POML: Prompt Orchestration Markup Language(github.com) |
POML: Prompt Orchestration Markup Language(github.com) |
I'm from a small group under Microsoft Research. POML originally came from a research idea that Prompt should have a view layer like the traditional MVC architecture in the frontend system. The view layer should take care of the data, the styles and rendering logic, so that the user no longer needs to care how some table needs to be rendered, how to present few-shot examples, how to reformat the whole prompt with another syntax (e.g., from markdown to XML).
I have to admit that I spent so much time on making POML work well with VSCode, building all the auto completion, preview, hover stuff. The time is long enough that the codebase is almost becoming a monster for an individual developer to handle. The outside environment is also changing drastically. The rise of Agentic AI, tool calls, response format. The models today are no longer sensitive to small changes in prompt format as they used to. AI-aided programming can simply give you code to read in PDFs, Excels and render them in any style you want. With all that in mind, I used to feel hopeless about POML.
Nevertheless, after several months of working on another projects, I recently noticed that the view layer can be more of just a view layer. With proper user interface (e.g., a VSCode live preview), it can deliver a very smooth experience in prompt debugging, especially in a multi-prompt agent workflow. I also noticed that the "orchestration" idea can go beyond a XML-like code. I'll share more details when I had a tutorial / screenshot to share.
Going through this thread, I saw a lot of thoughts that once went through my mind. We love markdowns. We love template engines like jinja. We need those response formats. I'm thinking what is the missing piece here. I've spend so much time writing prompts and building agents in the past few months. What's my biggest pain points?
I'm quite surprised that the news hit me first before I'm ready to hit the news. If you have tried POML, please send me feedbacks. I'll see what I can do; or maybe we end up not needing a prompt language at all.
Also, please good heavens hire a narrator for the demo video. That AI voice sucks in an extremely uncanny-valley way, as if the speaker is one second from collapsing of benzodiazepine overdose, and it makes me like your work less well with every word.
First, it's cool that you work on it. Creating a new language is not an easy task.
I would suggest to try to stand on the shoulders of giants instead of trying to come up with a completely new thing.
Have a look at dhall: https://dhall-lang.org/ - it is a language that was created for cases like yours. Or, if you want to make POML a fully fledged language (and turing complete, with for-loops etc.) then it would be advised to use an existing programming language and create a DSL-like library.
See react. React did it right JSX. It might look like XML, but that's just the syntax part. You can create components in pure javascript syntax, because JSX is just the wrapper. You could do the same with POML. That will future proof it and relieve you from a lot of headache when people will ask you for more features, but without breaking backwards compat.
A compulsion to give design notes without any reasoning on something you've just heard of?
If you want them to wait until everything is super ready (or dead) and then "throw it over the fence" into their public GitHub org, keep it up.
It is great that they were allowed to open source it.
This is not unlike the way the language of legal documents is highly formulaic, structured, and codified. When precise meaning is desirable, firmer structures tend to arise. With a bit more time, proper code languages may start to appear, to help tell LLMs exactly what we mean or want.
Conventional programs using structured templates with deterministic rules to construct output is... not new.
(Jinja templates have been widely used for communicating structure to assemble conversation history, tool calls, etc., into promots for open models for a while.)
the prompt.yaml format (which this project uses) suffers from the fact that it doesn't address the structured outputs problem. Writing schemas in yaml/xml is insanely painful. But BAML just feels like writing typescript types.
I'm one of the developers!
<let name="objVar" type="object" value="{{ { key: 'value' } }}"/>
<item for="item in ['apple', 'banana', 'cherry']">{{item}}</item>
<p if="isVisible">This paragraph is visible.</p>
I guess it doesn't prevent you from doing such things, but... well... there's some eyebrow-raising shoehorning in this one.
Creating a new language that looks like XML but is not XML is... kind of unforgivable. I'd go as far and call it amateur-like. We already have good configuration languages (such as dhall-lang) and when more power is needed, then just use a real language and provide a DSL inside of it.
Anyway quite an interesting project, XML's a better fit for "programmable" data.
https://docs.anthropic.com/en/docs/build-with-claude/prompt-...
No one is suggesting you write a prompt in docx..
And in the context of LLMs, this isn't just a matter of aesthetics. More verbose tags mean more tokens, and more tokens mean higher costs. It's a perfectly valid and practical piece of feedback.
I don't think I did.
Me! _o/
Compulsion to give feedback before thinking!
So happy to be here.