Neural Search enhances your catalog search by using advanced machine learning models (Large Language Models) to understand the semantic meaning and context of a user’s query, rather than just relying on exact keyword matches. The settings below allow you to configure how the engine generates and manages the vector embeddings required for this semantic search capability.

General Configuration

Parameter	Default value	Description
Enable Neural Search	Yes	Controls whether the Neural Search capabilities are active for your store.
Model Type	Pre-Trained	Defines the hosting location and type of the language model used to generate your catalog’s embeddings. Available options are Pre-Trained, Local, or Remote.
Pretrained Model	huggingface/sentence-transformers/all-distilroberta-v1	Specifies the exact pre-trained model used by the engine to compute vector embeddings.
Compute embeddings from a specific store	Yes	Yes: Allows you to compute embeddings from one primary store view and apply them across all other language store views. This is highly recommended if you are using a multi-lingual LLM and have one particular store view with exceptionally rich, high-quality content. It leverages your best data to power search across all stores. No: Embeddings are computed independently for each individual store view.
Source Store for embeddings	Default Store View	(Only applies if the setting above is set to Yes). Select the specific store view to act as your primary data source for computing embeddings. You should select the store view that has the most comprehensive and highest-quality catalog data.