# File intelligence

Opt-in per bucket. Platform-paid via Replicate by default (not BYOK unless you explicitly choose your OpenAI key).

## Controls

Per bucket you can toggle:

- **Features:** vision tags, descriptions, embeddings, similarity search, agent
- **File types:** image, video (poster/thumb only), document, audio, other
- **Spend caps:** total + daily USD
- **Auto-apply** tags/descriptions vs suggest-only (auto-apply default)

## Backfill

1. Enable / adjust toggles → **Estimate**
2. Review file counts + expected USD
3. **Confirm & start backfill** (requires `estimateToken`)
4. Pause / resume anytime

New uploads enqueue when intelligence is enabled. Jobs drain on Trigger.dev (`intelligence.drainBucket`) with a Replicate concurrency limit.

API: [Intelligence API](/docs/api/intelligence) · MCP: `get_intelligence`, `start_backfill`.

## Embeddings

Images and videos are embedded with **CLIP on the thumbnail/poster** (same vector space as text queries). Documents use CLIP text on the description/tags. Filename-only text embeds are not used for media.

## Video transcription

**Not in MVP.** Thumbnails/posters only for video. Speech-to-text is a future option.