A system prompt establishes the model’s operating behavior before it performs a task. In retrieval systems, it can shift query generation from shallow paraphrasing toward deliberate coverage of different facets and sub-intents, improving recall without changing the retriever itself. section 8.2
Effective system prompts make the intended behavior explicit:
- Specify the task and the dimensions to cover, such as distinct facets, sub-intents, or interpretations. section 8.2
- Preserve domain-critical details by naming what to extract, what to omit, and which terms or relationships require exact wording. section 10.6
- Define evaluation criteria and explicit reasoning steps when the prompt governs an evaluator rather than a generator. section 14.3.1
From vague rewriting to facet coverage
For “What is the best laptop for college?”, a weak prompt produces near-duplicate rewrites. A stronger system prompt requests distinct angles, yielding queries about durability, portability, price, and battery life. Each query can retrieve a different document set, widening coverage while remaining tied to the original intent. section 8.2


