Python 3.9 Support?

[davidmezzetti]

Hello,

I'm the author of txtai. A couple releases ago I added support for the Transformers Agents framework and it's been working well. I recently saw that this is going to be deprecated in favor of smolagents.

I have no issue migrating as smolagents looks basically the same but it's been actively developed with some new features. The only issue is that I've always followed the same pattern as Transformers with supporting the oldest but supported Python version, which is 3.9.

This library though requires Python 3.10. So there really isn't an upgrade path unless I force txtai to require 3.10 which seems heavy handed.

Thoughts?

[albertvillanova]

Hi, @davidmezzetti! Thanks for considering migrating to smolagents: it's great to hear that Transformers Agents has been working well for you.

The decision to require Python 3.10 was made by @aymeric-roucher, so I'll ping him here to consider adding support for 3.9. While 3.10 brings some useful language features, we understand the value of maintaining compatibility with the oldest supported Python version (Python 3.9 is scheduled for end of life in October 2025).

Appreciate your thoughts on this! Let’s see if we can find a good path forward.

[sysradium]

@davidmezzetti

I agree that we should probably move on from 3.9 since it is EOL fairly soon. Even though from my quick analysis, smallagents does not use 3.9 > features that much. There was a PR which tried to introduce a match operation, but I unknowingly raised a concern about Python 3.9 incompatibility and I think it did not get merged.

Is there something fundamental about txtai that blocks you from shifting to Python 3.10? I can assist if that's just time-consuming.

[davidmezzetti]

Thank you both for the responses!

txtai runs on Python 3.10 and above just fine. The main thing is I've basically made an implicit contract with users to support the earliest supported version of Python just like Transformers does. That's why Transformers Agents was working well.

If smolagents was not going back to Python 3.9, I'd understand that decision. But I would just hope then that the transformers issue I referenced above would perhaps reconsider deprecating Transformers agents until there is a path to migration for this use case.

[davidmezzetti]

@albertvillanova @aymeric-roucher Hello - Just wanted to follow up on this. I now see the deprecation message from Transformers, so I'm wondering what path I'll need to take. Thank you.

[albertvillanova]

Hi @davidmezzetti, thanks for your feedback!

We've looked into the possibility of supporting Python 3.9 alongside 3.10, but the main challenge comes from third-party dependencies. Some examples:

• mcp (and mcpadapt) require Python >= 3.10.
• openinference-instrumentation-smolagents also requires Python >= 3.10.
• We use gradio, and some necessary fixes are only available in gradio >= 5, which requires Python >= 3.10.

Given these constraints, I think the best solution would be for you to continue using transformers.agents. Since the code is no longer updated, you can pin a version before the deprecation warning to maintain stability. Would this approach work for you?

We will keep this consideration in mind as we ensure smolagents continues to support Python 3.10 and future versions.

[davidmezzetti]

It looks like I'm between a rock and a hard place on this one. I'll go ahead and close this.

For those who run into this same issue, these are the options I'll ultimately consider.

1. Require Python >= 3.10 (Looks like the least bad one)
2. Copy transformers.agents into a new package and maintain that until Python 3.9 EOL
3. Pin Transformers (AI/ML moves too fast, would miss too much new stuff over the next 6 months)