Embeddings
An embedding turns text into a numeric vector that represents its meaning. Capacitor uses embeddings for memories, judge-fact clustering, and Work Items semantic matching on plans that include it.
Open AI & models → Embeddings (/admin/settings/ai/embeddings). These features share one active provider, configured independently from the chat and judge credentials on AI Providers.
Configure a provider
Section titled “Configure a provider”- Choose Provider: Voyage, OpenAI, or Kurrent AI.
- Fill in the selected provider’s credential and model fields. Kurrent AI is deployment-managed and needs no key from you.
- Adjust Batch size if needed.
- Click Save. If the embedding identity changes, review the reset confirmation before choosing Save and switch.
Only the selected provider’s configuration section is shown.
| Configuration key | Range or default | Meaning |
|---|---|---|
Embedding:Provider | voyage, openai, kurrentai | One active provider. Runtime default is Voyage; a deployment can preconfigure another. |
Embedding:Voyage:ApiKey | Secret | Voyage credential. |
Embedding:Voyage:Model | Runtime default voyage-4-lite | Voyage embedding model. |
Embedding:OpenAI:ApiKey | Secret | OpenAI embedding credential. |
Embedding:OpenAI:Model | Runtime default text-embedding-3-small | OpenAI embedding model. |
Embedding:OpenAI:Dimensions | 0–3072 | Blank or 0 uses the model’s native dimensions. |
Embedding:BatchSize | 1–256; default 64 | Groups inputs into fewer requests. |
Secret values are encrypted. Leave an existing secret blank to retain it, or select Clear on save to remove it. Kurrent AI’s embedding credentials, endpoint, and model are managed by Kurrent and are available on Free as well as paid plans when configured on the deployment.
Changing provider, model, or dimensions
Section titled “Changing provider, model, or dimensions”Changing the active embedding identity invalidates stored vectors. The confirmation lists affected fields and the stored embedding count. Saving discards those embeddings, and retained judge facts are re-embedded in the background under the new identity without a restart. Clustering results rebuild as that work completes; provider usage can incur new charges.
A credential-only change does not change the embedding identity.
Enable judge-fact deduplication
Section titled “Enable judge-fact deduplication”An embedding provider alone does not turn deduplication on. Open Evals, enable Judge-fact deduplication, and save. Enabling it backfills findings retained while deduplication was off.