[Feature Request]: Leverage prompt caching by swapping instructions & content

[y0hnn]

What needs to be done?

We should first give the instructions in the prompt, then give the URL & Content : https://github.com/unclecode/crawl4ai/blob/main/crawl4ai/prompts.py

What problem does this solve?

We could leverage prompt caching. Depending on the provider, the cached tokens for input can ben up to 90% cheaper than normal input tokens. In my use-case, I'm crawling a lot of different pages with the same instructions set.

If we change the prompt to be :

1. Instructions
2. HTML Content
3. URL

We could also leverage on prompt caching for the HTML content which is quite the same (for example common body structure).

Target users/beneficiaries

Everyone using LLM Extraction. It would be cheaper at the end. For example with OpenAI : https://platform.openai.com/docs/pricing

Current alternatives/workarounds

No response

Proposed approach

No response

[hafezparast]

We implemented this — PR #1873.

All 4 extraction prompts in prompts.py are now reordered to put instructions before URL/HTML content, so the instruction prefix gets cached across calls.

If you want to try it before it's merged:

pip install git+https://github.com/hafezparast/crawl4ai.git@fix/prompt-caching-order-1699

To verify caching is working, check your provider's dashboard for cached token counts after running a batch extraction across multiple pages with the same LLMExtractionStrategy. Anthropic shows cache_read_input_tokens in their API response, and OpenAI shows cached_tokens in the usage object.

Would be great to hear if you see the expected cost reduction in practice — especially the actual cache hit rate across a real crawl session.

[richardchen874-sys]

Strong idea. For crawler-style extraction, the prompt structure is often more important than the model choice: keep stable instructions/cacheable schema separate from changing page content, then measure cost per extracted page rather than per request. I am testing an OpenAI-compatible multi-model API layer around official Chinese models, and prompt-caching plus model routing is exactly where these workloads can get much cheaper without changing the app architecture.