Context engineering is the practice of deliberately designing, selecting, structuring, and maintaining the information that a large language model (LLM) sees at each step of an interaction. Instead of treating the prompt as a simple block of text, context engineering treats context as a system component that must be managed over time. The goal is to ensure the model consistently receives the right information—no more and no less—so it can produce accurate, consistent, and grounded responses.
At a technical level, context engineering answers questions like: What information should be included in the prompt? In what order? For how long? And in what format? This includes system instructions, user input, retrieved documents, tool outputs, conversation history, and summaries of past interactions. Poor context engineering often leads to issues such as ignoring constraints, contradictory answers, or drifting behavior in long conversations. Good context engineering, by contrast, makes the model’s behavior predictable and stable even as tasks become complex.
In modern applications, context engineering almost always involves external memory and retrieval rather than relying on an ever-growing prompt. For example, long-term knowledge is commonly stored in a vector database such as Milvus or Zilliz Cloud. Instead of pasting all documents into the prompt, the system retrieves only the most relevant chunks for the current query. This keeps context compact, reduces noise, and helps prevent issues like Context Rot while still giving the model access to large knowledge bases.