Choose Models and Secrets
Quick reference
Model selectors start with DataBraid's versioned fallback catalog. After you select a provider secret, the editor refreshes models available to that credential for OpenAI, Groq, Anthropic, Gemini, ElevenLabs or Hugging Face. Refreshed lists remain private to the current DataBraid session and expire after five minutes.
Legacy models remain valid in saved Braids but are hidden from new selections. Voyage embedding models use a versioned catalog because Voyage does not expose the same account model-list contract.
Configure an AI Bead
- Select the provider model or model capability required by the Bead.
- Select a secret reference owned by your DataBraid account.
- Wait for the model selector to refresh when the provider supports discovery.
- Save the Braid so provider metadata for refreshed models travels with the document.
- Run the fixture and inspect the model preflight event before node execution.
Secret values never enter the Braid document. The server resolves the selected reference for the current owner immediately before execution.
Preflight errors
| Code | Meaning | Resolution |
|---|---|---|
unknown_model | The model is empty, unavailable in the fallback catalog or lacks authenticated refresh metadata. | Refresh the selector and choose an available model. |
model_provider_mismatch | The model belongs to a different provider than the Bead or persisted provider metadata. | Choose a model exposed by the intended provider. |
provider_secret_missing | The secret reference is empty, deleted or has no value. | Select or recreate the provider secret. |
Preflight failures happen before LiteGraph configures or executes the Braid, so they do not make a provider request.
The same contract applies to chat, vision, embeddings, OpenAI text-to-speech, OpenAI transcription and ElevenLabs text-to-speech. A model exposed for one capability cannot be selected for another capability.
Embedding models
Knowledge embedding selectors use the same catalog for OpenAI and Voyage. The configured vector index dimension must be supported by the selected model. Changing embedding families for an existing index can require rebuilding that index and re-vectorizing its documents.