I use baml for a multimodal ai feature extraction pipeline (for making ecommerce product listings based on product images). I see 'Jev' is this new ... thing? I dont' really understand it. Is it hype or the real deal? It's good for structured outputs but I don't yet understand how it would benefit me and how would it integrate into pipeline with BAML? I use structured enums for the feature extraction, usually a gemini LLM. Would I just leave same BAML prompts in place but instead route to Jev instead of gemini for structured enums? Or would it supplement? Could you guys do examples or talk about it in the next ai that works talk? Seems like Jev is text-only - so you'd still need a multimodal LLM like gemini to extract features, store them as text - like free text - then have Jev do the structured enums classification? So you'd reduce the prompt payload and input tokens costs to gemini - use it to strictly output free text. Then the gemini completion is used to create the Jev prompt which does the structured enums classification output? And then there are only 255 possible enums to use in Jev, so for >255 you'd have to run recursive decision-tree layers of Jev prompts? Not sure how this plays out.

i was actually thinking about doing our next AI That Works episode thing on jev possibly!

Sounds good man, can't wait to watch it

I just got access. Curious to see what you think, @Vaibhav (BAML 🐑)

Jev only really clicked for me with this demo https://www.loom.com/share/18c4dbcf8db546dfb2d7f2ef018e78e4 can't wait for Vaibhav's take on it! 🚀 Wondering if BAML lang would be a great substrate for composing Jev into AI systems just like it is for LLMs

soon 🙂 https://cdn.discordapp.com/attachments/1549804475424841760/1550400700247445544/Screenshot_2026-09-17_at_11.58.44_PM.png?ex=6aae32ab&is=6aace12b&hm=44883425de50c449990fce2d3312e17e48a457fa71f2dbda80d1362a8363bcaa&

Types. So hot right now.