum length set by the chunk_size parameter; and structurally having the same or greater than the threshold cosine similarity When a query object is received, the re-search operation of the vector database and knowledge graph database is omitted, and the embedding result of the previous search is immediately returned from memory; the final generated hybrid query and its resulting context chunk are stored as key-value pairs in the semantic cache and reused for subsequent identical or similar queries, thereby preventing query processing response delays; and the process for calculating the rightness parameter among the raw parameters calculated by the valuation agent comprises the patent analysis module identifying the technical requirement hierarchy of independent claims and dependent claims included in the intellectual property specification subject to evaluation in a tree form,A multi-agent artificial intelligence system comprising a control operation logic that quantifies the difficulty of a design-around by inversely comparing the density of similar entities on a multi-hop knowledge graph within a hybrid context chunk returned by the RAG agent with the depth of the tree, and reflecting this in the intellectual property value score, wherein the strategy synthesis agent is subject to a Graph-Conditioned Generation constraint that does not allow any additional imagination or external search outside of the assigned context. Claim 2 delete Claim 3 delete Claim 4 delete Claim 5 delete

Images