Summarize always outputs spanish for gemma3 models

[SteveCaruso]

Which version of assistant are you using?

2.4.0

Which version of Nextcloud are you using?

30.0.10

Which browser are you using? In case you are using the phone App, specify the Android or iOS version and device please.

Firefox, Chromium, etc.

Describe the Bug

• Output is always in Spanish.

Expected Behavior

• Output in the same language as input.

To Reproduce

• Set up gemma3:12b via LocalAI Integration via Ollama.
• Use the Summarize tab on any arbitrary text.
• Output is always in Spanish (occasionally 1 in 10 times in Italian).

Question:
Is there any way to view or alter the Chat User Instructions for the Summary tool like there is for Chat Completions and Title Generation (at /settings/admin/ai)?

It would also be useful for the Translate tool as well, since sometimes the model will give a bullet point line-by-line translation with the original text interleaved rather than just the results text. Giving a suitable pre-prompt in the Chat With AI or Generate Text tabs (such as "just output the translation, nothing else") allows for this trivially.

[SteveCaruso]

Quick update: This also appears to happen with the Change Tone tool occasionally as well, but not as consistently.

[janepie]

Moving this to the OpenAI integration as the system prompts are defined there.
Some prompt engineering probably needed here

[edward-ly]

I did encounter this sometimes too while I was doing some prompt engineering in #157. Granted, I admit I only tested the prompts with the GPT models, so finding a prompt that works on all supported models would be a bigger (if not impossible) challenge, for sure. If nothing else, while not ideal, we could have specify multiple system prompts and select the one that works best for the chosen model. Or, maybe we can make the system prompt available to the user as an advanced option so that they can tune the output themselves.