Skip to main content

Context Engineering (上下文工程)

How much contextual information AI can access and utilize—from single conversations to self-evolving knowledge.

Overview​

This dimension evaluates AI's ability to access, understand, and utilize contextual information. Context is a key determinant of AI coding quality—how much information AI can "see" directly determines the quality of code it can generate.

Levels​

Level 1: Ephemeral Context (瞬时上下文)​

Single conversation, no memory, all manual. Each interaction with AI is independent, with no context memory. All background information must be manually provided, resulting in low efficiency and easy omission of key information.

Level 2: File-Level Context (文件级上下文)​

Current file and nearby tabs. AI can sense the content of currently open files and adjacent tabs, providing file-level context understanding, but cannot correlate information across files.

Level 3: Repo-Level Retrieval (仓库级检索)​

Repo-level RAG, search historical Issues/PRs. Through RAG technology, repository-level knowledge retrieval is achieved, able to utilize information from historical Issues and PRs, greatly improving context breadth and depth.

Level 4: Agent Readability (智能体可读性)​

Codebase as record, "directory maps + progressive disclosure". The codebase itself becomes the agent's knowledge source, optimizing context acquisition efficiency through structured directory organization and progressive information disclosure.

Level 5: Self-Evolving Knowledge (知识自我演进)​

Dedicated "knowledge agents" autonomously maintain system memory. Dedicated knowledge management agents are responsible for continuously updating and maintaining the system knowledge base, achieving self-evolving context and forming organization-level AI memory.