Multi-user Artificial Intelligence Collaboration Orchestration System and Method in Controlled Enterprise Environment

Abstract

The present invention relates to a multi-user artificial intelligence collaboration orchestration system and method in a controlled enterprise environment. The system of the present invention normalizes natural language or file-based collaboration requests from multiple users into Task Card objects and dynamically routes Task Cards to Front LLM, Worker LLM, asynchronous queue, or idle time execution paths by comparing the number of active real-time AI sessions, token generation speed, GPU / VRAM or integrated memory usage, memory bandwidth load, context usage, and task cost. Furthermore, between logically isolated collaboration spaces, the original data is not directly transmitted; instead, an association summary token is generated, and only masked or generalized knowledge is mediated after approval by an authorization holder. Moreover, user-provided files or external data are not immediately embedded in the official RAG vector DB but are isolated in a quarantine repository, and are incorporated into official RAG knowledge only after passing through hash / file structure / hidden text / prompt injection verification, a primary summary report, authorization holder approval, and audit logs. Accordingly, real-time AI conversation quality, knowledge security, RAG integrity, and operational auditability can be improved in closed networks, on-premises, private cloud, VPC, hybrid, or tenant-isolated execution environments.

Description

Technology Field

[0001] The present invention relates to an artificial intelligence (AI) collaboration system, an enterprise knowledge management system, a Retrieval Augmented Generation (RAG) system, a multi-user task orchestration system, and a secure AI execution environment. More specifically, the invention relates to a system and method for a controlled enterprise AI execution environment comprising limited computational resources and a logically isolated knowledge repository per customer organization or tenant, wherein the system normalizes multi-user natural language collaboration requests into Task Card objects, dynamically routes AI tasks by recognizing the state of hardware resources to simultaneously protect the quality of real-time conversation sessions and long-term AI task execution, performs authorization-based knowledge brokerage between isolated collaboration spaces in a manner that does not expose original data, and performs multi-stage quarantine, approval, and audit procedures before incorporating user-provided files or external data into the official RAG knowledge repository.

[0002] In this specification, the "controlled enterprise AI execution environment" is not limited to a closed network environment in which communication with the external internet is completely blocked, but includes an execution environment comprising knowledge repositories isolated by restricted computing resources and access policies per customer organization or tenant, such as on-premises devices, internal networks with limited internet connectivity, customer-specific dedicated servers, private clouds, VPC-based dedicated instances, tenant-isolated SaaS instances, and hybrid on-premises-cloud environments. Background Technology

[0003] When utilizing generative AI for business operations within companies or organizations, users input requests of varying natures into the same AI system, such as real-time Q&A, review of contract or policy documents, inspection of large folders, OCR, vision analysis, generation of weekly reports, and internal knowledge retrieval. However, these requests differ significantly in computational cost, and if they share the same CPU, GPU, VRAM, Unified Memory (UMA), memory bandwidth, context window, or inference runtime, lengthy or high-cost requests can degrade the token generation rate (TPS) of real-time conversation sessions or cause out-of-memory (OOM).

[0004] Conventional simple chatbots or general business automation systems often stop at immediately executing user requests or simply registering them in a queue. This approach fails to comprehensively account for operational costs, model paths, context usage, memory pressure, the number of active sessions, and audit policies when operating real-time conversations and background tasks simultaneously in a multi-user collaboration space.

[0005] Furthermore, internal corporate knowledge is often managed separately at the level of departments, projects, customers, security levels, or tenants. Existing access control methods operate by blocking searches by unauthorized users or injecting only permitted data into queries. However, in actual collaboration, there is a need for AI to recognize the correlation between a request from one department and the knowledge of another, but to mediate the data as masked or generalized knowledge only after approval from the authorized holder, rather than immediately disclosing the original data.

[0006] Furthermore, PDFs, images, documents, tables, code, or other external data uploaded from internal collaboration spaces must be quarantined before becoming a basis for searching the official RAG vector DB. If uploaded files are embedded immediately, malicious scripts, hidden text, prompt injections, misinformation, or unauthorized internal documents may be incorporated into the official search base, affecting AI response quality and security.

[0007] Therefore, in a controlled enterprise AI execution environment, technical means are required to process restricted computational resources, multi-user requests, logical knowledge isolation, authorization, RAG inclusion status, audit logs, and policy enforcement on a single control plane, rather than simple UI or bulletin board configurations. The problem to be solved

[0008] The objectives of the present invention are as follows.

[0009] First, the goal is to provide a resource-aware AI task routing system that can handle heavy AI tasks such as large-scale RAG, OCR, Vision, folder scanning, and precision report generation, while preventing slowdowns in token generation speed and memory shortages in real-time 1:1 AI conversation sessions in a multi-user AI execution environment that shares a pool of limited computing resources per customer organization or tenant.

[0010] Second, it aims to normalize a user's natural language task request into an executable Task Card object and to improve the verifiability and reproducibility of the task by utilizing the input, pipeline, output, allow tool, block tool, approval policy, and execution history of the said Task Card.

[0011] Third, it is to provide a Context Broker that mediates only masked or generalized knowledge based on the explicit approval of the authorization holder, without directly transmitting original data between logically isolated departments, projects, customers, or tenant collaboration spaces, and only generates association summary tokens.

[0012] Fourth, it provides a Zero-Trust knowledge incorporation pipeline that does not immediately incorporate user-provided files or external data into the official RAG knowledge repository, but instead embeds them into the official RAG vector DB only after undergoing isolated storage, integrity and file structure verification, prompt injection and hidden text neutralization, generation of a primary summary report, authorization approval, and audit log recording.

[0013] Fifth, the purpose is to ensure that the above technical configuration is not limited to closed network on-premises environments, but can be applied with the same control principles in in-house networks with limited internet connectivity, private clouds, VPC-based dedicated instances, tenant-isolated SaaS instances, and hybrid environments. means of solving the problem

[0014] To achieve the above objective, a system according to one embodiment of the present invention is a system for processing AI collaboration requests from multiple users in a controlled enterprise AI execution environment, and may include a request receiving unit, a task card generating unit, a resource monitoring unit, a task cost evaluation unit, a dynamic routing unit, an execution queue or task run management unit, a front LLM path, a worker LLM path, a callback posting unit, a context broker, a zero-trust file quarantine and RAG inclusion unit, and a policy enforcement and audit log unit.

[0015] The request receiver receives natural language collaboration requests or file-based requests from a user interface, 1:1 AI session, collaboration space, channel, thread, file attachment, or API. The Task Card creator normalizes the received requests into Task Card objects that include a title, directive, trigger, input resources, pipeline, output format, allow tools, block tools, approval policies, execution history, outputs, and audit logs.

[0016] The resource monitoring unit measures at least one of the following: the number of active real-time conversation sessions, token generation rate, GPU / VRAM usage, integrated memory usage, memory bandwidth load, context window usage, queue backlog, number of concurrent executions per model, or number of running task runs. The task cost evaluation unit evaluates the RAG search volume of task cards, the need for OCR or Vision, the number and size of files, folder scan range, report length, the need for precise judgment, long text context usage, the need for approval, or the use of external tools.

[0017] The dynamic routing unit compares resource status, task cost evaluation results, and policy profiles to assign a task card to one of the following paths: immediate execution, front LLM path, worker LLM path, asynchronous queue path, or idle time / scheduled execution path. For example, if the token generation rate of a real-time session is predicted to drop below a threshold, or if VRAM or integrated memory usage exceeds a threshold, the dynamic routing unit may register the corresponding task card as an asynchronous queue or worker LLM path instead of executing it immediately.

[0018] The callback posting unit returns the output to the 1:1 AI session, collaboration space, channel, or thread where the original request originated after the asynchronous processing is complete. The execution queue or Task Run management unit can manage deterministic run ID, idempotency key, queue status, execution status, cancellation flag, timeout, retry / backoff, output, and audit log.

[0019] A context broker according to another embodiment of the present invention is a system for mediating knowledge between a plurality of logically isolated collaboration spaces, and may include an association detection unit that detects semantic association between a request from a first collaboration space and a knowledge object from a second collaboration space; a draft generation unit that generates an association summary token including association type, reliability, non-identifiable reference value, and reason for request, without transmitting original data from the second collaboration space to the first collaboration space; an approval broker that privately transmits the summary token to an authority holder in the second collaboration space to request explicit approval; a knowledge broker that masks sensitive information and generalizes business procedures or knowledge essences to transmit them to the first collaboration space only when an approval signal is received; and an audit log unit.

[0020] A Zero-Trust RAG incorporation method according to another embodiment of the present invention may include the steps of: storing data uploaded to a collaborative space by a user or an external data provider in an isolated quarantine volume or quarantine storage without immediately embedding it in an official RAG vector DB; verifying a file hash, digital signature, file type, path structure, file size, whether a script is included, hidden text, and prompt injection pattern; generating a primary summary report and risk score for the verified data, but not using it as a target for official RAG search before approval; displaying an approval request to an authorized person or administrator; performing normalization, chunking, and embedding only on the approved data to incorporate it into an official RAG vector DB; and storing the entire state transition as an audit log. Effects of the invention

[0021] According to the present invention, the following effects are achieved.

[0022] First, even if real-time conversations and heavy AI tasks share the same or related computational resources in a multi-user AI collaboration environment, the risk of degraded real-time response quality and memory shortage can be reduced by dynamically adjusting execution paths based on the number of active sessions, token generation speed, memory usage, context usage, and task costs.

[0023] Second, by normalizing natural language requests into Task Card objects and preserving allow tools, block tools, approval policies, execution history, and audit logs for each execution unit, verifiability, reproducibility, revocability, and operational transparency required in an enterprise environment can be enhanced.

[0024] Third, even if AI detects knowledge associations between departments or tenants, it does not immediately disclose the original data; instead, by mediating only masked generalized knowledge through a non-recoverable association summary token and an authorized person approval process, it is possible to simultaneously ensure collaboration and security between information silos.

[0025] Fourth, by preventing uploaded files from being immediately incorporated into the official RAG vector DB and by subjecting them to quarantine, integrity verification, prompt injection neutralization, approval, and auditing procedures, it is possible to reduce contamination of the RAG knowledge store, injection of incorrect evidence, and exposure of unauthorized information.

[0026] Fifth, since the independent billing structure is not limited to closed networks but can be applied across controlled enterprise AI execution environments, the same technological benefits can be achieved across on-premises devices, closed networks, private clouds, VPCs, hybrid, or tenant-isolated deployments. Brief explanation of the drawing

[0027] FIG. 1 is a configuration diagram illustrating the overall AI collaboration orchestration architecture of the present invention according to one embodiment of the present invention. Figure 2 is a flowchart illustrating a task card-based resource-aware AI task routing procedure. Figure 3 (FIG. 3) is a diagram illustrating the comparison structure of resource status signals and work cost signals used in dynamic routing. FIG. 4 is a sequence diagram illustrating the procedure in which a Context Broker mediates knowledge between logically isolated collaboration spaces using association summary tokens and authorized person approvals. Figure 5 (FIG. 5) is a flowchart illustrating the Zero-Trust quarantine and official RAG vector DB inclusion pipeline for user-provided files or external data. FIG. 6 illustrates a control plane that is common to examples of deployment variations to which the present invention can be applied, namely on-premises devices, closed networks, private clouds, hybrid and tenant-isolated SaaS instances. Figure 7 (FIG. 7) is a diagram illustrating the main fields of a Task Card object and the Task Run lifecycle. Specific details for implementing the invention

[0028] Embodiments of the present invention are described below with reference to the attached drawings. The following embodiments are merely examples for explaining the present invention, and the scope of the present invention is not limited to specific product names, UI types, model names, cloud status, or closed network status.

[0029] 1. Overall System Configuration

[0030] Referring to FIG. 1, a controlled enterprise AI execution environment (100) according to one embodiment of the present invention may include a collaboration interface (110) accessed by a user or employee (101), a 1:1 AI session (111), a collaboration space or channel (112), a file or data input unit (113), a request receiving unit (120), a task card creation unit (130), a resource monitoring unit (140), a task cost evaluation unit (150), a dynamic routing unit (160), a front LLM path (170), a worker LLM path (180), an asynchronous queue (163), a callback posting unit (190), an isolated knowledge domain management unit (200), a context broker (210), an official RAG vector DB (222), a quarantine volume (231), a Quarantine / RAG Gate (230), and a policy and audit log unit (240).

[0031] The collaboration interface (110) may be a web UI, a desktop application, a mobile application, an internal portal, a messenger-type channel, an API, or other user interface. The invention is not limited to specific UI formats such as Slack, Discord, a bulletin board, a portal, a chatbot, or an independent work screen.

[0032] The request receiving unit (120) receives natural language queries, work requests, file attachments, folder monitoring requests, periodic reporting requests, internal knowledge search requests, or API requests from external systems. The received requests are converted into executable task card objects (131) by the task card generating unit (130).

[0033] 2. Task Card Object

[0034] Referring to FIG. 7, a Task Card object (131) is an object that converts a user natural language request into a structured task unit that the system can validate and execute. Depending on the embodiment, the Task Card object may include at least one of title, instruction, trigger, inputs, pipeline, outputs, allowed_tools, blocked_tools, approval_policy, run_history, artifacts, and audit_log.

[0035] Triggers may include manual, schedule, watched_folder, email_digest, or equivalent execution conditions. Inputs may include uploaded files, folder paths, RAG search targets, existing reports, email inboxes, or user input data. Pipelines may include extract, classify, normalize, summarize, table, report, risk_review, or equivalent processing steps. Outputs may include reports, dashboards, internal alerts, draft emails, CSV, Excel, DOCX, PDF, or other output formats.

[0036] The Task Validator (133) can verify whether the input path of the Task Card is within the allowed workspace, whether the allowed tools and blocked tools do not conflict with the policy, whether additional approval is required for high-risk actions such as sending externally or modifying / deleting the original file, and whether the task cost does not exceed the limits of the operational profile.

[0037] 3. Resource-Aware Dynamic Routing

[0038] Referring to FIGS. 2 and 3, the resource monitoring unit (140) can measure the number of active real-time 1:1 AI sessions (141), token generation rate or TPS (142), GPU / VRAM / integrated memory usage (143), memory bandwidth load (144), context window usage (145), asynchronous queue backlog (146), etc. The task cost evaluation unit (150) can calculate the RAG search cost (151) of the task card, OCR or Vision cost (152), file scan scale (153), report length or need for precision judgment (154).

[0039] The dynamic routing unit (160) may include a resource-cost comparison unit (155) that compares resource status and task cost. The comparison unit (155) considers the policy profile, operation class, time zone, approval policy, or user priority together and assigns the corresponding task card to one of the front path (161), worker path (162), asynchronous queue (163), or idle time execution scheduler (164).

[0040] The front LLM route (170) can be configured to handle lightweight requests where real-time responsiveness is important, such as general conversations, short document summaries, and UI usage guides. The worker LLM route (180) can be configured to handle requests requiring high cost or precision processing, such as OCR, Vision, contract or tax risk reviews, long reports, large volume RAGs, and Task Card compile / validates. The asynchronous queue (163) can hold requests that could cause real-time session quality degradation or out of memory upon immediate execution and pass them to a worker route or idle time execution.

[0041] For example, the dynamic routing unit (160) may register a task card in an asynchronous queue (163) without immediately executing it if it is predicted that the TPS of a real-time conversation session will fall below a predefined lower limit, or if the GPU / VRAM / UMA memory occupancy is above a predefined threshold. Additionally, OCR / Vision, bulk folder scanning, long-text report generation, or recurring risk reports may be assigned to a worker route or a deferred route by default according to the operation profile.

[0042] The Task Run management unit (250) can assign a deterministic run ID or idempotency key (242) to each execution. Accordingly, duplicate processing of the same execution can be prevented even if scheduler tick duplication, network retries, duplicate user requests, or queue re-registrations occur. The Task Run management unit (250) manages states such as queued, running, waiting_approval, succeeded, failed, cancelled, etc., and can process execution cancellation requests, timeouts, retry / backoffs, and output saving.

[0043] The callback posting unit (190) returns the output or summary result to the collaboration space (112), channel, thread, or 1:1 AI session (111) where the original request occurred when the asynchronous execution is completed. The Task Run status and output can be maintained even if the user moves the screen while working.

[0044] 4. Example of an Operational Profile

[0045] In one embodiment, in a single-user device with limited computational resources, the front LLM and worker LLM may share the same local inference runtime, and the dynamic routing unit (160) may defer heavy task cards to sequential execution or after-hours execution. In another embodiment, in a team operation device or a dedicated GPU server, the front LLM and worker LLM are separated so that the front LLM handles general conversations and light summaries, and the worker LLM handles precision reports, OCR / Vision, and long-duration tasks.

[0046] Operational profiles can be represented as off, standard_lite, premium_pending, premium_pro, or equivalent modes. For example, in standard_lite mode, due task enqueues are limited to a low frequency, and OCR / Vision, recurring risk reports, heavy RAG, and long-term scheduled tasks can be skipped or executed after user confirmation. In premium_pro mode, team-based task cards are processed using a validated scheduler and worker routes, while providing notifications for long-term execution, queue delays, and failures to administrators or team managers.

[0047] 5. Context Broker

[0048] Referring to FIG. 4, a context broker (210) mediates the association of knowledge between a first collaboration space (201) and a second collaboration space (202) that are logically isolated. The isolation criteria may be a department, project, customer, tenant, security class, work subject, or access policy.

[0049] When a request occurs in the first collaboration space (201), the association detection unit (211) can detect semantic associations with knowledge objects stored in the second collaboration space (202). At this time, the original document, original conversation, original file, or sensitive data of the second collaboration space (202) is not directly injected into the LLM context of the first collaboration space (201).

[0050] The Draft generation unit generates an association summary token (212). The association summary token (212) may include at least one of a non-identifying reference value that cannot restore the original data, an association type, a reliability level, a request purpose, a requester identifier or role, a related policy, an authorized person identifier, an approval expiration time, and a reason for requiring approval. The association summary token may be generated so as not to include sentences, filenames, customer names, amounts, proper nouns, or secret information of the original data.

[0051] The approval broker (213) privately transmits the association summary token (212) to the authority holder (203) of the second collaboration space (202) and requests explicit approval for viewing or brokering. If the authority holder does not approve or the approval time expires, knowledge brokering is not performed.

[0052] When an approval signal is received, the masking / generalization module (214) removes or masks sensitive information from the original data of the second collaboration space (202) and generates only business procedures, judgment criteria, checklists, generalized rules, or summary knowledge. The knowledge brokerage unit (215) transmits the generalized knowledge to the first collaboration space (201). In this process, the entire original data or unauthorized sensitive data is not transmitted.

[0053] The context broker (210) records approval, rejection, expiration, masking results, version of transmitted knowledge, reason for brokering, and requester / approver information in the audit log (241).

[0054] 6. Zero-Trust File Quarantine and Official RAG Inclusion

[0055] Referring to FIG. 5, when a user uploads external data or a user-provided file through the collaboration space (112) or file input section (113), the Quarantine / RAG Gate (230) does not immediately embed the data into the official RAG vector DB (222) but stores it in a quarantine volume or quarantine storage (231).

[0056] The verification unit (232) can check the file hash, digital signature, file type, path structure, file size, allowed extensions, internal structure of the compressed file, whether a script is included, whether a macro is included, hidden text within the image / document, or file creator information. The sanitization module (233) can remove, mask, or disable prompt injection patterns, system directive disguise phrases, hidden text intended to deceive the model, external URL call instructions, scripts, or executable code.

[0057] The Draft Report generation unit (234) can generate a primary summary report including a summary of the file, source, hash value, risk score, verification result, estimated classification, recommendation processing status, and whether to recommend inclusion in RAG. The approval gate (235) displays the Draft Report to an administrator, department head, document owner, or policy-authorized person and waits for an approval or rejection input.

[0058] Original source data and Draft Reports prior to approval are not used as targets for official RAG searches. In other words, while the system may place pre-approval data in a temporary quarantine state or a review index, it is not exposed as an official RAG retrieval result for general user queries.

[0059] Only when approval input is confirmed, the normalization / chunking / embedding module (236) normalizes, chunks, and embedding the original data or the data after sanitization according to a parser recipe and reflects it in a designated collection of the official RAG vector DB (222). The RAG collection manager (237) can manage the collection ID, document version, approver, approval time, hash before and after sanitization, parser recipe, embedding recipe, and revocation status.

[0060] If approval is rejected, verification fails, or approval expires, the Reject / Archive / Revoke module (238) transitions the file to a discard, archive, return, or review state. All state transitions are recorded in the audit and state transition log (241).

[0061] 7. Distribution Variants

[0062] Referring to FIG. 6, the control surface (300) of the present invention can be applied to an on-premises device (301), a closed network or air-gapped environment (302), a private cloud or VPC (303), a hybrid on-premises-cloud environment (304), and a tenant-isolated SaaS instance (305).

[0063] In a closed network embodiment, the system may be configured to use only the LLM runtime, RAG repository, and workspace within the customer organization without external LLM API calls. In an on-premises embodiment, the system may be deployed on a single device or on a customer internal server. In a private cloud or VPC embodiment, the system may use a customer-specific dedicated GPU / worker pool and a logically isolated RAG repository. In a hybrid embodiment, some management control planes may be connected to a central management server, but customer data or official RAG knowledge repositories may be stored in a customer-specific isolated area.

[0064] 8. Example of linkage with existing security layers

[0065] In one embodiment, the present invention may operate on a security layer such as measurement-based integrity verification, hardware-based authentication, virtualization or container-based logical isolation, quarantine volumes, security volumes, gateway and queue / channel-based data flow restriction, restriction of external network outbound communication, and blocking of bypass paths. However, the independent technical feature of the present invention lies not in whether it is a closed network, but in the orchestration control aspect, which normalizes multi-user AI collaboration requests into task cards, controls execution paths by comparing resource status and task costs, performs original non-exposure approval mediation between isolated knowledge spaces, and performs quarantine-approval-audit state transitions before formal RAG inclusion.

[0066] 9. Audit Logs and Policy Enforcement

[0067] The policy and audit logbook (240) can record task card creation, verification, activation, deactivation, execution, failure, retry, timeout, cancellation, input file access, output creation, download, context broker approval / rejection, RAG inclusion approval / rejection, license or feature gate verification, user permission verification, etc. in the audit log (241).

[0068] Policy enforcement may include allowed_tools, blocked_tools, approval_policy, workspace path restrictions, RAG collection access permissions, blocking external sending, blocking modification / deletion of source files, blocking unauthorized document retrieval, excluding files awaiting approval from official RAGs, and holding off on the execution of tasks requiring administrator approval. Explanation of the symbols

[0069] 100: Controlled Enterprise AI Execution Environment 101: Users 110: Collaboration Interface 111: 1:1 AI Session 112: Collaboration Space or Channel 113: File or Data Input 120: Request Receiver 130: Task Card Generator 131: Task Card Object 132: Task Card Fields 133: Task Validator 140: Resource Monitor 141: Active Session Count Monitor 142: Token Throughput Monitor 143: GPU / VRAM / UMA Memory Monitor 144: Memory Bandwidth Load Monitor 145: Context Window Monitor 146: Queue Backlog Monitor 150: Cost Evaluator 151: RAG Retrieval Cost 152: OCR / Vision Cost 153: File Scan Volume 154: Report Length / Precision Need 155: Resource-Cost Comparator 160: Dynamic Router 161: Front Route 162: Worker Route 163: Async Queue 164: Deferred / Idle-Time Scheduler 170: Front LLM Route 180: Worker LLM Route 190: Callback Posting 200: Isolated Knowledge Domain Manager 201: First Collaboration Space 202: Second Collaboration Space 203: Authorized Holder 210: Context Broker 211: Relation Detector 212: Relevance Summary Token (Draft Token) 213: Approval Broker 214: Masking / Generalization Module 215: Brokered Generalized Knowledge Module 220: Knowledge Store 221: Isolated RAG Store 222: Official RAG Vector DB 230: Quarantine / RAG Gate 231: Quarantine Volume or Inbox 232: Integrity / Type Validator 233: Prompt Injection Sanitizer 234: Draft Report Generator 235: Authority Approval Gate 236: Normalization / Chunking / Embedding Module 237: RAG Collection Manager 238: Reject / Archive / Revoke Module 240: Policy and Audit Book (Policy + Audit) 241: Audit Log 242: Idempotency Key / Deterministic Run ID 250: Task Run Manager 251: Run State 252: Artifact Store 300: AI Collaboration Orchestration Control Plane 301: On-Premise Appliance 302: Closed Network or Air-gapped Environment 303: Private Cloud or VPC 304: Hybrid Deployment Environment 305: Tenant-Isolated SaaS Instance

Claims

Claim 1 A system for processing multi-user AI collaboration requests in a controlled enterprise AI execution environment, comprising: a request receiving unit receiving a natural language or file-based collaboration request from at least one of a user interface, a 1:1 AI session, a collaboration space, a channel, a thread, a file attachment, or an API; a task card generating unit normalizing the collaboration request into a task card object comprising at least one of a directive, a trigger, an input resource, a processing pipeline, an output format, an allow tool, a block tool, and an approval policy; a resource monitoring unit measuring at least one of the number of active real-time AI sessions, a token generation rate, GPU / VRAM or integrated memory usage, memory bandwidth load, context window usage, and a queue backlog; a task cost evaluation unit calculating at least one of the RAG search volume of the task card object, whether OCR or Vision is required, file scan scale, long text context usage, report generation cost, and the need for precision judgment; and comparing the measurement result of the resource monitoring unit with the calculation result of the task cost evaluation unit to determine the task card object among an immediate execution path, a front LLM path, a worker LLM path, an asynchronous queue path, or an idle time / scheduled execution path. A dynamic routing unit that assigns a single item; a Task Run management unit that executes the Task Card object according to the assigned path and manages the execution status, cancellation status, and output; a callback posting unit that returns the output to the AI ​​session, collaboration space, channel, or thread where the original request occurred after the execution of the Task Card object is completed; and an audit log unit that records the history of creation, verification, execution, cancellation, failure, output creation, and return of the Task Card object.A multi-user artificial intelligence collaboration orchestration system comprising, wherein the controlled enterprise AI execution environment comprises restricted computing resources allocated to a customer organization or tenant and a knowledge repository isolated by access policies. Claim 2 A multi-user artificial intelligence collaborative orchestration system according to claim 1, wherein the limited computational resources include a CPU, GPU, NPU, VRAM, unified memory (UMA), memory bandwidth, context window, LLM runtime, worker pool, or a combination thereof. Claim 3 A multi-user artificial intelligence collaboration orchestration system according to claim 1, wherein the Task Card object comprises at least one field among title, instruction, trigger, inputs, pipeline, outputs, allowed_tools, blocked_tools, approval_policy, run_history, artifacts, and audit_log. Claim 4 A multi-user artificial intelligence collaboration orchestration system according to claim 1, wherein the dynamic routing unit assigns the Task Card object to an asynchronous queue path or an idle time / scheduled execution path without immediately executing it when it is predicted that the token generation speed of a real-time AI session will drop below a predefined lower limit value, or when GPU / VRAM or integrated memory usage is above a predefined threshold value. Claim 5 A multi-user artificial intelligence collaboration orchestration system according to claim 1, wherein the dynamic routing unit prioritizes assigning Task Card objects, including OCR, Vision, bulk file scanning, bulk RAG search, repetitive risk report, or long report generation, to a worker LLM path or an asynchronous queue path. Claim 6 A multi-user artificial intelligence collaboration orchestration system according to claim 1, characterized in that the front LLM path is configured to process general conversation, short summaries, or UI guidance, and the worker LLM path is configured to process task card verification, OCR, vision, detailed reports, long text context processing, or expert judgment review. Claim 7 A multi-user artificial intelligence collaboration orchestration system according to claim 1, wherein the Task Run management unit generates a deterministic run ID or idempotency key to prevent duplicate execution of the same scheduled execution or retry request. Claim 8 A multi-user artificial intelligence collaboration orchestration system according to claim 1, wherein the Task Run management unit manages at least one execution state among queued, running, waiting_approval, succeeded, failed, and cancelled, and processes a timeout, retry / backoff, or user cancellation request. Claim 9 A multi-user artificial intelligence collaboration orchestration system according to claim 1, wherein the dynamic routing unit adjusts the due task enqueue frequency, the number of concurrent executions, whether OCR / Vision is allowed, whether repetitive risk reports are allowed, or whether heavy RAG is allowed. Claim 10 A multi-user artificial intelligence collaboration orchestration system according to claim 1, wherein the controlled enterprise AI execution environment is at least one of an on-premises device, a closed network device, an internal network with restricted internet connectivity, a private cloud, a VPC-based dedicated instance, a customer-specific dedicated worker pool, a hybrid environment, or a tenant-isolated SaaS instance. Claim 11 A multi-user artificial intelligence collaboration orchestration system according to claim 1, characterized in that the system is configured to use only the LLM runtime, internal RAG repository, and internal workspace within the customer organization without external LLM API calls. Claim 12 A context broker system for mediating knowledge between multiple logically isolated collaboration spaces comprises: an isolated knowledge domain management unit that manages a first collaboration space and a second collaboration space separated by an access policy; an association detection unit that detects semantic associations between a request generated in the first collaboration space and a knowledge object stored in the second collaboration space; a draft generation unit that generates an association summary token including at least one of a non-identifying reference value, an association type, a reliability, and a reason for the request regarding the semantic association, without transmitting the original data of the second collaboration space to the first collaboration space; an approval brokerage unit that privately transmits the association summary token to an authority holder of the second collaboration space and requests an explicit approval signal; a masking / generalization module that generates mediating knowledge by masking sensitive information from the original data of the second collaboration space or generalizing business knowledge only when the approval signal is received; and a knowledge brokerage unit that transmits the mediating knowledge to the first collaboration space. A context broker system characterized by including: an audit logbook that records the association detection, approval request, approval result, masking result, and brokerage history. Claim 13 A context broker system according to claim 12, characterized in that the association summary token is generated so as not to include the original sentence, original file, original conversation, or sensitive identification information of the second collaboration space. Claim 14 A context broker system according to claim 12, characterized in that the original data or brokerage knowledge of the second collaboration space is not injected into the LLM context of the first collaboration space before the approval signal is received. Claim 15 A context broker system characterized in that, in Clause 12, the authority holder is at least one of a department head, project manager, document owner, tenant manager, security manager, or policy approver. Claim 16 A context broker system according to claim 12, characterized in that the approval broker generates an approval request including an approval expiration time, and stops knowledge brokerage if an approval signal is not received within the approval expiration time. Claim 17 In Clause 12, the above-mentioned masking / generalization module is characterized by removing, replacing, or categorizing personal names, customer names, project names, amounts, contract conditions, identification numbers, location information, or security level information. Claim 18 A context broker system according to claim 12, characterized in that the knowledge broker conveys at least one of a business procedure, judgment criteria, checklist, summary knowledge, or policy-allowed reference link, rather than the entire original document. Claim 19 A context broker system according to claim 12, characterized in that the first collaboration space and the second collaboration space are logically separated by department, project, customer, tenant, security class, work subject, or access policy. Claim 20 A Zero-Trust based RAG knowledge incorporation method for incorporating external or user-provided data into an official RAG knowledge repository in a controlled enterprise AI execution environment, comprising: receiving external or user-provided data uploaded by a user to a collaboration space or file input section; storing the data in an isolated quarantine repository without immediately embedding it into an official RAG vector DB; verifying at least one of a file hash, digital signature, file type, path structure, file size, presence of a script, hidden text, and prompt injection pattern for the data; generating at least one of a primary summary report, risk score, source information, and recommended processing status for the data; displaying the primary summary report or risk score to an authorized person and waiting for approval input; normalizing, chunking, and embedding the data and incorporating it into an official RAG vector DB only when the approval input is confirmed; and storing the quarantine status, verification result, approval result, normalization result, RAG collection ID, and audit log of the data. Claim 21 A Zero-Trust based RAG knowledge incorporation method according to claim 20, characterized in that the above data and primary summary report before approval input is verified are not used as official RAG search targets for general user queries. Claim 22 A Zero-Trust based RAG knowledge incorporation method according to claim 20, characterized in that the data stored in the quarantine repository is managed with at least one status value among uploaded, quarantined, sanitized, pending_approval, approved, embedded, rejected, archived, or revoked. Claim 23 A Zero-Trust based RAG knowledge incorporation method according to claim 20, characterized in that the verification step includes the step of detecting and removing, masking, or disabling system directive disguise phrases, model directive change phrases, external URL call instructions, hidden characters, white characters, comment area text, macros, or execution scripts within the document. Claim 24 A Zero-Trust based RAG knowledge incorporation method according to claim 20, characterized in that the step of incorporating into the above-mentioned official RAG vector DB includes the step of recording together at least one of a parser recipe, a normalization rule, an embedding model identifier, a chunking rule, an approver identifier, and an approval time. Claim 25 A Zero-Trust based RAG knowledge incorporation method according to claim 20, characterized in that if the above-mentioned approval input is rejected or verification fails, the above-mentioned data is not incorporated into the official RAG vector DB but is converted to a state of rejection, storage, disposal, or review. Claim 26 A Zero-Trust based RAG knowledge incorporation method according to claim 20, characterized in that the method further includes the step of storing both the hash value before sanitization and the hash value after sanitization for use in subsequent auditing or rollback review. Claim 27 A Zero-Trust based RAG knowledge incorporation method according to claim 20, characterized in that data incorporated into the official RAG vector DB can also be switched to a revocation state by an authorized person or a policy change, and data in the revocation state is excluded from subsequent retrieval targets. Claim 28 A Zero-Trust based RAG knowledge incorporation method according to claim 20, wherein the quarantine repository operates in conjunction with at least one of a hardware-based Root of Trust, a measurement-based integrity verification, a virtualization isolation, or a gateway-based data flow restriction in a closed network or on-premises embodiment. Claim 29 A method for orchestrating multi-user AI collaboration in a controlled enterprise AI execution environment comprises: receiving natural language or file-based collaboration requests from multiple users; normalizing said collaboration requests into Task Card objects; measuring at least one of the number of active real-time AI sessions, token generation rate, GPU / VRAM or integrated memory usage, memory bandwidth load, context window usage, and queue backlog; calculating the RAG, OCR, Vision, file scanning, report generation, or long text context cost of said Task Card objects; comparing said measurement result with the calculated result to determine the execution path of said Task Card objects as one of a front LLM path, a worker LLM path, an asynchronous queue path, or an idle time / scheduled execution path; executing said Task Card objects according to said determined execution path; generating an association summary token that does not transmit original data when said Task Card objects are associated with knowledge in another isolated collaboration space, and mediating masked or generalized knowledge only after authorization; and said Task Card objects external A multi-user artificial intelligence collaboration orchestration method characterized by comprising: a step of, if a user-provided file is included, not immediately embedding the file into the official RAG vector DB but incorporating it into the official RAG vector DB only after quarantine, verification, generation of a summary report, and authorization by an authorized person; and a step of returning the execution result of the Task Card object, the knowledge brokerage result, or the RAG incorporation result to the AI ​​session, collaboration space, channel, or thread where the original request occurred, and recording an audit log. Claim 30 A non-transient computer-readable recording medium storing a program for executing the method according to paragraph 20 or 29 on a computer.

Images