LLM Concepts

LLM concept는 agent와 wiki-style knowledge base를 지탱하는 language model, context, retrieval, workflow pattern을 설명합니다.

LLM note는 “모델이 텍스트를 생성한다”에서 멈추지 않고, 어떤 context와 constraint 아래에서 어떤 output을 만들고 어떻게 검증하는지 봅니다.

$$\hat{y}\sim p_{\theta }(y\mid I,C,T,F)$$

where $I$ is instruction, $C$ is context, $T$ is optional tool or retrieval evidence, and $F$ is output format or constraint.

LLM Contract

LLM note는 prompt trick이 아니라, 입력과 증거와 출력 제약을 분리한 contract로 써야 합니다.

$${\mathcal{L}}_{\mathrm{LLM}}=(I,C,R,D,F,V)$$

This makes a useful separation:

$$\text{fluent answer}\ne \text{grounded answer}\ne \text{verified answer}$$

Route Map

Core Concepts

LLM vs Agent Boundary

The boundary is action. A model that proposes a tool call is still an LLM concept; a workflow that executes tools, observes results, updates state, and verifies completion belongs under Agents.

$$\text{LLM}:x\to y$$ $$\text{Agent}:x\to a_t\to o_t\to s_{t+1}$$

Context-Evidence Pipeline

For LLM Wiki writing, the important path is not only token generation. It is evidence selection and claim verification.

$$\text{source}\to \text{chunk}\to \text{retrieve}\to \text{pack context}\to \text{generate}\to \text{verify}$$

RAG and Grounding Template

RAG note should preserve enough detail to debug retrieval and generation separately.

Claim Types

Common Failure Modes

Checks

• Is the model being used for generation, classification, extraction, retrieval, or tool orchestration?
• What context is provided, and what evidence is missing?
• How is token budget allocated?
• What decoding and output constraints are used?
• Are retrieved documents trusted as data, not instructions?
• Is retrieval unit, query rewriting, and reranking behavior visible enough to debug?
• Are generated claims grounded in evidence or marked to verify?
• Is the output verified outside the model?
• Is the task actually LLM-only, or does it need an agent/tool/workflow note?
• Does the claim distinguish language-model behavior from product feature behavior?
• Does the note specify context source, output format, and verification boundary?

Related

• Agents
• Agent architecture
• LLM Wiki
• Context engineering
• Inference contract
• Decoder-only Transformer
• Autoregressive model