Embeddings

An embedding turns a piece of text — a judge finding, a memory, a work-item title — into a numeric vector that captures its meaning, so Capacitor can tell when two things are about the same topic even when the wording differs. That’s what lets near-duplicate judge findings collapse into a single cluster on the Facts tab instead of piling up, and what powers semantic matching for memories and Work Items correlation.

Settings → Embeddings (/admin/settings/ai/embeddings) is where you pick the provider that produces those vectors and enter its credentials. All three features that rely on them — session memories, eval judge-fact deduplication, and Work Items correlation — use this one provider: one set of credentials, one place to change them.

Embedding credentials used to live on AI Providers and the provider selector on Evals; both now live here.

Provider and credentials

Embedding:Provider selects the active provider; the matching credential block below must be filled in for the features that use embeddings to work.

Key	Type / range	Default	Notes
Embedding:Provider	enum voyage / openai / kurrentai	voyage	kurrentai only on plans that allow it.
Embedding:Voyage:ApiKey	secret	—	Required when judge-fact dedup is enabled and Voyage is active. Stored encrypted at rest.
Embedding:Voyage:Model	string	voyage-4-lite (runtime)	Form renders blank; the runtime default applies.
Embedding:OpenAI:ApiKey	secret	—	Stored encrypted at rest.
Embedding:OpenAI:Model	string	text-embedding-3-small (runtime)	Form renders blank; the runtime default applies (EmbeddingOptions.DefaultOpenAIModel).
Embedding:OpenAI:Dimensions	int 0–3072	unset = model native	0 or blank uses the provider’s native dimension count.
Embedding:BatchSize	int 1–256	64	Groups inputs into fewer provider requests.

Kurrent AI, on plans that allow it, uses deployment-managed credentials — the API token, endpoint, and gateway are set by Kurrent at deployment and aren’t admin-editable.

Saved changes take effect immediately; no restart is required.

Caution

Changing identity fields — the provider, its model, or (for OpenAI) the dimension count — discards all stored embeddings. The backfill service re-embeds retained judge facts under the new identity in the background. Until it finishes, judge-fact clustering has nothing to show.

Related

• Evals — enables the judge-fact deduplication that consumes these embeddings, and sets the server-runner run limits.
• AI providers — the judge-vendor credentials the eval runner uses.
• Work items — the correlation agent, which also uses this embedding pipeline.