Representing LLM inputs as HTML instead of JSON to cut input tokens by 11%

blog.skyvern.com

15 pointssuchintan2 comments

Very interesting! I'd love to see a breakdown of various formats including YAML, TOML, CSV, and HCL with their relative cost vs accuracy.

I suspect accuracy is going to vary between domains that LLMs are being applied to

Interesting call about validating this against YAML / TOML / CSV / HCL. HTML was natural for us because we're dealing with web data, so the translation ended up being surprisingly intuitive

Reply on news.ycombinator.com