Question
📝 **Question:** Which part of a RAG prompt should you NOT mark as cacheable?
📋 Pick the right answer.
💡 **Hint:** Re-read the theory above if unsure.
A The system prompt — it's typically the longest stable block of tokens but its size is usually dwarfed by the conversation history, making it the least cost-effective segment to cache aggressively B The user's question — it changes every call, so caching it adds the 25% write surcharge for zero benefit. Cache the stable system prompt, tools, and reference docs instead C The tool definitions — they change every time you add or remove a tool, and the write surcharge on the rest of the prompt cascades whenever tools shift, defeating the purpose of caching upstream D The retrieved document chunks — they're the largest segment by token count but they also change per query, so caching them adds the write surcharge for content the next call will never reuse anyway