Embeddings

Embeddings turn titles, abstracts, and other text into vectors that semantic search can compare.

Embeddings tab showing scope, text source, skip-existing option, Run Embeddings button, and stored vector summary.
Embeddings are generated by scope and text source, then stored for semantic search and related workflows.

Before you run

Configure an Embeddings model in Settings: Runtime and model presets. Embedding models use an embeddings endpoint, not the chat model route. If no embedding model is configured, Semantic Search will not have vectors to compare.

Controls

Scope

Choose which project records to embed, such as Pending or Included. Use smaller scopes for testing and larger scopes when the settings are confirmed.

Text source

Choose the text that becomes the embedding input. Title and abstract is the usual first choice for screening-stage search.

Skip rows with existing vectors

Leave this on when you only want missing vectors. Turn it off when you intentionally want to regenerate vectors after changing model or text source.

Run Embeddings

Queues an embeddings job. Large scopes can take time, so progress is shown in Jobs.

Refresh

Reloads the stored vector summary after jobs finish or after another workflow adds records.

Stored vectors

Shows what embedding sets already exist for the project and how many rows are covered.

What happens next

When the job finishes, Semantic Search can score records against a query. If records are added later, run embeddings again for the missing rows. If you change embedding models, regenerate vectors for the relevant source so scores are comparable.

If Semantic Search says no embeddings are available, check Settings first, then Jobs for a failed or still-running embeddings job.