Embedding Models

An embedding model gives your search a sense of meaning: without one, search matches words; with one, it matches meaning.

Each record becomes a list of numbers — a numeric fingerprint of its meaning — so similar meanings sit near each other.
“Running shoes” then brings up “trainers”, and “laptop for travel” finds a “lightweight ultrabook” — matches with no words in common.
You bring the model, with your own API key.

What an embedding model does

An embedding model gives your search meaning. It turns each record into a list of numbers that captures what it is about, so a search for “running shoes” can also find “trainers”. Without one, search matches words; with one, it matches meaning. They live under Models → Embedding Models.

Add one, bring your own key

Click Add Embedding Model and pick a provider — OpenAI, Gemini, Voyage or Microsoft Foundry. You bring your own API key, so embedding runs on your account: your models, your limits, your bill.

Pick a model, attach it to a list

Choose the model and paste your key — checked with a live call before it is saved. Attach the model when you create a list, or add it to a group later: records re-embed in the background with no search downtime while the change settles in.

Auto-plays · use Back / Next to step through at your own pace.

Add an embedding model

Go to Models → Embedding Models and click Add Embedding Model.
Pick a provider: OpenAI, Gemini, Voyage or Microsoft Foundry. Embedding runs on your own key: your models, your limits, your bill.
Choose the model and paste your provider API key. The key is checked with a live call before it is saved.
Click Add. The model can now be attached to lists and groups.

The Add Open AI Model dialog with a model name and API key

Attach it: what meaning-based search unlocks

Attach a model when you create a list, or add one to a group later. Once records have fingerprints:

Meaning-based matching: “warm winter coat” also finds a “padded parka”.
Related results: “more like this”, from the records closest in meaning.
Image and multimodal search: with a model that reads images too, search by picture as well as words.

Changing the model happens in the background

Every record needs a new fingerprint — minutes on a large list — so it runs as a background job on a list and a group alike. Change List Model saves and closes straight away.
Search keeps working throughout: records not yet re-fingerprinted are still found on their text.
On a group the switch is all-or-nothing — it keeps searching with the old model until every list has caught up. No half-way state.

Re-imports and monthly budget

A re-imported record whose fingerprinted text has not changed keeps its fingerprint — no new call to your model. Only new and changed records are sent.
Monthly budget… on a model’s ⋯ menu caps the tokens it embeds for imports in a calendar month (UTC), and shows roughly what it has used so far. Empty means no limit.
At 80% the account owner gets an email. At 100% a second email, and new or changed records are saved without a fingerprint until next month — still found on their text. Raise the budget, then rebuild (below).
Usage is estimated at about four characters per token; your provider’s bill is the exact figure.

Records with no vector

Causes: the provider refused, rate-limited or could not be reached; the budget ran out; the list’s model or its credential was deleted or unreadable; or a model change’s rebuild could not start. Account health shows the list and the reason.
Fix the cause first, then rebuild: Options → Rebuild vectors… on the list, Rebuild after fixing it on the Account health row, or POST list/{account}/{list}/rebuild-vectors (MCP list_rebuild_vectors).
A rebuild re-embeds every record with the same model and key. It is queued; the list’s records_without_vectors updates a few minutes after it lands.
GET list/{account}/{list} returns records_without_vectors (null until counted) and vectors_incomplete_reason (null when nothing failed).