Select representative documents, questions and permitted actions.
02
Measure each stage
Assess extraction, indexing, retrieval and generation separately.
03
Agree operations
Set permissions, logging, latency targets and error handling.
Practice questions · Proposed by the research center
Map external APIs and telemetry
Apply document permissions to retrieval results
Provide a way to report and correct wrong answers
Assign responsibility for model updates and patching
REFERENCE ARCHITECTURE / PROPOSED SCOPE
Beyond a model. Toward a system you can evaluate.
SYSTEM DESIGN / 01—07
A reference scope for evaluating an institutional local LLM project. Actual delivery, performance and acceptance conditions depend on the institution’s documents, security environment and agreed requirements. These items do not claim a completed deployment or certification.
INTERACTIVE REFERENCE / 01—07
Select an area to explore its scope and review criteria.
REFERENCE MODULE
04
Inference for institutional infrastructure
Review models, hardware, concurrent use and external connections as a single operating scenario.
Model terms, import checks and update procedures
Methods for measuring task latency and concurrent use
An inventory of external API and telemetry paths
REFERENCE MODULE
03
Permission-aware retrieval
Define how document access rights carry through indexing and retrieval results.
Mapping users, departments and document permissions
Checks after access changes or revocation
Links between answer evidence and document versions
REFERENCE MODULE
01
Document intake and provenance
Design a document path that connects originals, versions and processing history to each answer.
Rules for supported, encrypted and damaged files
Original retention, deduplication and version identity
Failure records and controlled reprocessing
REFERENCE MODULE
02
Korean document structure
Evaluate relationships between text, tables, notes and images using representative administrative documents.
Comparison samples for merged cells, columns and notes
Identification of OCR needs and extraction omissions
Evidence links back to source pages or sections
REFERENCE MODULE
05
Quality tested against real work
Define how to assess evidence accuracy, omissions, wrong answers and human correction effort.
Agreed documents, questions and reference answers
Separate records of retrieval and generation failures
Deferral or clarification when evidence is insufficient
REFERENCE MODULE
06
Observable operations and accountability
Define logging, correction and incident-response boundaries and owners before adoption.
Log fields, sensitive-data handling and retention
An owned path from error reports to corrections
Recorded incident, backup and recovery checks
REFERENCE MODULE
07
Delivery, handover and accessibility
Plan the handover needed to verify, change and operate the system after installation.
Installation, environment and operations documentation
Approval, release and rollback procedures
Keyboard, zoom, contrast and reduced-motion checks
View the complete module checklist
04LOCAL INFERENCE
Inference for institutional infrastructure
Review models, hardware, concurrent use and external connections as a single operating scenario.
Model terms, import checks and update procedures
Methods for measuring task latency and concurrent use
An inventory of external API and telemetry paths
03GOVERNED RETRIEVAL
Permission-aware retrieval
Define how document access rights carry through indexing and retrieval results.
Mapping users, departments and document permissions
Checks after access changes or revocation
Links between answer evidence and document versions
01DOCUMENT INTAKE
Document intake and provenance
Design a document path that connects originals, versions and processing history to each answer.
Rules for supported, encrypted and damaged files
Original retention, deduplication and version identity
Failure records and controlled reprocessing
02HWPX STRUCTURE
Korean document structure
Evaluate relationships between text, tables, notes and images using representative administrative documents.
Comparison samples for merged cells, columns and notes
Identification of OCR needs and extraction omissions
Evidence links back to source pages or sections
05EVIDENCE EVALUATION
Quality tested against real work
Define how to assess evidence accuracy, omissions, wrong answers and human correction effort.
Agreed documents, questions and reference answers
Separate records of retrieval and generation failures
Deferral or clarification when evidence is insufficient
06OPERATIONS & AUDIT
Observable operations and accountability
Define logging, correction and incident-response boundaries and owners before adoption.
Log fields, sensitive-data handling and retention
An owned path from error reports to corrections
Recorded incident, backup and recovery checks
07DELIVERY & ACCESS
Delivery, handover and accessibility
Plan the handover needed to verify, change and operate the system after installation.
Installation, environment and operations documentation
Approval, release and rollback procedures
Keyboard, zoom, contrast and reduced-motion checks
DELIVERY FRAMEWORK
From an idea to acceptance and operation.
REFERENCE DELIVERY PLAN
01 / DEFINE
Define the task and boundaries
Identify users, documents, permissions and data paths, then select the priority workflow.
Requirements, data flows and evaluation plan02 / PROVE
Prove the approach with samples
Evaluate extraction, retrieval and answers separately and document their limits.
Sample results, issues and improvement priorities03 / INTEGRATE
Integrate access and operations
Connect permissions, logs, approvals and error handling in the agreed environment.
Integrated configuration, procedures and acceptance scenarios04 / TRANSFER
Accept and transfer operations
Review acceptance evidence and agree training, documentation and ongoing support.
Acceptance records, handover package and ownership matrix
Acceptance should rest on reproducible evidence.
Can the agreed documents, questions and test environment be rerun?
Have document-level access changes and revocation been checked?
Are source evidence and error-handling records retained?
Have release, recovery and operations documents and owners been handed over?
QUESTIONS & ANSWERS
About this resource
Is this an official K-Moonshot website?
An independent private information and analysis service by Sejong AI Research Center. Not a government website or the official operator of K-Moonshot. Facts, media reports and our analysis are labeled separately.
Will a new article appear every day?
We aim to review sources daily. Briefs are added or corrected when a verified substantive development warrants it. No-news days do not generate filler articles, and the displayed review date follows actual source checks.
Where do consultation links go?
They lead to the existing Sejong AI Research Center contact page at SEJONG-AI.COM. This information site does not collect inquiry details or documents.
LOCAL LLM · ON-PREMISE
Before choosing a local LLM, define the conditions of the task.
Describe your documents, tasks, network constraints and expected concurrency. Discuss the deployment scope, evaluation and operating responsibilities.
You can start with the current constraints before selecting hardware or a model.Prepare your inquiry
We are evaluating a local LLM in [network/on-premise environment]. Expected concurrency is [range], tasks are [tasks], available hardware is [hardware/undecided], and document permissions use [approach]. We would like to discuss scope, evaluation and operations.