UpStart Commerce Search
Semantic (Neural) Search

Setting Up Semantic (Neural) Search with PIM

9min

Prerequisites

Before enabling neural search, ensure you have:

Step 1: Create Required Attributes

Create Neural Description Attribute:

curl


Create Neural Name Attribute

Similar configuration as above, but with "name" as the attribute name.

Step 2: Commit Attributes

Curl


Step 3: Create Product Template (screenshots will be added in the published doc)

1. Upon logging in to the UpStart Commerce portal, navigate to PIM on the left navigation and select Product Templates.

Document image


2. Click on the +Create Template button and enter the template name and description.

Document image


3. Scroll down in the Create Product Template pop-up and select the Product Type.

Document image


4. Once all the details are added, click on Create & Open.

5. From the Attributes list, search for Neural/Semantic attributes and select all that are applicable.

Document image


6. Once selected, click on the Add button to add the Neural Search attribute to the template.

Document image


7. Once added, the attribute will be shown under the Added Attributes section.

Document image


8. Click on the Publish button to publish the product template.

For a detailed guide on how to create a product template, go to Product Templates. a

Step 4: Update Catalog Route

The catalog route needs to be updated to optimize the retrieval of KNN vectors. This step is crucial because it prevents the system from retrieving KNN vectors on every search request, which would be inefficient.

Curl


Important Notes about KNN Vectors:

  1. The excludes field in the payload specifically excludes the KNN vector fields (neural_description_semantic_catalog_model_embedding and name_semantic_catalog_model_embedding).
  2. This exclusion is important because:
    • KNN vectors are large data structures that aren't needed in the search response
    • Retrieving them with every search would increase response time and bandwidth usage
    • The vectors are only needed for the internal neural search computation
  3. The vector embeddings are still used in the search process, but they're not returned in the response payload
  4. Once the ML model and index are set up, these vectors will be automatically generated and stored for:
    • Every new product created
    • Any existing product that gets updated
    • The fields you've designated for neural search (name and description in this case)

Step 5: Final Setup

Contact the UpStart Commerce team to reindex your catalog indexes to complete the Neural Search setup.