Filament Ai Actions Laravel Package

Getting Started

• Install the package with Composer, then register its Filament panel plugin in your PanelProvider so the model() and enabled() settings can be resolved:

  ->plugin(AiActionsPlugin::make()
      ->model('gpt-4o-mini')
      ->enabled())
  

• Look first at the package README/config file for the exact action class names and required LLM credentials. Publish the config (php artisan vendor:publish) and confirm the API key and default model are set before touching actions.
• Start with a single table column action to validate the end-to-end flow (prompt → model call → saved result). A good first use case is SummarizeAction on a TextColumn:

  TextColumn::make('body')->actions([
      SummarizeAction::make(),
  ]);
  

• Once that works, add a form-side GenerateAction or TranslateAction to confirm form actions behave correctly. TranslateAction is the easiest demo of the new release because it exposes the searchable 183-language dropdown out of the box.

Implementation Patterns

• Table vs form actions. Summarize/Rewrite/Classify are most naturally used on table columns against existing records; Generate/Translate fit form fields where you want to fill or transform input before saving.
• Generate in forms. Use sibling field state to build the prompt and applyResultUsing() to decide where the output lands:

  GenerateAction::make()
      ->applyResultUsing(fn (array $state, string $result) => ['summary' => $result]);
  

• Preview before persisting. saveResult(false) turns the action into a dry-run: run it, inspect the output, and apply it yourself later. This is the right pattern for sensitive/large fields.
• Classify with enums. With structured output parsing and invalid-label rejection in v1.1.0, pass a discrete label set and persist to an enum column; invalid responses are rejected rather than silently written.
• Translate with restrictions. The full ISO 639-1 list is available by default. Restrict it per action when your domain only supports a subset:

  TranslateAction::make()->languages(['en', 'fr', 'de', 'es']);
  

• Per-action tuning. Use systemPrompt(), temperature(), and maxTokens() on individual actions so creative rewrites and strict classifications can use different settings without touching the plugin config.
• Chain with Filament conventions. Gate actions with visible()/authorize() policies, use ->afterFormFilled() or table callbacks for post-processing, and prefer applyResultUsing() over custom observers.
• Panel-wide defaults. Set model/credentials once via the panel plugin and override only where an action needs a different model or is disabled.

Gotchas and Tips

• Compatibility is promised, but verify persistence defaults. Existing Summarize/Rewrite/Classify usage should remain compatible; still confirm whether saveResult defaults to true in your version, because that changes whether results are written back to the record.
• applyResultUsing() replaces default persistence. If you supply a callback, the action won't automatically save for you — call $record->save() or return the right state explicitly.
• Plugin vs action precedence. model()/enabled() on the panel plugin are defaults; per-action systemPrompt()/temperature()/maxTokens() and any model override should win. If an action ignores your model, check the panel plugin isn't forcing it.
• Structured output needs a capable model. Classify's invalid-label rejection depends on the model honoring structured output; smaller or older models may return prose and get rejected. Test classification against your actual configured model, not just the default.
• Language dropdown is large. 183 options are fine for searchability but heavy if rendered many times; restrict with ->languages([...]) on dense tables/forms.
• Token limits and cost. Long source text plus a generous maxTokens() can hit context limits or inflate costs; set maxTokens() deliberately and consider queueing actions for bulk/batch use.
• Debugging. Log the raw model response alongside the parsed result; invalid-label errors and malformed structured output are much easier to diagnose with the raw payload. Wrap calls in try/catch or rely on Filament notifications for surfacing failures.
• Create vs edit context. Form actions used on create screens have no record, so callbacks should not assume $record exists.
• Extension points. Subclass an action to bake in a default systemPrompt()/languages(), or centralize model/credential selection in the panel plugin; avoid duplicating long prompts across resources.