Product recommendation limitations

This guide explains the known restrictions and trade-offs you may encounter when using product recommendations in Bloomreach.

Restrictions

The following limitations apply to product recommendations:

1. A metric-based engine returns a maximum of 500 products or items in a single request.
2. A metric-based category engine returns a maximum of 500 products or items in a single request with one category in the categoryNames attribute.
3. Catalog filter supports neither regex filters nor list attributes and filters.
4. Catalog properties longer than 256 UTF-8 bytes aren't fully indexed. Only the first 256 UTF-8 bytes of a catalog property are indexed and searchable. Values exceeding this limit won't match the contains filter for the portion beyond 256 bytes. To resolve this, shorten the property value or split it into multiple properties.
5. Catalog filter works as a post-filter. This means the recommendation model may return 10 suitable items, but a strict catalog filter can reduce that number further.
6. In the Filter-based template, the dynamic catalog filter from the get-recommendation request is ignored.
7. In the Advanced template, request latency increases with each recommendation engine added to a single advanced template instance. Factor this in when designing a recommendation engine.
8. In all AI models that use events for training, a minimum threshold of 2 events per customer is required. For example, a customer with 1 purchase_item event in the last 90 days won't be considered for training. A customer with 1 purchase_item and 1 view_item will be considered. This threshold reduces model skewness.
9. AI models support a maximum of 1,000 products in the size parameter.
10. Textual similarity trains only on models with fewer than 1,000,000 rows.
11. Combining multiple recommendation engines inside the advanced engine may increase response time.

Challenges and trade-offs

You may encounter the following challenges when developing and using recommendations:

• Exploitation vs. exploration: The system needs to balance recommending items a customer has already interacted with (exploitation) and surfacing new items they haven't seen yet (exploration).
• Popularity bias: Most customers view very few items, and most items are seen by only a small number of customers. This creates significant data gaps that the recommendation model must account for.
• Cold-start problem: When you first start using Bloomreach, there's usually not enough historical customer-item interaction data for personalized recommendations to perform well. Recommendation quality improves as more data is collected. The same applies when a new product is added to the catalog.