https://www.boundaryml.com logo
Rewriting BAML
# general
a
g
- A lot of our PRs now are very large because they touch many golden tests, and there are golden tests for each stage of our multi-stage compiler. So a small change can touch M tests x N stages. - And, Yes. 🙂
What lead to the decision: we are working on a very large extension to BAML that allows you to write more interesting functions than just wrappers around single LLM calls. When we used this feature internally, we concluded that it needs a lot more polish and a different architecture, to be up to the dev-experience standards of tools in that category. You can expect to see many strange and exotic PRs in the coming weeks!
p
Omg, this is epic!! Greg pls do write up some stuff on the BAML blog on this when you're ready
a
I hear prompt optimisation and gepa coming to baml. 🙂 Want to read more !
p
It's already here, check out the docs 😀 https://docs.boundaryml.com/guide/baml-advanced/prompt-optimization
Currently only supports optimizing the prompts, descriptions and aliases in a single function
I'm eager for the day compound optimization is supported, then we're really cooking 😄
@Abhishek Tripathi as for why BAML is being rewritten, as you may have heard, the goal was always to be a Turing-complete language. And to make it as easy as possible for humans and LLMs to manipulate BAML code
The transition right now is from DSL to Turing-complete language. No mean feat
a
Ah. I see.
Thanks !
p
And just like any other language syntax and core features are carefully thought out, to minimize bloat and tech debt long term
So that's why some features are taking longer than others. Well, I'm patient, so I'll wait 😀