Wow this is fabulous. One of those things you wonder "why didn't I do this myself?". I'm sure this is going to be big. Thanks so much for covering this!
Interesting, I’ve been building out relational neural networks that use structured data for enhanced query efficiency. Keen to see how it challenges LMMs
We don't need no frameworkation. aider's CONVENTIONS markdown text file has done all this for AI-years, by adding a single sentence: "If a prompt seems to want to write a short story, {insert system prompt or executable script path}." Epic presentation, fantastic video, love your style and explanatory precision, but every time I think a framework can do more than aider I get a cold bucket of water and realize that aider is the pinnacle of human achievement, never to be underdone. Wait until you use aider to edit it's own CONVENTIONS markdown based on user input. Or give your PHP an aider() function which can edit the self-same app's own source code.Then users can make changes to the web app's behaviour that weren't preordained by the developer. Interesting to explore.
it's nice but you save few lines of code hiding the actual interaction with openai api, which can be a limitation when you are trying to bend your code to particulary funky specs. I wonder how this library would behave when I need functions and structures output
These videos are beneficial. I have not yet had the opportunity to review the documentation in detail. Are there any prompt tuning or fine-tuning capabilities within this framework?
I was playing with ell the past days. It's a very fresh approach. Very clean code. But I find the multi-turn tool usage quite complicated and I was unable to use the structured output with OpenAI compatible APIs. I hope these two points are being addressed.
I would have understood a simplification by returning a tuple, where it can be just two strings, with the first one being system and the second user, by default, overridable by decorating it with ell.system etc. for complex situations, but using the docstring, idk if I find that clean, otherwise very interesting toolset
Thanks for finding and explaining this, its such a different model from langchain (which I like. The side-by-side comparison with openai, langchain, ell was great. Ell is so much simpler for what you showed. Would be interesting to see how to use it in complex agent architectures, like plan-and-execute or CRAG.
With a standardized syntax it would probably be possible to move towards using the web interface to construct flows into code. That would be in incredibly powerful system and spell the end of dev jobs even faster!
@@ufffd Hmmm, never is a word that doesn't age well in IT. After 35 years as a programmer I see 95% of dev jobs being replaced in 10 years, and the definition or task types of dev being so fundamentally changed that it will probably require a new title.
Ths seems like an amazing tool for building the prompts and debugging but I can't see this being used in any large project in production. You need to have full control over the API calls and often optimize on that level.
Can I store and run standardized queries on pdfs? Say I have 4 or 5 standard questions to ask of a pdf, maybe the even follow on one from the next. Can I store the queries and then load in a pdf for the queries to run on?
I would have a standard set of prompts like 1) summarise, 2) give me all the key points, lessons, takeaways, conclusions of the book, 3) what is the core problem / issue that is tackled here, 4) what solution/s are presented by the author/s to the core issues, problems, challenges identified in the book, 5) what does the book claim to be its unique, novel insight or offering, 6) summarise the book in 3 (or 5) key points, with a thesis statement, and provide me dot points as evidence to support each
Wow this is fabulous. One of those things you wonder "why didn't I do this myself?". I'm sure this is going to be big. Thanks so much for covering this!
And then you see the tooling quality and think “ahhhh, I’m glad I didn’t just roll my own thing for this”
@@TheUserbladewise words
thanks for introducing a new tool for my work! Subscribed
I love this as much as I hate the long lines in my code for a simple prompt.
How is this not known out there??? Great find.
Interesting, I’ve been building out relational neural networks that use structured data for enhanced query efficiency.
Keen to see how it challenges LMMs
Great! this looks super slick! will use in one of the project I am working on!
You're very resourceful Ian! Nice find.
Ell puts into code how I've been thinking about prompts and LLM interactions, as functions. Its very elegant, can't wait to see where it goes.
I cannot wait to give this a spin. Ell is so Elegant. :D
this combined with Langroid would be so wonderful. I guess i need to refactor my code again.
😂😂
We don't need no frameworkation. aider's CONVENTIONS markdown text file has done all this for AI-years, by adding a single sentence: "If a prompt seems to want to write a short story, {insert system prompt or executable script path}." Epic presentation, fantastic video, love your style and explanatory precision, but every time I think a framework can do more than aider I get a cold bucket of water and realize that aider is the pinnacle of human achievement, never to be underdone.
Wait until you use aider to edit it's own CONVENTIONS markdown based on user input. Or give your PHP an aider() function which can edit the self-same app's own source code.Then users can make changes to the web app's behaviour that weren't preordained by the developer. Interesting to explore.
it's nice but you save few lines of code hiding the actual interaction with openai api, which can be a limitation when you are trying to bend your code to particulary funky specs. I wonder how this library would behave when I need functions and structures output
Excellent framework review, Ian. Thanks for sharing.
I was literally planning to build something like this. Thank you!
Excellent demo!
These videos are beneficial. I have not yet had the opportunity to review the documentation in detail. Are there any prompt tuning or fine-tuning capabilities within this framework?
I was playing with ell the past days. It's a very fresh approach. Very clean code. But I find the multi-turn tool usage quite complicated and I was unable to use the structured output with OpenAI compatible APIs. I hope these two points are being addressed.
love it, thanks for sharing!! a new follower to your content! :)
Fantastic Ian
Can we have Typescript support please!
I would have understood a simplification by returning a tuple, where it can be just two strings, with the first one being system and the second user, by default, overridable by decorating it with ell.system etc. for complex situations, but using the docstring, idk if I find that clean, otherwise very interesting toolset
Thanks for finding and explaining this, its such a different model from langchain (which I like. The side-by-side comparison with openai, langchain, ell was great. Ell is so much simpler for what you showed. Would be interesting to see how to use it in complex agent architectures, like plan-and-execute or CRAG.
Cheers, yes exploring more complex use cases is something I'm going to be doing in the coming weeks.
With a standardized syntax it would probably be possible to move towards using the web interface to construct flows into code. That would be in incredibly powerful system and spell the end of dev jobs even faster!
dev jobs will never end
@@ufffd Hmmm, never is a word that doesn't age well in IT. After 35 years as a programmer I see 95% of dev jobs being replaced in 10 years, and the definition or task types of dev being so fundamentally changed that it will probably require a new title.
@@74Gee We'll be "software coordinators" then...
Her name was Lila
She was a coder
.....
with yellow feathers in her hair...
how can I use my own client? as many company has their own client originate from openai. I try to use it, but it tell me absent of provider
Why not use DSPy?
Ths seems like an amazing tool for building the prompts and debugging but I can't see this being used in any large project in production. You need to have full control over the API calls and often optimize on that level.
From the github repo, I think that's it's intended use case - which it works really well at.
Can I store and run standardized queries on pdfs? Say I have 4 or 5 standard questions to ask of a pdf, maybe the even follow on one from the next. Can I store the queries and then load in a pdf for the queries to run on?
I'm curious what are your standard questions
I like, summarize and include specific details
I would have a standard set of prompts like 1) summarise, 2) give me all the key points, lessons, takeaways, conclusions of the book, 3) what is the core problem / issue that is tackled here, 4) what solution/s are presented by the author/s to the core issues, problems, challenges identified in the book, 5) what does the book claim to be its unique, novel insight or offering, 6) summarise the book in 3 (or 5) key points, with a thesis statement, and provide me dot points as evidence to support each
how do we stream?
It's awesome and it's missing a context manager
Looks amazing. I just wish it worked with other languages besides just python.
It is amazing. Not sure how viable other languages are this early on.
Richard Mountains
Wow
soy devs are gonna love this
Schaden Ways
prompt engineer joke. Engineer, sure! More like prompt monkey.