You are viewing a plain text version of this content. The canonical link for it is here.
Posted to general@rai.apache.org by Jarek Potiuk <ja...@potiuk.com> on 2026/09/26 09:13:02 UTC
[DISCUSS] Jev in our architecture and any impact on RAI direction?
Hello everyone,
Jev has been gaining a lot of traction recently, and for good reason. It is
much faster and more cost-effective than standard LLMs for tasks that
require classifying free-text input into structured output based on a
predefined set of choices. This covers the core requirements for the vast
majority of agentic app tasks.
Here is a nice article from towards data science describing what Jev is [1]:
Airflow's common.ai package with Jev support will be released this Monday.
Our newest committer is already adding Jev support to Magpie (it fits
perfectly because a huge part of the maintainer tasks implemented in Magpie
involves classifying PRs, Issues, and Security issues with a predefined,
small set of choices): https://github.com/apache/magpie/discussions/1378 .
That might heavily improve Magpie's affordability.
This is happening just weeks after Jev was released.
And undoubtedly there will be a flood of similar services will arrive in
coming weeks - because the promise of having faster, cheaper and more
reliable classification than LLMs for specialized tasks is far too
attractive to ignore (even for frontier models). We will likely need
support for features like confidence calibration, and these features should
be generic enough to apply to other solutions.
Maybe it's a good idea if our llmao will be accompanied by jevao :) ?
[1]
https://towardsdatascience.com/jev-vs-llms-when-ai-moves-from-generation-to-decision-making
Best regards,
Jarek
Re: [DISCUSS] Jev in our architecture and any impact on RAI direction?
Posted by Andrew Musselman <ak...@apache.org>.
Good call Jarek! It's not an open-weight model so we wouldn't be able to
self-host it but we can connect to upstream providers just as easily from
llmao.
When per-project token spend budgeting is implemented in llmao Jev could be
one of several options for proprietary models projects can use.
Best
Andrew
On Sat, Sep 26, 2026 at 2:13 AM Jarek Potiuk <ja...@potiuk.com> wrote:
> Hello everyone,
>
> Jev has been gaining a lot of traction recently, and for good reason. It is
> much faster and more cost-effective than standard LLMs for tasks that
> require classifying free-text input into structured output based on a
> predefined set of choices. This covers the core requirements for the vast
> majority of agentic app tasks.
>
> Here is a nice article from towards data science describing what Jev is
> [1]:
>
> Airflow's common.ai package with Jev support will be released this Monday.
> Our newest committer is already adding Jev support to Magpie (it fits
> perfectly because a huge part of the maintainer tasks implemented in Magpie
> involves classifying PRs, Issues, and Security issues with a predefined,
> small set of choices): https://github.com/apache/magpie/discussions/1378 .
> That might heavily improve Magpie's affordability.
>
> This is happening just weeks after Jev was released.
>
> And undoubtedly there will be a flood of similar services will arrive in
> coming weeks - because the promise of having faster, cheaper and more
> reliable classification than LLMs for specialized tasks is far too
> attractive to ignore (even for frontier models). We will likely need
> support for features like confidence calibration, and these features should
> be generic enough to apply to other solutions.
>
> Maybe it's a good idea if our llmao will be accompanied by jevao :) ?
>
> [1]
>
> https://towardsdatascience.com/jev-vs-llms-when-ai-moves-from-generation-to-decision-making
>
> Best regards,
> Jarek
>