上下文工程解析
Context Engineering, Explained
隨着上下文窗口增至數十萬個 token，團隊開始把模型接收的所有內容視為需要精心設計的上下文預算，而非只追求完美提示詞。預算包括系統提示詞、RAG 檢索文件、工具結果、對話歷史及記憶；無關內容會分散模型注意力並降低準確度。建立 RAG 或 Agent 系統時，應精簡及整理上下文，而非一味擴大模型或累積內容。
What happened: As context windows grew to hundreds of thousands of tokens, teams stopped obsessing over the perfect single prompt and started treating everything a model sees — system instructions, retrieved documents, tool results, conversation history, memory — as one deliberate budget to design.
Why it matters: A bigger window makes it tempting to dump everything in "just in case." That backfires: irrelevant context dilutes attention and measurably hurts accuracy, even when the answer is technically present in the window. More space isn't the same as better use of it.
What's actually in the context budget
- System prompt: role, constraints, output format — set once, reused every call
- Retrieved documents: the RAG chunks pulled for this specific question — relevance beats volume
- Tool results: what a function call or search actually returned — often the noisiest part of the window
- Conversation history: prior turns — prune aggressively in long agent loops, don't just append forever
- Memory: facts carried across sessions — curated, not a dump of every past interaction
Common mistakes
Pasting full documents instead of the relevant chunk. Letting tool output accumulate unpruned across a long agent loop until it crowds out the actual task. Assuming the model "sees" everything with equal weight — position and relevance both matter, not just presence.
Who should care
Anyone building RAG or agents past the demo stage. If your system prompt keeps growing and accuracy keeps dropping, the fix is usually curation, not a bigger model.
Originally published on ayraix.com, practical AI for enterprise builders.
來源：DEV Community · dev.to