Technical Prerequisites
Before going further, ensure your technical team completed the pre-requisites.
⚠️ For now, the vector search has only been tested with french and english languages. It might not work properly with other languages (english, spanish, german etc…), and might even need to use language-specific models for some languages, rather than relying on the pre-trained model shipped with Opensearch.
Installing the module
The module responsible for neural search is smile/module-elasticsuite-neural-search and this module is not part of the standard Premium package. You’ll need to install it manually prior going further.
Enabling neural search feature
You can head to Stores > Configuration > ElasticSuite > Neural Search
| Parameter | Default value | Description |
|---|---|---|
| Enable Neural Search | No | Set this to “Yes” to enable the neural search feature |
| Model Type | Pre-trained | ⚠️ For now, only Pre-Trained models are supported |
| Pretrained Model | A list of pre-trained models available in Opensearch. The default one is recommended. | |
| Compute embeddings from a specific store | No | When using a LLM capable of dealing with multiple languages, you can choose to compute the embeddings from a particular store view and use them for all the other languages. That’s particularly valuable if you have a store view which is greatly contribued. It can then be used to leverage the data of the other stores. |
Enabling attributes for Neural Search
Enable the “Use for Vector Search” option on the attributes you plan to use with Neural Search. Refer to the attribute configuration page to do so.
Proceed with reindex
When you’ll have configured your attributes for Vector Search, and enabled the feature, you’ll be able to reindex the embeddings data.
This is done by rebuilding first the elasticsuite_products_embeddings index which will compute the embeddings, and then the catalogsearch_fulltext which will copy the embeddings into the fulltext index.
Doing so, you don’t need to rebuild the embeddings index as often as your search index, because embeddings are more costly to rebuild.
Configuring neural search feature
You can head to ElasticSuite > Search Relevance > Neural Search
| Parameter | Default value | Description |
|---|---|---|
| Use vector search as a fallback when there is no search result. | No | We recommend to set this to “Yes” to use the neural search as a fallback |
| Use vector search in spellchecked | No | When set to yes, the neural search will be used as part of the spellchecking. We don’t recommend to use this. |
| Use vector search in exact matches | No | When set to yes, the neural search will be used as part of the exact matching. We really don’t recommend to use this. |
| K Value | 100 | We recommend to let this value to the default one. |
| Use Min Score | No | We don’t recommend enabling this |