KDG Technical Paper
What is the 4-Layer Knowledge Architecture
This paper describes the technical structure of KDG's core 4-layer knowledge architecture. It introduces the 'Text-Driven Principle' — which keeps domain knowledge as human-readable text rather than locking it into black boxes — and explains the role of each of the four knowledge layers (Original / Formal / Insight / Meta), comparing them with common AI approaches.
1.What is the 4-Layer Knowledge Architecture
1-1KDG's Design Philosophy
KDG is a framework for organizations to leverage their domain knowledge through AI. At its core is the philosophy of treating knowledge not as material for inference, but as the foundation for deterministic execution. Business judgment, expert consultation, technical analysis — many AI use cases in organizations aim to systematize tacit expertise into a form where anyone can execute it with consistent quality. Achieving this requires explicitly managing where and in what form the knowledge exists.
1-2The Text-Driven Principle
KDG's most important design principle is the Text-Driven Principle: organizational domain knowledge must be explicitly managed as human-readable and editable Markdown text — not locked into forms like model weights or vector spaces that humans cannot directly read. This principle is grounded in three requirements that cannot be avoided when using AI in business operations.
Visibility
Knowledge must be verifiable and auditable by anyone in the organization.
Editability
Knowledge updates must be reflected instantly through organizational decisions alone.
Accountability
The basis for AI conclusions must be explainable to applicants, users, and auditors.
Making domain knowledge visible to the customer — this is KDG's core philosophy. By avoiding black-boxing and returning organizational knowledge to the hands of the organization, AI can be integrated into business operations as a transparent collaborative tool, rather than an inscrutable judgment engine.
1-3Structure of the 4-Layer Knowledge Architecture
KDG implements the Text-Driven Principle through four knowledge layers. Each layer is responsible for a different aspect of organizational domain knowledge. All four layers are managed as human-readable Markdown text.
- ▸
Original Layer (Source Retention)
Retains source data owned by the organization (regulations, manuals, daily reports, past cases) as-is. Always serves as the wellspring of context.
- ▸
Formal Layer (Structured Data)
Organizes source text into structured formats such as decision rules, procedures, and patterns. Provides the criteria that AI directly references at execution time.
- ▸
Insight Layer (Derived Knowledge)
Extracts relationships and patterns across multiple items and data as automatic decision rules. Turns organizational experience into reusable assets.
- ▸
Meta Layer (Values & Policy)
Defines the organization's decision criteria, priorities, and policies. The highest-priority layer that can override decisions from all other layers.
1-4Flexibility Across Domains
The 4-layer knowledge architecture is not designed for any specific business domain. It offers flexibility that extends to any business where knowledge representation is needed. KDG's product family deploys this architecture across a wide range of business applications.
- ▸BOM cross-analysis in manufacturing
- ▸Knowledge management for large-scale software development
- ▸High-expertise estimation
- ▸Intelligent monitoring
- ▸Customer service
- ▸BPO workflow analysis
All of these are business operations that can only deliver value when organizational domain knowledge is explicitly managed. The 4-layer knowledge architecture provides a consistent answer to the common challenge of knowledge structuring, regardless of the specific business domain.
2.Comparison with Common AI Approaches
The major approaches for leveraging organizational knowledge with AI can be broadly classified into five categories. This chapter summarizes the characteristics of each and shows where the 4-layer knowledge architecture stands in relation to them.
Machine Learning (Domain-Specific Models)
An approach that trains a classification model on historical data and outputs decisions for inputs. Because knowledge is embedded in model weight parameters, it is difficult for humans to verify what the decision is based on. Moreover, retraining is required every time the decision criteria change.
ML + LLM Pipeline
A two-stage configuration where an ML model handles decisions and an LLM handles explanations. Since the LLM writes 'plausible reasons' without knowing the basis for the decision, the decision logic and the explanation logic are structurally separated.
RAG + LLM
Documents are chunked and vectorized, and similarity search results are passed to an LLM for inference. Because the knowledge being searched exists in vector space and the LLM performs inference on each query, decision reproducibility is limited.
Graph-RAG
Manages knowledge in graph structures, preserving relationships between entities. It enables more structural search than RAG, but 'judgment criteria themselves' are difficult to express as edges, and final decisions still require LLM inference.
LoRA / Fine-Tuning
Fine-tunes an LLM with domain-specific knowledge. Since the learned knowledge is buried in model weights, humans cannot audit what was learned, and retraining is required every time policies change.
2-2The Common Flaw: Black-Boxing of Knowledge
These five approaches, despite differing technical methodologies, share a common structural feature: knowledge is locked in some form that humans cannot directly read. Model weights, vector spaces, graph-based inference results — in every case, the knowledge itself cannot be verified or edited by humans in the organization.
| Approach | Knowledge Form | Decision Basis | Human Editable |
|---|---|---|---|
| Machine Learning | Model weights | Probability score | × |
| ML + LLM | Model + general knowledge | Post-hoc writing | × |
| RAG + LLM | Vector space | Similarity + inference | × |
| Graph-RAG | Graph + inference | Relationships + inference | △ |
| LoRA / FT | Model weights | Trained inference | × |
| KDG | Markdown text | Explicit rule matching | ○ |
Table 2-1: Knowledge form and editability across approaches
2-3Position of the 4-Layer Knowledge Architecture
The 4-layer knowledge architecture is a design that manages knowledge explicitly as human-readable Markdown text, rather than locking it into vectors or model weights. By using the LLM's excellent text comprehension capability for 'reading and applying explicit rules' rather than 'inferring rules,' it achieves deterministic execution while leveraging the LLM's strengths. This is not a design that denies LLM capabilities. On the contrary, it is a design that deploys LLM capabilities in their most effective form. Only when knowledge exists in a form visible to humans can an organization integrate AI into its operations as a trustworthy tool.