【人工智能】什么是上下文工程Context Engineering | 上下文Context | Agent的缺点 | 提示词工程 | RAG | MCP | 写入 | 选取 | 压缩 | 隔离

Understanding Context Engineering in AI Agents

Introduction to Context Engineering

• The speaker introduces the concept of AI agents, noting that while they appear advanced, they often fail during complex tasks.

• The failure is attributed not to model limitations but rather to failures in "Context Engineering."

• Context Engineering is defined as a crucial aspect of optimizing how information is provided to language models for task completion.

Defining Context

• The speaker clarifies what "context" means, emphasizing it as a comprehensive set of information given to language models for reasoning or generating tasks.

• Three core categories of context are introduced:

• Guiding Context: Sets frameworks and rules for model behavior (includes system prompts and task descriptions).

• Informational Context: Provides necessary knowledge and facts (includes RAG and memory types).

• Actionable Context: Informs the model about possible actions and their outcomes (includes tool definitions).

Understanding Context Engineering

• The speaker explains that context engineering involves designing a dynamic system that optimally assembles context for each task step.

• Tobi Lütke describes it as an art form focused on providing all necessary context for effective problem-solving by language models.

• Andrej Karpathy adds that it's both an art and science aimed at filling the context window with appropriate information.

Distinction from Other Concepts

• The relationship between context engineering, prompt engineering, and RAG is clarified; they are complementary rather than mutually exclusive.

• Prompt engineering focuses on optimizing single interactions, while RAG retrieves relevant external information.

Importance of Context Engineering

• Without effective context engineering, agents may output subpar results due to missing critical contextual information.

• An example illustrates how a lack of sufficient context can lead to ineffective responses from AI assistants when handling simple requests.

Challenges in Providing Context

• Simply providing all possible contexts isn't feasible due to limitations in model capacity and potential performance degradation over time.

• A second example highlights issues arising from excessive historical data being passed into the model during long-term tasks.

Solutions Offered by Context Engineering

• Effective context engineering aims to manage and compress contextual data intelligently, ensuring only high-value information is included.

Best Practices in Context Engineering

Key Components of Effective Context Management

Writing (Write)

• Writing involves persisting contextual data beyond immediate use through session-level or persistent writes.

Selecting (Select)

• Selection dynamically pulls relevant information before each model call, ensuring high signal-to-noise ratios. It includes deterministic selection based on preset rules or retrieval-based methods.

Compressing (Compress)

• Compression reduces the amount of data entering the context window while retaining essential signals. Techniques include auto-compression strategies used by systems like Claude Code.

Isolating (Isolate)

• Isolation sets boundaries between different streams of information within multi-agent systems, allowing sub-processes to digest details before presenting key insights back to main agents.

Conclusion on Best Practices