Customizable AI templates

[InAnYan]

Now, all LLM templates can be customized by users. Previously, only the system message was customizable in a limited way.

I used Apache Velocity template engine.

Mandatory checks

- [ ] Change in CHANGELOG.md described in a way that is understandable for the average user (if applicable)
- [ ] Tests created for changes (if applicable)

• Manually tested changed features in running JabRef (always required)
• Screenshots added in PR description (for UI changes)
• Checked developer's documentation: Is the information available and up to date? If not, I outlined it in this pull request.
• Checked documentation: Is the information available and up to date? If not, I created an issue at https://github.com/JabRef/user-documentation/issues or, even better, I submitted a pull request to the documentation repository.

[koppor]

Side comment: This refs koppor#392, where we first wanted Apache Velocity - but then wanted Thymeleaf (and did not want Apache FreeMarker).

[calixtus]

I think it I necessary now to finally create an ADR now for the template engine. Please create a simple one, with your decision drivers for velocity and we can fill in the others for the next dev-call to decide on it together.

I think we need to decide on it, since we should not have multiple templating engines running in favor of resources.

[koppor]

Minor comments.

Regarding the wording...

Should it be

• User message for a summary of a chunk of text
• User message for a combinding chunk summaries

?

[koppor]

I don't understand that - this should be generated?

[InAnYan]

About that. There are two ways of asking AI to generate text:

1. In form of chat: a list of messages. For this situation we need both system + user message templates.
2. In form of text that should be completed. This is only one template + I've seen that this approach is typically chosen for summarization. And I think it's easier.

For the summarization I chose 2nd approach. That is why you are seeing "OVERVIEW" and "FINAL OVERVIEW"

[InAnYan]

Because without those words AI won't understand what should it do

[koppor]

I don't understand that - the FINAL OVERVIEW should be generated? Why is it part of the prompt?

[HoussemNasri]

We can actually define fields in enums, so here we can define a field called localizedName and define the relevant constructor. After that we can write something like this: CHATTING_SYSTEM_MESSAGE(Localization.lang("System message for chatting"))

[InAnYan]

But I'm not sure, if the names are properties of templates, and especially localized ones

[HoussemNasri]

If that is the case we can do something else, but if all strings are static having it as a field is better IMO.

[koppor]

Just call Localization.lang in the enum. Why not? -- The initialization of the enum should take care about the localiaztion. When using toString, the stored localized string is returned.

I know this will cause troubles when users change the UI language. Since language switch is not done that often, I accept this.

---

I think, a related discussion was going on at #8304 and ADR-0023.

[koppor]

I think it I necessary now to finally create an ADR now for the template engine.

Coming from #11900, Velocity was the right choice. We need to improve the ADR itself, but it is good enough to show that this was the right choice.

[koppor]

@InAnYan Can you resolve the conflicts? I think, then, this can go in - and JabRef/user-documentation#519 can also be merged?

[InAnYan]

Cool, the ADR is merged!

I'll fix everything, but in the evening

[koppor]

Comments on naming.

Moreover, can you change from dropdown to tabs?

JabRef uses tabs as UX pattern for options. (Moreover, it is one click for change to another template - for instanc,e when trying to make them consistent, comparing them, ... this is useful)