วิศวกรรมพรอมต์และหน้าต่างบริบท

Every LLM has a fixed context window — the max tokens it reads and writes per call. System prompt, retrieved docs, history, and your question all have to fit.

A production RAG prompt packs in system instructions, retrieved context, prior history, and the current question. Exceed the window and something gets cut.

A clear prompt structure helps the model tell instructions apart from data. Here's a clean layout separating context from the question.

To cut hallucination, add grounding instructions: tell the model to answer only from the provided context and to admit when it doesn't know.

A prompt template makes prompts reusable and safe to fill with variables. LangChain's ChatPromptTemplate does exactly this.

Before sending, count tokens so you don't overflow the window. Rough rule for English: about 4 characters per token.

When context is too big, shrink it: fewer chunks, smaller chunk sizes, or summarize. Never silently truncate the middle — you might drop the answer.

Watch the lost-in-the-middle effect: models attend best to the start and end of context, worst to the middle. Put your most relevant chunks first or last.

Input and output share the window. Fill it all with input and there's no room to generate — so reserve output space for the expected answer length.

A few few-shot examples can steer format and tone, but they eat tokens. Weigh their value against the space they consume.

Prompt engineering is empirical: change one thing at a time, test on real questions, and measure. Small wording tweaks can shift answer quality a lot.

Test your understanding of context windows.

Recap: the context window holds system, context, history, question, and output. Structure prompts, ground them, count tokens, beat lost-in-the-middle, and reserve output room.