RAG & Security
An enterprise AI security platform that brings Korean-language document handling, vector encryption, and knowledge graphs together
A summary of the core technology for reviewers and practitioners. Detailed specifications are provided in a separate proposal.
The data processing pipeline.
Upload
Bulk-load documents with a security level assigned.
Document parsing
Parsed by 13 format-specific engines.
Chunking
Split into chunks at Korean sentence boundaries.
Semantic conversion
Sentences are converted into retrieval vectors.
Indexing
Semantic and keyword indexes are built at the same time.
Retrieval accuracy and information security are designed at the same layer.
A single query passes through four layers in order. Security level determines the retrieval candidates; the quality gate determines the final answer.
Hybrid retrieval
Unified query across semantic search (vectors) and exact keyword matching (BM25)
Multi-document reasoning across a knowledge graph (GraphRAG)
Re-ranking narrows 50 candidates down to 8 precise ones
RAG-Guard vector security
AES-256-GCM with our own vector transform (SRXT)
Search runs while the data stays encrypted, with no loss of result quality
Per-customer independent encryption keys · processed in a trusted execution environment
Security levels and access control
Four levels — public, department, organisation, confidential — aligned to the NIS N2SF standard
Permissions filtered in real time at query time
Tamper-evident retention of access logs · integrated with a cryptographic verification module
Answer quality and safeguards
10-stage quality reinforcement · automatic query expansion (HyDE)
Prompt-injection detection and blocking · automatic de-identification of personal data
Citation verification plus automatic scoring on three answer-quality metrics
Your security policy decides how it runs
Documents and models both run on your own servers. Air-gapped deployments are supported.
Documents stay in-house and only inference is separated. Routes are split by security level.
Start immediately on a tenant dedicated to your company. Keys and indexes are kept separate.
Choose between our own lightweight model (xLLM) and open LLMs, and assign a different model to each security level.
Large document sets are processed concurrently, and repeat queries are answered straight from cache.