Refactor guardrails once they are moved upstream

[edeandrea]

I'm planning to move the guardrail implementation upstream to LangChain4j (see langchain4j/langchain4j#2549). Once implemented there I'd like to retrofit the Quarkus implementation to re-use what's done in upstream.

This will take some time of course, so I'm just creating this now as a placeholder for future work.

[edeandrea]

@cescoffier / @geoand Should we deprecate at least the existing guardrail annotations & interfaces, as they will most likely be replaced by similar things upstream? That way people know that something is coming?

If so, I can create a sub-issue to do that work now.

Or if you prefer we can hold off until the upstream implementation is done (or closer to done).

[geoand]

Let's not deprecate until upstream has something in place

[edeandrea]

Also tracking langchain4j/langchain4j#2571

[edeandrea]

@geoand / @mariofusco Now that guardrails are upstream I will start working on this in the next couple of weeks.

Question - should we leave the existing Quarkus implementation there and just deprecate for a few versions leaving both implementations? That would allow users an opportunity to migrate (by migrate, essentially just change package names).

Or do we want to rip the existing out & replace?

[cescoffier]

The problem is the behavioral difference for streaming. I don't believe we can use the upstream implementation, or we would have to find a way to implement our current behavior.

We can start using the upstream annotation and hopefully the input guardrail implementation.

[edeandrea]

Yes we will have to do something using upstream on `Multi`. I think we can make it work.

[edeandrea]

@cescoffier Do you have an opinion on this?

Question - should we leave the existing Quarkus implementation there and just deprecate for a few versions leaving both implementations? That would allow users an opportunity to migrate (by migrate, essentially just change package names).

Or do we want to rip the existing out & replace?

I'm thinking we leave the existing stuff there for now & deprecate it for a few releases before ripping it out.

[cescoffier]

If we can handle multi using upstream it would be great. However:

• complete accumulation is a no go
• you cannot implement back pressure with plain listeners (without blocking)