Replace AI output-bar language model and fix Croatian generation issues

[martinbedouret]

Context

User feedback from Kristijan Cegur reports several incorrect Croatian outputs produced by the AI output bar. The current language model appears to mishandle Croatian morphology/context and, in at least one case, seems to interpret a Croatian token using its English meaning.

We plan to replace the current language model in the next Cboard release. These examples should be used as regression cases when validating the new model.

Reported cases

1. "pet" (Croatian: five) is interpreted as English "pet"

   • Intended interaction: the user builds a phrase such as "Ja želim" + "bombon" and then selects a number to indicate quantity.
   • Numbers 2–4 work correctly.
   • When selecting 5 / "pet", the generated output can introduce "pas" (dog), apparently interpreting pet as the English noun pet rather than the Croatian number five.
   • Expected: a grammatically correct Croatian phrase expressing five candies.

2. "puzle" is changed to "puzzle"

   • Croatian input uses "puzle".
   • The AI output changes it to the English spelling "puzzle", which is then pronounced incorrectly (reported as sounding like "pucle").
   • Expected: preserve/use the correct Croatian form and pronunciation context.

3. "Ja želim" + "umetaljka" is semantically rewritten

   • "umetaljka" works correctly in isolation.
   • In combination with "Ja želim", the AI outputs "Ja želim umjetničku sliku", changing the intended meaning.
   • Expected: preserve the intended concept and produce a grammatically correct Croatian sentence containing "umetaljka".

Proposed change

Replace the current LLM used by the AI output bar with the new language model planned for the next Cboard version.

Acceptance criteria

• Validate the new model specifically in Croatian.
• Add the three examples above to manual or automated regression testing.
• The model must not translate or reinterpret Croatian words based on English homonyms.
• The model should preserve the intended AAC symbols/concepts while applying Croatian grammar.
• Verify that repeated concepts + numeric quantities produce correct Croatian morphology.
• Confirm that the reported cases no longer occur before releasing the new version.

Reference

Feedback received by email on September 21, 2026. Screenshots supplied by the user:

• Ja_zelim_puzzle.png
• Pet_ja-zelim-psa.png
• Tri_bombona.png
• Bombon_5_za-psa.png