Embedding

Embedding sets the one model that turns your content into vectors, so search can find things by meaning. There is one embedding setup for the whole brain.

• Model: the embedding model name.
• Primary route: the provider, base URL and API key. Click Test dimensions to check the model returns 768 numbers per vector.
• Backup route (same model): turn on Enable failover to reach the same model another way.
• Performance & throughput: hardware profile, concurrency and batch size for indexing.
• Click Save embedding config.

Do not change the model without a reason. All vectors must come from the same model to be comparable. After a change, click Rebuild index to re-embed everything. Search is weaker until that finishes.

Assistant

The assistant cannot change these settings. Set them here.

Technical

• Every vector is stored at 768 dimensions. A model with another size cannot be used without a database change.
• The backup must be the same model, or its vectors would not match the index.
• Embeddings are cached by model and text, so unchanged text is never embedded twice. A new model invalidates the whole cache.
• With no setup, a local keyless model is used.
• See Models and API keys and Local models.