1 document across the knowledge base.
Practical ways to cut LLM API spend in 2026: prompt caching, model routing, batching, streaming, and retries that do not multiply cost.
