Filament Ai Actions Laravel Package

Product Decisions This Supports

• Shipping AI-assisted content ops inside the admin panel, not as a separate product. Summarize, Rewrite, and the new Generate action let editors produce and clean up copy where the content already lives (Filament resources/forms), which removes a tool-switch and a data-sync problem from the roadmap.
• Localization roadmap acceleration. The Translate action now targets a searchable dropdown covering all 183 ISO 639-1 languages. This turns "expand to N markets" from a per-market engineering project into a configuration and editorial-workflow decision — the cost driver becomes translation review quality, not language coverage.
• Human-in-the-loop as the default, not an afterthought. saveResult(false) and applyResultUsing() let you preview, transform, or route AI output before it is persisted. That is the difference between an AI feature you can ship to a regulated or brand-sensitive customer base and one you cannot. It directly supports review queues, approval gates, and audit logging.
• Operational triage and tagging. The hardened Classify action (structured output parsing, rejection of invalid labels) makes automated routing — ticket categorization, content tagging, priority scoring — a realistic build rather than a demo. Structured output + label validation is the specific property that turns a probabilistic model into something a PM can put behind a business rule.
• Build vs buy. This is a thin, opinionated layer over a model provider inside a framework you already run. It replaces an estimated multi-sprint "provider abstraction + prompt management + Filament UI + structured output validation" workstream with a dependency. Buy, and spend the saved sprints on prompts, evaluation, and workflow design — the parts that actually differentiate.
• Configuration as a product surface. Panel plugin model() / enabled() wiring plus per-action systemPrompt(), temperature(), and maxTokens() mean cost/quality/behavior can be tuned per panel or per tenant without code changes. That supports tiered offerings (cheap model for internal drafts, stronger model for customer-facing generation) and per-tenant AI toggles for enterprise deals.
• Restricted language lists as an entitlement lever. ->languages([...]) lets you scope Translate to a plan tier or a contract's supported locales — a pricing/packaging decision expressible in config.
• Low switching cost. The release is explicitly backward compatible for existing Summarize/Rewrite/Classify usage. That materially lowers the risk of committing now and reduces the "what if we need to rip it out" objection.

When to Consider This Package

Consider it when most of the following hold:

• You are on Laravel and Filament. The value is concentrated in the Filament actions and panel plugin. Outside that stack, you are paying for integration you will not use.
• The job is text transformation, not conversation. Summarize, Rewrite, Generate, Translate, and Classify are discrete, one-shot, form-adjacent tasks. If your requirement is a streaming chat assistant, multi-turn agent, or tool-calling orchestration, this is the wrong layer.
• You want AI output to land in your own models. The result-handling API is built around previewing and applying results to your Eloquent records — good for CMS, support, catalog, and internal-ops domains.
• You need structured classification with guardrails. Invalid-label rejection and structured parsing are the reason to pick this over hand-rolling HTTP calls to a provider.
• You need broad locale coverage cheaply. 183 ISO 639-1 targets out of the box is a strong fit for content platforms with long-tail market ambitions.
• You are a small-to-mid team without a dedicated AI platform group. The package absorbs provider plumbing so product engineers can own the feature.

Look elsewhere when:

• You need retrieval, embeddings, or vector search. No indication this package addresses RAG, semantic search, or knowledge grounding. A "summarize these documents" feature backed by a corpus is an architecture decision, not an action.
• You need fine-tuning, custom model hosting, or strict data residency / on-prem inference. Provider abstraction here is a convenience, not a compliance story. Verify where prompts and content are sent before committing in health, finance, legal, or EU-resident contexts.
• You need deterministic output or hard accuracy SLAs. Classification is now more robust, but it remains probabilistic. Anything that must be exactly right needs a deterministic path or a human gate.
• Cost and rate-limit control are first-order. There is per-action maxTokens() and temperature(), but no evidence of token budgeting, per-tenant quotas, caching, or spend dashboards. At high volume you will build that yourself or pick a gateway.
• You are not on Filament, or you are mid-migration off it. The action layer is the product; without it you are adopting a thin client you could write in a day.
• You need deep observability and evaluation. Prompt versioning, A/B testing, offline evals, and trace-level logging are not in this release's scope. Plan to add them.

How to Pitch It (Stakeholders)

For executives (one paragraph). This release turns a Laravel admin panel into an AI-assisted content and operations surface without adding a new system, a new vendor integration project, or a new team. Concretely: editors can generate, rewrite, summarize, and translate content in place — across all 183 ISO 639-1 languages — and operations can auto-classify and route records with validated, structured labels. The two commercially relevant points are speed and control. Speed: localization and content-throughput capabilities that would otherwise be a multi-sprint build are now configuration. Control: results can be previewed and transformed before they are saved, so AI output passes through human review and your existing permissions rather than around them. The upgrade is backward compatible, so there is no migration cost to adopting it and no lock-in cost to reversing it.

For engineering (one paragraph). v1.1.0 adds Generate and Translate actions alongside the existing Summarize/Rewrite/Classify set, wired into Filament tables and forms and respecting the panel plugin's model() and enabled() settings. The API surface you will actually consume is deliberately small: saveResult(false) for preview-before-persist, applyResultUsing() for custom mapping into your models, and per-action systemPrompt() / temperature() / maxTokens() for behavior and cost tuning. Classify now parses structured output and rejects invalid labels, which is what makes it safe to drive routing logic. Translate defaults to the full ISO 639-1 set and can be narrowed with ->languages([...]) for plan tiers or contract scoping. Existing usage is unchanged. The gaps to plan around explicitly: no token accounting, caching, rate-limit handling, prompt versioning, or eval harness — treat those as our responsibility, and gate any PII-bearing content behind a data-flow review before rollout. Suggested first milestone: one internal, review-gated Summarize or Translate action to validate prompt quality and cost per operation before broadening.