ai schema generator

1. open the ai schema tab on any project at /dashboard/projects/<p_id>/ai-schema
2. type a description of your data model in the textarea (up to 4000 chars)
3. click generate schema. expect 5-15 seconds for a typical response — Qwen runs on a single DGX so larger prompts take longer
4. click copy on the result, paste into your project's briven/schema.ts, and review every column before committing
5. run briven deploy as usual — the AI output goes through the same schema-diff + migration path as a hand-written change

• be specific about relationships: "users have many posts, posts have many comments, comments can reply to other comments" ports cleanly. "a blog" doesn't.
• name your domain entities: "projects, tasks, time-tracking entries" beats "a productivity app"
• call out denormalised fields: "each post stores its current comment count alongside the comment rows" saves a follow-up query
• mention indexes you know you'll need: "most queries are by user_id and created_at descending" signals the right index

the model is primed with a system prompt that:

• pins the available column helpers: text(), bigint(), boolean(), timestamp(), jsonb<T>(), uuid()
• pins the modifiers: .primaryKey(), .notNull(), .default(...), .nullable(), .references(table, column), .unique()
• insists on a primary-key column per table; prefers text() for ULIDs; uses bigint() only for counters
• adds indexes only where a non-trivial query would scan — no over-indexing
• returns only the code (no markdown fences, no explanation) so it pastes cleanly into your editor

• write your functions — schema only. function bodies are too app- specific to template
• guess your auth model — every table that needs per-user scoping still needs an explicit user_id column and the function-level guard
• refactor an existing schema— the prompt assumes you're generating from scratch. for incremental changes, edit the schema by hand and let briven deploy compute the diff
• understand your data privacy constraints— review the output for anything you wouldn't store; the AI doesn't know your jurisdiction

your prompt and the generated schema are notlogged. only the prompt length, response length, model name, and elapsed milliseconds are recorded for operational monitoring. the request never leaves briven's infrastructure — the Ollama instance runs on a dedicated DGX VPS in EU-Central.

we do not train or fine-tune the model on your prompts. there is no "telemetry to improve the service" pipeline.

if the dashboard shows "AI assistant offline", the operator has not configured the Ollama endpoint (BRIVEN_OLLAMA_URL is unset on the api container). this is the default state on self-hosted briven — the feature is opt-in. self-hosters who want it should:

1. run Ollama on a machine with at least 24GB GPU VRAM (qwen2.5-coder:32b quantized fits in ~22GB)
2. ollama pull qwen2.5-coder:32b
3. set BRIVEN_OLLAMA_URL=http://your-ollama-host:11434 on the api container
4. restart the api