System and method for providing professional information based on generative ai with enhanced reliability and safety
The professional information provision system addresses hallucinations and regulatory ambiguities in generative AI by using a multi-agent structure with hybrid search and verification modules, ensuring accurate and compliant responses in specialized fields.
Patent Information
- Application Number
- KR1020250208485
- Authority / Receiving Office
- KR · KR
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2025-12-23
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2045-12-23
AI Technical Summary
Conventional generative AI systems face issues with hallucination, lack of personalized long-term memory, ambiguity in regulatory boundaries, and suboptimal search results, posing risks in specialized knowledge fields like healthcare and finance.
A professional information provision system with a supervisor module, generation module, and verification module to ensure accuracy, compliance, and personalized responses, using a multi-agent structure that includes hybrid search methods and dynamic weight adjustments.
Provides fact-based expert information, minimizes legal and ethical risks, and offers customized responses by dynamically adjusting search weights based on query attributes, ensuring regulatory compliance and user-specific context.
Smart Images

Figure 112025146219778-PAT00001_ABST
Abstract
Description
Technology Field
[0001] The present invention relates to the field of artificial intelligence technology, and more specifically, to a technology that provides specialized information requiring high accuracy and safety, such as in healthcare, law, and finance, using a Generative AI Model and a multi-agent structure. Background Technology
[0002] The recent emergence of Large-Scale Language Models (LLMs) has significantly improved natural language understanding and generation performance, and attempts are being made to utilize them in specialized knowledge fields to provide answers to user queries. However, conventional generative AI-based systems have the following limitations.
[0003] First is the phenomenon of hallucination, where non-existent information is presented as fact. In the field of specialized information, high reliability is required because even minor errors can cause economic or physical harm to users.
[0004] Second, the boundary between providing information and regulated activities (e.g., medical practice, legal consultation, etc.) is ambiguous. If AI uses definitive language or makes direct prescriptions or judgments, there is a possibility of violating relevant laws.
[0005] Third, there is a lack of a personalized long-term memory structure that continuously manages and reflects the user's individual background (underlying conditions, preferences, past history, etc.).
[0006] Fourth, when searching external databases, using a fixed search method (keyword, vector, etc.) often results in failure to derive optimal search results according to the characteristics of the query.
[0007] Therefore, there is an urgent need for an integrated system architecture capable of verifying the accuracy of expertise, determining regulatory compliance, and dynamically reflecting the unique context of users. Prior art literature
[0008] Korean Patent Publication No. 10-2023-0030511 (Published on March 6, 2023) The problem to be solved
[0009] The problem that the present invention aims to solve is to ensure the reliability of information by identifying and removing illusory sentences that occur during the process of providing professional information, and to enhance the safety of the system by detecting and replacing dangerous sentences based on safety policies and regulatory standards. means of solving the problem
[0010] The professional information provision system of the present invention for solving the above problem includes a supervisor module, a generation module, and a verification module.
[0011] The supervisor module receives and analyzes natural language queries from users and dynamically orchestrates the execution status, order, and number of iterations for additional actions (search, regenerate, edit, etc.) required to generate answers. In particular, if the verification result is unsatisfactory, it directs re-execution within the dynamically set number of iterations to improve the completeness of the final information.
[0012] The generation module creates a draft response using 'integrated information' that combines the query with the user's unique information (profile, preferences, goals, risk sensitivity, etc.) or external database search results. In this process, the external database search combines two or more keyword-based, graph-based, and vector-based semantic searches, and the results are reordered by applying dynamic weights based on query attributes and source reliability under the control of the supervisor module.
[0013] The verification module performs a two-step verification to determine whether regulatory standards have been violated by referring to the safety policy database.
[0014] Detect sentences that may violate regulations, such as diagnoses or prescriptions, and delete or replace them with alternative phrases.
[0015] Identify sentences that conflict with the integrated information or lack basis, and modify them into general explanations or remove them.
[0016] Furthermore, when new unique user information is discovered, the present invention updates the unique information database asynchronously based on the verification module's determination to allow long-term storage, and protects personal information by applying anonymization or storage restriction policies. Each module is implemented in a structure that repeatedly performs reasoning and action through function-specific sub-prompts. Effects of the invention
[0017] According to the present invention, the following effects can be expected.
[0018] It provides fact-based expert information through a hallucination verification process that compares with external evidence (integrated information).
[0019] It minimizes legal and ethical risks by automatically modifying risk expressions based on the safety policy database and providing safety guidance templates when non-conformity judgments are repeated.
[0020] It provides customized information that reflects the user's past history and risk sensitivity, and enables merging based on time-of-date information.
[0021] By dynamically adjusting search weights based on query attributes, optimal references are secured in both technical terms and general language.
[0022] By adopting a supervisor-based multi-agent structure, it can be universally applied to various professional fields such as healthcare, law, and finance. Brief explanation of the drawing
[0023] FIG. 1 is a block diagram showing the overall configuration of a professional information provision system according to one embodiment of the present invention. FIG. 2 is a flowchart showing the step-by-step flow of a professional information provision method according to one embodiment of the present invention. Specific details for implementing the invention
[0024] Hereinafter, embodiments of the present invention will be described in detail with reference to the attached drawings. The present invention is not limited to the following embodiments and various modifications are possible by those skilled in the art. Furthermore, terms such as “part,” “module,” “agent,” “terminal,” and “database” used herein refer to hardware or software that performs a specific function, or a combination thereof, and multiple modules may be implemented together in a single physical device, or a single module may be distributed and implemented across multiple physical devices.
[0025] 1. Definition of Terms
[0026] First, the key terms used in the present invention are defined.
[0027] (1) “Generative artificial intelligence model” means a deep learning-based model trained to receive natural language input (text, optionally images / tables / numeric strings) and generate a corresponding natural language output (answer sentences, summaries, instructions, etc.). For example, a Transformer-based autoregressive language model, an instruction tuning model, or a model including a reasoning process may be included.
[0028] (2) “Expert information” includes knowledge or guidance on a specific domain (e.g., healthcare, law / patent, finance / tax, industrial technology, compliance, etc.), but includes (i) content whose basis can be verified from external databases or internal knowledge, and (ii) content provided in a form that complies with regulatory standards and service policies.
[0029] (3) “Supervisor module (110)” means a module that receives a user query, analyzes the query and produces an analysis result, and dynamically controls whether to execute additional actions (e.g., additional search, prompt reconstruction, answer regeneration, verification strength adjustment, alternative response switching), the order of execution, and the number of repetitions based on the analysis result.
[0030] (4) “Generation module (120)” means a module that generates a draft answer using a generative artificial intelligence model with the query and integrated information under the control of the supervisor module (110). The generation module (120) can generate metadata including “sentence-reason mapping” or “source tags” along with the draft answer if necessary.
[0031] (5) “Verification module (130)” means a module that, under the control of the supervisor module (110), verifies whether the draft answer is factually grounded and whether it complies with policies by referring to integrated information, and determines the final professional information by modifying or approving the draft answer based on the verification results. The verification module (130) can identify false (unfounded or conflicting) sentences and dangerous sentences and perform modification processing such as deletion, substitution, or generalization.
[0032] (6) “User’s unique information” refers to information including at least one of profile information, preferences, goals, constraints, past interaction history, and risk sensitivity stored per user, and may include structured data (key-value, JSON, table) and / or unstructured data (summary text, embedding vector, document snippet).
[0033] (7) “External database (130a)” means a knowledge repository including at least one of a professional document repository (guidelines / regulations / papers / manuals / FAQ), graph structure data (entity-relationship knowledge graph), and / or a vector index (embedding-based semantic search).
[0034] (8) “Integrated information” means a set of information that includes (i) search results (documents / paragraphs / snippets / nodes / relationship paths, etc.) from an external database (130a) and / or (ii) results that selectively reflect parts related to the query context from user-specific information, and is used as an input context for the generation module (120) to generate a draft answer.
[0035] (9) “Hybrid search” is a search method that combines two or more of keyword-based search (e.g., inverted index and BM25 / TF-IDF), vector-based semantic search (e.g., embedding similarity), and optionally graph-based search (entity-relationship path search) in parallel or stepwise, and dynamically adjusts the weight of each search method according to query characteristics (slang, ambiguity, need for recency, need for relationship inference, etc.) to select the final reference document.
[0036] 2. System Configuration
[0037] Referring to FIG. 1, a professional information providing system (100) according to one embodiment of the present invention may include a user terminal (10), a network (20), and one or more servers (200). The user terminal (10) may be implemented as a smartphone, tablet, desktop PC, web client, etc., and includes a user interface (UI) that allows a user to input a natural language query and receive and display final professional information.
[0038] A supervisor module (110), a creation module (120), and a verification module (130) may be deployed in the server (200). Additionally, the server (200) may include a user unique information storage unit (121), an external database (130a), a safety policy database (150a), and (optional) a message queue / event bus (170). Each module may be implemented as a single process, and in other embodiments, may be deployed separately as a container-based microservice or serverless function and interconnected via asynchronous communication.
[0039] 3. Overall Operation Flow
[0040] Referring to FIG. 2, the system (100) according to the present invention provides professional information in approximately the following flow.
[0041] S210: Receive a natural language query from the user terminal (10).
[0042] S220: The supervisor module (110) analyzes the query and produces an analysis result.
[0043] S230: Based on the analysis results, the supervisor module (110) dynamically determines (i) whether / scope of user-specific information is reflected, (ii) selection / combination of search methods for the external database (130a), (iii) prompt reconstruction, (iv) verification strength, (v) number of repetitions and suspension / alternative response conditions.
[0044] S240: The generation module (120) uses the query and integrated information to generate a draft answer as a generative artificial intelligence model.
[0045] S250: The verification module (130) verifies the draft answer in terms of policy compliance and factual basis by referring to the integrated information, and determines the final professional information by modifying or approving the draft answer based on the verification results.
[0046] S260: The supervisor module (110) controls additional search / regeneration / range reduction / alternative response switching when necessary, and improves reliability within a pre-set or dynamically set number of iterations.
[0047] S270: Provides final professional information to the user terminal (10).
[0048] 4. Query analysis and workflow control of the supervisor module (110)
[0049] When the supervisor module (110) receives a query, it can produce the following query attributes (analysis results).
[0050] · Domain Attributes: Classification into Healthcare / Legal / Patent / Finance / Technology / Compliance, etc.
[0051] Risk / Regulatory Attributes: Potential for Misleading (Medical, Legal, Investment), Prohibited Subjects, Potential for Privacy / Rights Infringement
[0052] Specificity / Expertise Attributes: Proportion of technical terms, density of proper nouns, presence of mention of figures, standards, or legal provisions
[0053] Ambiguity attribute: Whether additional questions are needed
[0054] Recency Attribute: Whether recently revised regulations or the latest guidelines are required
[0055] Personalization Need Attribute: Need to Reflect Unique User Information (High / Medium / Low)
[0056] Relationship Inference Attributes: “Requirement-Exception”, “Cause-Effect”, “Relationship between A and B”, etc. Whether graph traversal is necessary
[0057] The supervisor module (110) can dynamically control, based on the analysis results above, (i) whether to call user unique information lookup / summary (121), (ii) combination and weight of hybrid search strategies (keyword / vector / graph), (iii) output format to be provided to the generation module (120) (summary / step / caution / disclaimer), (iv) priority of the verification module (130) (policy priority vs. factual priority vs. both reinforced), and (v) action in case of non-compliance (additional search, regeneration, scope reduction, template replacement).
[0058] 5. User unique information storage unit (121) and optional reflection
[0059] The user unique information storage unit (121) stores profiles, preferences, goals, constraints, history, risk sensitivity, etc., for each user by item. The supervisor module (110) may omit the reflection of unique information for queries with low need for personalization, and may selectively extract only relevant items for queries where personalization is important and include them in the integrated information.
[0060] In one embodiment, unique information may be distributed and stored in (a) a relational DB table (structured), (b) a document-type repository (summary log), and (c) a vector DB (embedding), and the supervisor module (110) may provide only the “user context related to this query” to the generation module (120) in a compressed summarized form with a limited number of tokens.
[0061] Additionally, the system (100) may add or update unique information newly identified during a conversation to the storage unit (121). At this time, the update may be limited to (i) priority of the user's explicit statement, (ii) priority of the latest timestamp, and (iii) reflection of only information that the verification module (130) has determined to be 'allowed for long-term storage'.
[0062] 6. External database (130a) and hybrid search algorithm
[0063] Document Repository (130a-1): statutes / notices / guidelines / papers / manuals / FAQs / policy documents, etc.
[0064] Vector index (130a-2): Document or paragraph / chunk embedding index (e.g., ANN / HNSW)
[0065] Knowledge Graph (130a-3): A graph in which entities (people / organizations / laws / requirements / exceptions / drugs / diseases / technological elements, etc.) and relationships (citations / exceptions / requirements fulfilled / contraindications / substitutions / super-subordinates, etc.) are represented as nodes / edges.
[0066] Recency / Reliability Metadata (130a-4): Date of creation, revision history, publisher, document type, baseline reliability score
[0067] 6.1 Query Preprocessing and Entity Ratio-Based Dynamic Weight Calculation
[0068] Before performing a hybrid search, the system (100) can perform the following on the query and selected user unique information summary.
[0069] Normalization: Standardization of notation (abbreviations / full names), standardization of units, correction of typos
[0070] Entity Recognition (NER) and Entity Linking: Extracting proper nouns / technical terms as entity candidates and mapping them to graph nodes or ontologies.
[0071] Query Decomposition: In the case of complex queries, they are broken down into sub-questions (requirements, exceptions, procedures, risks, etc.).
[0072] In addition, the “professional entity ratio” within the query can be calculated. For example, among all tokens or phrases, the ratio of (i) entities capable of matching, such as article numbers, standards, drug names, disease names, patent numbers, and institution names, can be calculated as r_entity, and the ratio of life descriptions, situation descriptions, and abstract expressions can be calculated as r_narrative.
[0073] In one embodiment, if r_entity is greater than or equal to a threshold (e.g., 0.7), the keyword-based search weight α can be set large and the vector-based search weight β small, and if r_narrative is large, the opposite can be set. Additionally, if the relationship inference attribute is high, the graph-based search weight γ can be further increased. For example,
[0074] α = f1(r_entity), β = f2(r_narrative), γ = f3(relational inference property)
[0075] Weights can be determined dynamically as shown, and can be normalized to α+β+γ=1.
[0076] 6.2 Keyword-based search (lexical retrieval)
[0077] Keyword-based search can be performed using an inverted index structure, and BM25 or TF-IDF scores (score_keyword) can be calculated. In particular, in domains where articles, numbers, and terms are important, such as law and patents, the weight of keyword-based search can be set relatively high.
[0078] 6.3 Vector-based Semantic Retrieval
[0079] Vector-based search can calculate a score_vector based on the similarity (cosine similarity, dot product, etc.) between the query embedding vector and the document / chunk embedding vector. It is advantageous for finding semantically similar paragraphs in ambiguous natural language description-centered queries and can be performed with low latency using an ANN index.
[0080] 6.4 Graph-based search (relationship / graph retrieval)
[0081] Graph-based search can perform at least one of the following in an entity-relationship graph (130a-3).
[0082] Path search: Explores relationship paths within 1-hop or 2-hop (or limited depth) from a specific entity node to expand related entity / document candidates.
[0083] Constrained search: Prioritizes specific relationship types, such as “requirement-exception” relationships and “ban-substitution” relationships.
[0084] Backtracking: Finding similar nodes through vector search and tracing back to the parent concept or basis node.
[0085] The graph-based search results preserve the relationship path itself (e.g., [Entity A]—(Exception)—[Entity B]—(Requirement)—[Entity C]) as metadata so that the generation module (120) or the verification module (130) can use it as a basis.
[0086] 6.5 Hybrid score integration and reflection of source reliability / latest
[0087] For each candidate document / chunk, after calculating the keyword score (score_keyword), vector score (score_vector), and graph score (score_graph), the combined score can be calculated in the following form.
[0088] base_score = α·score_keyword + β·score_vector + γ·score_graph
[0089] In addition, weights based on document source and recency can be reflected.
[0090] w_source: Basic reliability weights based on publisher / document type
[0091] w_fresh: Newness (revision / publication date) weight
[0092] penalty_dup: Penalty for repeating duplicate / similar content
[0093] For example,
[0094] final_score = (w_source · w_fresh) · base_score - penalty_dup
[0095] The final score can be calculated in the form shown above. Here, w_fresh can be set so that its influence increases as the freshness attribute increases.
[0096] 6.6 Candidate selection, re-ranking, diversity control (MMR, etc.)
[0097] After selecting the top K candidates from the hybrid search results, they can be reordered through a re-ranking module. The re-ranking module can be implemented in at least one of the following.
[0098] (a) Relevance evaluation based on a trained ranking model or cross-encoder
[0099] (b) Sentence-query consistency evaluation using small language models (sLM)
[0100] (c) Rule-based reordering (e.g., documents containing provisions first, newest revision first)
[0101] In addition, to reduce duplication of the same topic and secure diverse evidence, Maximal Marginal Relevance (MMR) or similar diversity selection criteria can be applied. This can be implemented by pushing candidates to a lower priority if they are highly relevant but excessively similar to already selected documents.
[0102] 6.7 Chunk (Paragraph) Extraction and Creation of “Evidence Packages”
[0103] Considering the LLM context window limitations, it is preferable to extract only the paragraphs or chunks most relevant to the query from the document instead of delivering the entire final selected document. The following metadata may be added to each chunk.
[0104] Document ID, Document Type, Publisher, Date of Creation / Revision, Reliability, Citation Availability
[0105] Graph path (if any), query subitem mapping tags
[0106] The hybrid search results are finally packaged into structured integrated information including a “query summary,” “relevant user-specific information summary (optional),” and “foundation snippet / relationship path and source metadata” and provided to the generation module (120).
[0107] 7. Configuration of prompts and generation of answer drafts for the generation module (120).
[0108] The generation module (120) configures prompt data to be input into a generative artificial intelligence model using integrated information and queries under the control of the supervisor module (110). The prompt data may include, for example, the following.
[0109] System Instructions: Source-based answers, limit assertive statements, safety / policy compliance, mark “Unknown” when uncertain
[0110] Context: Integrated information (foundation snippet, source metadata, relationship path)
[0111] User Query: Original or Normalized Query
[0112] Output format instructions: Summary / Step-by-step / Notes / Disclaimers / Induce additional questions, etc.
[0113] After receiving the draft answer, the generation module (120) can generate “ground mapping metadata” on a sentence or paragraph basis. For example, for each sentence, it can tag (i) which snippet ID was referenced, (ii) whether a graph path was used, and (iii) whether user-specific information (optional) was used. This can be directly used by the subsequent verification module (130) to determine whether there is a hallucination.
[0114] 8. Two-stage verification and modification / approval of the verification module (130).
[0115] The verification module (130) verifies the draft answer by referring to the integrated information, and in one embodiment may include the following two steps.
[0116] 8.1 Policy / Regulatory Compliance Verification
[0117] The verification module (130) determines a risk statement by referring to regulatory standards or service policies stored in the safety policy database (150a). The risk statement is, for example,
[0118] Expressions that compel, direct, or assert specific actions (“necessarily,” “certainly,” “unconditionally”)
[0119] Misleading expressions in medical / legal / investment (“confirmed illegality,” “no need to visit hospital,” “guaranteed returns”)
[0120] Expressions with a high likelihood of infringing on personal information / rights
[0121] It may include the back.
[0122] When a dangerous sentence is detected, the verification module (130) can delete the sentence, replace it with a template-based mitigation phrase, or call a generative artificial intelligence model to rewrite it to conform to the policy intent.
[0123] 8.2 Verification of Factuality (Hallucination) (Determination of Absence of Evidence / Conflict)
[0124] The verification module (130) checks for the existence of grounds within the integrated information for each sentence of the answer draft. If there are no grounds, if it conflicts with the grounds snippet, or if it (optional) contradicts the user's unique information, the sentence is identified as a hallucinatory sentence.
[0125] When a hallucinatory sentence is detected, the verification module (130) can provide feedback to the supervisor module (110) to (i) rewrite it by generalizing it to a valid range, (ii) delete the sentence, or (iii) request additional searching to secure the basis.
[0126] 9. Repetitive Control and Alternative Response Switching
[0127] When the verification module (130) outputs a non-conformity judgment, the supervisor module (110) can be controlled to perform at least one of the following as an additional action.
[0128] Additional search (change weights, limit / expand sources, enhance graph exploration)
[0129] Prompt reconfiguration (narrowing range, removing assertions, generating additional questions)
[0130] Regenerate answer
[0131] Safe template-based alternative response switching
[0132] If the non-conformity judgment is repeated more than a preset number of times, the supervisor module (110) can stop the response generation loop and control the provision of a template (general guidance / expert consultation recommendation / limitation notice) stored in the safety policy DB (150a) as the final expert information.
[0133] 10. User Interface and Transparent Orchestration Visualization
[0134] A user terminal (10) according to one embodiment of the present invention provides, as an optional option, a ‘Thinking Process’ verification mode that visually provides not only the final professional information but also the logical procedure leading up to the derivation of the final information.
[0135] (1) Real-time streaming of process status information
[0136] Whenever the internal operation state changes, the supervisor module (110) generates a 'process state event' corresponding to that stage and transmits it to the user terminal (10). The communication method may be WebSocket, SSE (Server-Sent Events), or Long Polling.
[0137] The above process status event may include the following step-by-step status messages.
[0138] · Analysis Phase: “Analyzing user query intent and risk...”
[0139] · Context Loading: “Loading user profile and underlying condition information...” (When unique user information is reflected)
[0140] · Knowledge Search: “Searching the latest case law and statute databases (mixed vectors / keywords)...”
[0141] · Reasoning and Generation: “Drafting answer based on searched evidence...”
[0142] · Policy Verification: “Verifying for violations of the Medical Service Act and safety policies...”
[0143] · Loop: “Performing an additional detailed search due to insufficient evidence...”
[0144] (2) Visual Interface (UI) Configuration
[0145] When the user activates the 'view process' function, the user terminal (10) displays the status message sequentially or in a timeline form in one area of the screen (e.g., a loading area before answer generation or a separate accordion panel).
[0146] In addition to simple text logs, a 'check () icon' or progress bar is updated whenever each step is completed, allowing users to intuitively recognize which stage (search vs. verification) the system is currently at.
[0147] (3) Exposure of intermediate products (Artifacts)
[0148] In addition to simple progress updates, the supervisor module (110) can provide intermediate outputs to the user within the scope allowed by the security policy.
[0149] For example, in the “searching for knowledge data” stage, the reliability of the source of information can be predicted even before the answer is generated by listing and displaying the top 3 “reference titles” or “legal provision numbers” that are actually searched and adopted.
[0150] Additionally, if the verification module (130) determines a specific sentence as a “safety policy violation” and modifies it, it may display a notification message “An unsafe expression was detected and modified into a softened expression” along with the final response as metadata (depending on user settings).
[0151] This transparent orchestration disclosure process demonstrates to users that the system does not merely combine words probabilistically, but undergoes a rigorous 'search-inference-verification' procedure, thereby significantly enhancing the credibility of professional information services.
[0153] 11. Examples
[0154] [Example 1: Healthcare (Symptom-centered natural language query)]
[0155] This embodiment details specific actions in which, when a user inputs a symptom-described (unstructured natural language-centered) query such as “I have been waking up frequently at night and thirsty lately, what could be the cause?”, “I have a persistent burning sensation in my stomach, what is the reason for that?”, “I have been getting easily tired and dizzy recently, is this a big problem?”, the system (100) evaluates ambiguity and risk / regulatory sensitivity highly, dynamically adjusts the weights of the hybrid search algorithm (specifically vector-based weights β), limits the final response to the range of ‘general health information’, and combines a visit recommendation template if necessary.
[0157] 1. Reception and preprocessing of input queries (S210~S219)
[0158] When a natural language query is received from a user terminal (10) (S210), the supervisor module (110) can perform the following preprocessing steps.
[0159] 1. Normalization
[0160] Correction of spacing / typo (optional), punctuation cleanup, reduction of redundant expressions
[0161] Arrange synonyms such as “wakes up often at night / wakes up frequently / wakes up in the middle of the night” as standard expressions
[0162] 2. Extraction of time, frequency, and intensity parameters (temporal parsing)
[0163] Convert “these days / recently / one month / since yesterday”, “often / sometimes / every day”, “severely / slightly”, etc. into quantifiable hints
[0164] Example: “Frequently at night these days” → Assign the same estimation parameters as Period = Last 2 weeks to 1 month (estimated), Frequency = 3 or more times per week (estimated) (however, if a user-specified value exists, it takes precedence)
[0165] 3. Red-flag candidate keyword detection
[0166] Generates a flag to increase the strength of safety verification in subsequent stages by primarily detecting high-risk patterns such as chest pain, shortness of breath, decreased consciousness, hemiplegia, and suicide / self-harm.
[0167] This preprocessing result is packaged with the source query as “preprocessing metadata” and used in subsequent analysis (S220).
[0169] 2. Query Analysis of Supervisor Module: Calculation of Ambiguity / Regulation / Risk Scores (S220~S229)
[0170] 2.1 Domain and Intent Classification
[0171] The supervisor module (110) can classify the query into a healthcare domain and then subdivide the intent. For example,
[0172] (a) Request for cause estimation (listing possibilities)
[0173] (b) Request for a definitive determination of a specific disease (“Is it diabetes?”)
[0174] (c) Request for treatment / medication / dosage instructions (“What should I take? How many mg?”)
[0175] (d) Request for urgency assessment (“Is this dangerous?”)
[0176] Symptom-centered queries often involve a mixture of (a) and (d), and users frequently unconsciously insert (b) and (c) (e.g., “Do I have diabetes? Do I need to take medicine?”), so the system can determine this as a multi-intent.
[0177] 2.2 Calculation of Ambiguity Score
[0178] The ambiguity score increases as the query leans toward “general and broad expressions” and lacks objective clues. For example, the following features can be used.
[0179] Lack of specific indicators: Absence of numerical values (body temperature, blood sugar, blood pressure), test name, duration / frequency, and information on associated symptoms
[0180] Increase in the proportion of symptom descriptions: Proportion of polysemous expressions such as “tired / dizzy / thirsty”
[0181] Question format: Open-ended questions such as “Why is this happening?” or “Is this a big deal?”
[0182] Lack of contextual context: Cases where deterministic context is absent, such as age group, pregnancy status, underlying conditions / medications.
[0183] The supervisor module (110) can calculate an ambiguity score A (0 to 1) based on the above features, and if A exceeds a threshold (A ≥ 0.6, etc.), it sets a “needs additional questioning” flag or strengthens the search / verification strategy.
[0184] 2.3 Calculation of Regulatory / Safety Risk Score
[0185] In the healthcare domain, risk scores include not only the "potential for dangerous diseases" but also the risk that a service might be mistaken for a medical act (diagnosis / prescription). For example:
[0186] Users demand confirmations: “You have diabetes, right?”, “You really need to take your medicine, right?”
[0187] Drug / Dosage Request: “How many mg do I take?”
[0188] Whether emergency / critical danger signal (red-flag) is included
[0189] If user-specific information (optional) exists, whether to include high-risk group contexts such as pregnancy, pediatrics, advanced age, or anticoagulant use.
[0190] The supervisor module (110) synthesizes this to calculate a risk score R (0 to 1), and controls that as R increases, (i) policy compliance verification is strengthened, (ii) the expression strength of the response is lowered (hedging), and (iii) template-based recommendations are actively combined.
[0192] 3. Hybrid Search Algorithm: β (Vector Weight) Upward and Multi-stage Search (S230~S249)
[0193] The key point of symptom-centered query is that “the expression is ambiguous and the keywords are not precise.” Therefore, the supervisor module (110) can set the vector-based semantic search weight (β) relatively high when searching the external database (130a).
[0194] 3.1 Determining Hybrid Search Weights
[0195] In this embodiment, the search scores can be fused as, for example, as follows.
[0196] score_keyword: BM25 / TF-IDF based keyword score
[0197] score_vector: Embedding-based similarity (cosine similarity / dot product, etc.)
[0198] score_graph(optional): Path / connectivity score of the knowledge graph (symptom-disease-test-treatment relationship)
[0199] And the integrated score can be calculated as follows.
[0200] base_score = α·score_keyword + β·score_vector + γ·score_graph
[0201] Here, α+β+γ = 1 (normalization)
[0202] In symptom-centered queries, β can generally be set to a large value (e.g., 0.6–0.8) and α to a small value (e.g., 0.2–0.4). However, if specialized entities such as “HbA1c”, “metformin”, or “fasting blood glucose 130” are mixed into the query, adaptive mixing to partially restore α or γ is also possible.
[0203] 3.2 Multiple Query Generation (Query Rewriting & Expansion)
[0204] Since ambiguous symptom expressions may result in fluctuating recall / precision when searched as is, the supervisor module (110) or the generation module (120) may use the following multiple query generation strategy.
[0205] Maintain original inquiry Q0
[0206] Synonym / Expression Conversion Q1: “I am thirsty” → “Dry mouth,” “Polyphagia / Polyuria,” “Increased fluid intake,” etc. (Note: Use only as search extensions, not as definitive statements)
[0207] Cause Categorization Q2: Questions regarding top categories such as endocrine (glucose metabolism), sleep disorders, dehydration, drug side effects, etc.
[0208] Red Flag Quarter Q3 (Optional): Search with keywords for accompanying symptoms such as “rapid weight loss / polyuria / fatigue / blurred vision”
[0209] At this time, Q1~Q3 are for internal search purposes to avoid making definitive conclusions in the answer, and can be managed separately so that they are not directly reflected in the final answer sentence until passing through the verification module (130).
[0210] 3.3 Candidate Document Collection and Re-ranking
[0211] Search can collect (i) vector search candidates, (ii) keyword search candidates, and (iii) graph path-based candidates in parallel, and then perform the following.
[0212] First integration using RRF (Reciprocal Rank Fusion) or weighted summation method
[0213] Deduplication and diversification (e.g., MMR)
[0214] Re-evaluate “Query-Document Match” with small ranking models or cross-encoder-based reranking
[0215] Reflects priority of source reliability (w_source), recency (w_recency), and document type (guidelines / textbooks / FAQs / case studies).
[0216] Consequently, “final K” snippet chunks are selected, and each chunk is assigned metadata such as document source, time of creation, document type, and reliability tag.
[0217] 3.4 Integrated Information Packaging (Context Configuration)
[0218] The selected basis chunk is packaged into “integrated information” and delivered to the generation module (120). This integrated information includes, for example:
[0219] Query Summary (Normalized Version)
[0220] Uncertainty / Risk Score (A, R)
[0221] List of Source Chunks (Source, Paragraph, Summary, Keyword)
[0222] Hints on prohibited expressions (policy summary, such as prohibition of definitive diagnosis / prescription)
[0223] Recommended conditions for visit (mandatory inclusion upon detection of red flags, etc.)
[0224] In this way, if policy compliance hints are included in the context from the search stage, the generation module (120) can be encouraged to generate fewer dangerous sentences in the first place (pre-safety shaping).
[0226] 4. Generate response draft for generation module: Structure “General Information Restriction” (S240)
[0227] The generation module (120) can force the following output structure under the control of the supervisor module (110).
[0228] (i) Present the “categories” of possible causes (list possibilities, avoid conclusions)
[0229] (ii) Additional verification questions (duration / accompanied symptoms / medication / fluid intake, etc.)
[0230] (iii) Warning signals (red flags) requiring immediate medical attention / emergency recommendation
[0231] (iv) “Safe general measures” of self-care (hydration, sleep hygiene, etc.)
[0232] (v) Disclaimer of Exclusions / Limitations (Not a Diagnosis / Prescription)
[0233] In particular, when the risk score R is high, the generation module (120) can configure the prompt to avoid assertive and directive expressions such as “doubtful / certain / take” and to use probabilistic and conditional expressions (conditional & hedged statements).
[0235] 5. Enhanced policy compliance verification + factual verification of the verification module (S250~S269)
[0236] In symptom-centered querying, it is important to prevent “plausible words (hallucinations)” and “misinterpretation of medical practice” simultaneously. Therefore, the verification module (130) can reinforce the following two steps.
[0237] 5.1 Policy Compliance Verification (Regulatory / Service Policy)
[0238] The verification module (130) refers to the safety policy database (150a) and breaks down the draft answer into sentences / clauses, and then:
[0239] Confirmed diagnosis (“It is diabetes”)
[0240] Prescription / Dosage Instructions (“Take ○ mg”)
[0241] Medical Substitute (“No Need to Go to the Hospital”)
[0242] Advertising / Exaggeration (“Cure”, “100%”)
[0243] Detects back patterns.
[0244] Detected sentences are processed using one of the following methods: (i) deletion, (ii) weakening (“there is a possibility”), or (iii) replacement of the consultation recommendation template. In this case, the replacement phrase can be rewritten not only by pasting the template as is, but also by calling a generative model to ensure policy compliance while maintaining contextual naturalness.
[0245] 5.2 Verification of Factual Verification (Evidence): Evaluation of Evidence-Sentence Consistency
[0246] The verification module (130) checks which chunk of integrated information each key sentence is supported by. The method is, for example:
[0247] If sentence-basis mapping metadata exists, use it first.
[0248] If not present, evaluate “whether the evidence supports the sentence” using NLI (Natural Language Inference) or LLM-based entailment judgments.
[0249] If support is insufficient, generalize or delete the sentence, and request additional search if necessary (S260 re-loop)
[0250] In this process, considering the characteristics of the symptom query, which is prone to producing “plausible estimations with weak evidence,” the verification module (130) can apply the evidence coverage threshold more strictly.
[0252] 6. Conditions and Operations for Combining Visit Recommendation Templates
[0253] In this embodiment, the visit recommendation template is not an “option” but can be hard constraintd under the following conditions.
[0254] Red flag candidate detection result is positive
[0255] Users directly request “medicine / prescription / diagnosis”
[0256] When A (ambiguity) and R (risk) are both high (A≥0.7, R≥0.7, etc.)
[0257] When sufficient supporting documents are not secured (deterioration of search result quality)
[0258] When combining templates, the system distinguishes between “emergency” and “general care recommendations” to avoid increasing user anxiety. For example:
[0259] Emergency Recommendation: Immediate response required for chest pain, difficulty breathing, loss of consciousness, etc.
[0260] General medical recommendation: Consultation recommended if symptoms persist, worsen, or are accompanied by symptoms.
[0261] In other words, instead of simply saying "Go to the hospital," it controls the system to provide step-by-step recommendations through conditional triage logic.
[0263] 7. Action Example (Connected to a Simple Scenario)
[0264] User: “Lately, I've been waking up frequently at night and feeling thirsty. What could be the cause?”
[0265] Supervisor (110): Ambiguity A↑ (symptom description-centered), Risk R↑ (disease conclusion-centered), β upward (vector search-centered), policy verification strengthened
[0266] Search: Widely collected guidelines, FAQs, and case studies related to “night awakening + dry mouth” based on vectors, reinforced documents in the “polyuria / fluid intake / drug side effects / sleep environment” categories using select keywords, removed duplicates, and re-ranked.
[0267] Generation (120): Possible Causes “Category” + Check Questions + Warning Signals + General Care Tips + Disclaimer
[0268] Verification (130): If there is an assertion such as “it is likely diabetes,” weaken / delete; if there is an unfounded assumption, generalize; if there is a red flag item, combine templates.
[0269] · Final: Safely provided within the scope of “general information”
[0271] 8. Summary (Technical effect points of this embodiment)
[0272] The core of this embodiment is to simultaneously address the uncertainty of symptom-centered queries and medical compliance risk,
[0273] 1. The supervisor module (110) quantifies the ambiguity and risk scores to dynamically control the workflow, and
[0274] 2. In the hybrid search algorithm, increase the weight of vector search (β) while improving evidence quality through multiple query expansion / reranking, and
[0275] 3. After the verification module (130) strengthens policy compliance verification and removes hallucinations through evidence-sentence consistency,
[0276] 4. Safety is ensured by forcibly combining visit recommendation templates according to conditions.
[0278] [Example 2: Legal / Patent / Compliance]
[0279] In queries containing provisions / numbers / requirements / exceptions, r_entity is high, which strengthens keyword-based search (α) and graph-based search (γ). A “requirement-exception-procedure” path is secured through graph navigation, and the generation module (120) packages the basis provision snippet together. The verification module (130) applies a “legal advice misrecognition prevention” template to mitigate assertive expressions.
[0280] [Example 3: Finance / Tax]
[0281] If the recency attribute is high, it is reordered to increase the influence of w_fresh, and expressions guaranteeing returns or making assertions are blocked as risky sentences. If the evidence is weak, the scope is narrowed or it is switched to a consultation recommendation template.
[0282] As such, the system according to the present invention can substantially reduce hallucinations and regulatory / policy risks in the process of providing expert information based on generative AI, and improve the reliability and safety of the final expert information by organically combining dynamic workflow control (execution status / sequence / number of repetitions) based on query analysis, selective reflection of user-specific information, hybrid search combining keyword / vector / graph-based search and dynamic weighting / re-ranking algorithms, integrated information packaging based on evidence snippets, and risk sentence control and hallucination sentence removal based on safety policy DB. Explanation of the symbols
[0283] 10: User terminal 20: Network 100: Professional Information Provision System 110: Supervisor Module 120: Create module 121: User unique information storage unit 130: Verification Module 130a: External database 130a-1: Document Repository 130a-2: Vector Index 130a-3: Knowledge Graph 130a-4: Recency / Reliability Metadata 150a: Safety Policy Database 170: Message Queue (or Event Bus) 200: Server
Claims
Claim 1 A system for providing specialized information using a generative artificial intelligence model, comprising: a supervisor module that receives a natural language query from a user, analyzes the query to produce an analysis result, and dynamically controls the execution status, execution order, and number of repetitions of at least some additional operations for providing specialized information based on the analysis result; a generation module that generates a draft answer using the query and integrated information generated from the user's unique information or external database search results according to the control of the supervisor module; and a verification module that determines the final specialized information by verifying the factual basis of the draft answer by referring to the generated integrated information according to the control of the supervisor module, and modifying or approving the draft answer based on the verification result, wherein the supervisor module is configured to improve the reliability of the final specialized information within a dynamically set number of repetitions by controlling at least some of the additional operations to be re-executed based on the verification result. Claim 2 A system according to claim 1, wherein the verification module determines whether the draft response contains a risk sentence corresponding to the regulatory standard by referring to the regulatory standard stored in the safety policy database, and if the risk sentence is detected, deletes the risk sentence or replaces it with an alternative phrase to generate modified final text information. Claim 3 A system characterized in that, in paragraph 2, the verification module identifies, for each sentence constituting the draft answer or the modified professional information, a hallucinatory sentence that conflicts with the integrated information or cannot be found as a basis in the integrated information, and when the hallucinatory sentence is detected, generates final professional information from which the hallucinatory sentence has been removed by using the generative artificial intelligence model to remove the hallucinatory sentence or modify it into a more general explanatory sentence. Claim 4 A system characterized in that, in claim 1, the unique information of the user is stored associated with a plurality of information items for each user, and the information items include at least one of profile information, preferences, goals, constraints, past interaction history, and risk sensitivity; the supervisor module selectively reflects the part of the unique information of the user that is relevant to the query context into the integrated information according to the analysis result of the query; adds new unique information obtained during the user query and the operation process of the system by associating it with the information items, or, if it contradicts existing unique information, merges and updates based on the latest information based on time information and reliability; and converts the updated unique information into a natural language summary text or structured form suitable for the query context and provides it to the generation module. Claim 5 A system characterized in that, in claim 1, the integrated information includes search results for an external database, and the supervisor module controls the generation module to search the external database using two or more search methods among keyword-based search, graph-based search, and vector-based semantic search based on the analysis results for the query, and controls the generation module to dynamically determine weights assigned to the results of each of the two or more search methods according to the attributes of the query included in the analysis results, and to rearrange the search results by reflecting the weights and the source reliability of the documents, and then generate the integrated information from the rearranged search results. Claim 6 A system according to claim 1, wherein the supervisor module controls the execution of at least one of the additional actions when it receives a determination from the verification module that the draft answer is unsuitable due to at least one of the possibility of violating regulatory standards or service policies, the possibility of containing hallucinogenic sentences, or the risk of infringing on user safety or rights, wherein the additional actions include at least one of performing an additional search to obtain additional supporting information from an external database, performing an edit to mitigate the expression intensity of the draft answer or remove specific assertive expressions, reconstructing prompt data using the query or the integrated information, and regenerating the draft answer using the generative artificial intelligence model, wherein if the determination of unsuitability is repeated more than a preset or dynamically set number of times, the supervisor module controls the discontinuation of the provision of the draft answer and the provision of an alternative response based on a safety guidance template or an expert consultation recommendation template stored in a safety policy database as the final expert information. Claim 7 In claim 1, the supervisor module determines whether new unique information regarding the user is included in the answer draft or the final text information based on metadata received from the generation module and the verification module, and controls the updating of the storage structure of the user-specific unique information asynchronously only when the new unique information exists and a determination is received from the verification module that long-term storage of the new unique information is permitted, and controls the application of at least one storage restriction policy among anonymization, partial storage, or storage prohibition for unique information for which long-term storage is not permitted or the retention period is limited, and the supervisor module, the generation module, and the verification module may each be implemented in a structure that repeatedly performs inference and action by configuring function-specific sub-prompts to input into the generative artificial intelligence model, interpreting the model response to update the internal state or instructing the execution of at least some of the additional operations, and the system further includes a user interface implemented on a user terminal, wherein the supervisor module provides status messages indicating the progress status of each step in real time while the query analysis, integrated information generation, answer draft generation, and verification process by the verification module are performed. A system characterized by generating and transmitting to the user terminal, wherein the user interface has a process verification mode that is activated based on the user's selective input, and when the process verification mode is activated, displaying the status message on the screen in the form of a sequential or visual object before the final professional information is displayed to provide an orchestration process of the supervisor module, and displaying at least one intermediate output information along with the status message, such as a list of references identified in the step where the supervisor module performs the external database search, or a notification regarding the reason for modification when the verification module modifies the answer draft. Claim 8 A method for providing expert information executed by a computer, comprising: receiving a natural language query from a user; analyzing the query to produce an analysis result; dynamically controlling whether to perform, the order of execution, and the number of repetitions of at least some additional actions for providing expert information based on the analysis result; generating a draft answer using a generative artificial intelligence model using the query and integrated information generated from the user's unique information or external database search results; and determining the final expert information by verifying the factual basis of the draft answer by referring to the integrated information and modifying or approving the draft answer based on the verification result. Claim 9 In claim 8, the step of generating the integrated information from the external database search results comprises: a step of performing a search by selecting or combining two or more search methods among keyword-based search, graph-based search, and vector-based semantic search based on the attributes of the query included in the analysis results; a step of dynamically determining a weight assigned to the search result of each of the two or more search methods according to the attributes of the query; a step of rearranging the search results by reflecting the weight and the source reliability of the document; and a step of generating the integrated information from the rearranged search results, and further comprising, prior to the step of determining the final professional information, a step of activating an orchestration visualization function according to a user's request; a step of generating a process log whenever each step is started or completed by a supervisor module; and a step of visualizing and providing the generated process log as a step-by-step progress on the screen of a user terminal. Claim 10 A computer-readable recording medium having a program for executing the method according to paragraph 8.
Citation Information
Patent Citations
Multi-Stage Verification-Based Error-Blocking Artificial Intelligence Response Generation System and Operating Method Thereof
KR1020250120232A
Automated plan generation and execution system for large language models
US20240289545A1
Response generation using a retrieval augmented ai model
US20240346256A1
