Reference
Cloudflare AI capability map
Separate Workers AI, Vectorize, AI Gateway, and Agents before composing an AI application.
Edited and verified by Orange Book Editorial Team ·
REFERENCEAI BUILDERS18 minutesVerified 2026-08-26
| Capability | Solves | Is not |
|---|---|---|
| Workers AI | Runs Cloudflare-hosted models through a Worker binding | A model catalog frozen forever |
| Vectorize | Stores and queries vectors for semantic retrieval | The only store for source documents |
| AI Gateway | Centralizes model traffic, observability, and policy | A guarantee that answers are correct |
| Agents | Builds stateful, interactive agent applications | A mandatory layer for every chat page |
Three starting points
One inference call
Start with a Worker and Workers AI binding; measure input, output, latency, and failure.
Your own knowledge
Add Vectorize and map every match back to auditable source text.
Multi-provider governance
Evaluate AI Gateway after defining secrets, logs, caching, and data-retention boundaries.
AI application safety floor
- Treat model output as untrusted text: escape it for display and authorize before action.
- Do not log user questions, private documents, or tokens by default.
- Say that no context was found instead of inviting fabricated citations.
- Recheck model names, prices, and quotas against current official pages before release.
Next: build the bilingual knowledge assistant.
Primary sources
Did this page help you complete your goal?
Beta feedback is generated in this browser and is never uploaded automatically.
Technology guide network for runtimes, frameworks, and data
Continue from Cloudflare platform boundaries into Bun, Deno, TanStack, PostgreSQL, MariaDB, and Valkey, then return to an Orange Book implementation path.
Verifying plans, limits, and cost
Combine official limits, an observed request model, and budget alerts instead of treating a free tier as an architecture guarantee.