AI Features

Ondros ships guideline-aware AI: every feature retrieves relevant passages from your ingested brand guidelines (vector search when embeddings are available, keyword retrieval otherwise) and grounds the model in them. Works with free providers — Groq, Gemini, Ollama (fully local), OpenRouter — as well as OpenAI/Azure. GET /ai/status reports the active provider and retrieval mode.

All endpoints are management-plane (Authorization: Bearer <JWT>).

Generate an entry from a brief

POST /ai/generate
{ "space_id": "…", "environment": "master", "content_type_id": "…",
  "brief": "Landing page announcing our Q3 developer conference" }

Returns a complete fields draft matching the content-type schema — including nested assembly blocks — ready to review in the editor's AI sidebar.

Rewrite / shorten / expand / tone

POST /ai/rewrite
{ "text": "…", "instruction": "shorten" }   // or "expand", "seo tone", free-form

The editor exposes these as one-click actions on any text field.

Title suggestions & SEO metadata

• POST /ai/suggest-titles — headline options for the current draft.
• POST /ai/seo-meta — meta title + description within length limits.

Compliance check

POST /ai/check-compliance
{ "space_id": "…", "text": "…" }

Audits copy against your ingested guidelines and returns pass/fail findings with the guideline passages that triggered them.

Translate between locales

POST /ai/translate-fields
{ "entry_id": "…", "source_locale": "en-US", "target_locale": "hi-IN" }

Translates every localized field of the entry between any pair of active locales — the editor's locale tabs surface this as "AI translate".

Guidelines (RAG corpus)

Upload brand/style documents under Guidelines in the editor; they're chunked, embedded (pgvector), and retrieved per request. No guidelines ingested? Features still work — they just run without brand grounding.

AI endpoints return 503 when no provider is configured. Set AI_PROVIDER (+ AI_API_KEY) in the backend environment — the free options are listed in the repo's .env.example.