Models
Flarepipe: Chat uses two Workers AI models: one to turn text into vectors, and one to write answers. Both are settings rather than hard-coded values, which matters more than it sounds.
The defaults
| Role | Default model |
|---|---|
| Answering | @cf/meta/llama-4-scout-17b-16e-instruct |
| Fallback | @cf/meta/llama-3.3-70b-instruct-fp8-fast |
| Embedding | @cf/google/embeddinggemma-300m (768 dimensions) |
The answering and fallback models are editable in Flarepipe Chat > Settings. They are free-text fields taking any Workers AI model identifier, so you are not waiting on us to add an option.
Why there is a fallback
Cloudflare retires Workers AI models. When that happens, a chatbot with one hard-coded model identifier simply starts returning errors, and on a distributed plugin there would be no way to hot-fix thousands of installations at once. So the Worker tries the primary model, and on failure tries the fallback before giving up.
Alongside that, a daily health check exercises the Worker and its bindings and raises a WordPress admin notice the moment something breaks. A chatbot that quietly stopped working is worse than one that never launched, and this is the mechanism that stops that happening silently.
Changing the answering model
Safe to do at any time. Save the setting, the Worker is redeployed with the new configuration, and the next message uses it. Your index is unaffected.
Worth trying a different one if answers are too terse, too verbose, or keep missing the point of questions your content clearly covers. Bigger models are generally better at following the "decline when you do not know" instruction, which is the behaviour that matters most.
Changing the embedding model
This one requires a rebuild
The embedding model and the index are welded together by dimension. Changing the model without rebuilding leaves you comparing new vectors against old ones, which produces retrieval that is quietly wrong rather than obviously broken. Always follow an embedding change with Rebuild index.
The default index is created with 768 dimensions to match the default embedding model. A model with a different output size needs a fresh index of the matching size, which is exactly what the rebuild does.
If a model gets retired
- You will see an admin notice from the daily health check, or answers will start failing.
- Check Cloudflare's Workers AI model catalogue for a current replacement.
- Put its identifier in the answering model field and save. The Worker redeploys.
- Ask the bot a question to confirm it is answering again, or use Run diagnostics.
- If it was the embedding model that was retired, also run Rebuild index.
What this costs
Inference runs on your Cloudflare account against its Workers AI allocation, so a larger model consumes that allocation faster. See limits and costs.