Problem
Native Instruments makes over a thousand virtual instruments, textures and effects. Many are hard to describe in words. The old metadata held only product name, description and category, so a query that describes a sound found nothing unless it matched those fields. About 20% of searches returned no results. Manual tagging could not cover the catalog or capture what each product sounds like.
Example from the article: the instrument Electric Mint emulates a 1960s solid-body electric guitar. A producer who hears a funky riff types "funk", "guitar" or "surf rock", never "Electric Mint", and gets nothing.
Solution in brief
An LLM writes a list of search keywords for every product record, directly inside the data warehouse. The keywords go into a new column, and the warehouse connector syncs that column to the search index as one more searchable field. Queries then match genres, artists and related products that the old metadata never mentioned.
Questions this case answers
- How to make catalog items retrievable by the words users actually type when item metadata cannot describe them and manual tagging does not scale
- How to make products findable by search when their names and descriptions do not match how users describe what they want
- How to cut the share of searches that return nothing without tagging a large catalog by hand
- How to add LLM-generated search keywords to a product catalog that lives in a data warehouse
Tools and methods
- BigQuery — warehouse that holds web shop, CMS and file data and runs the generation step
- dbt — SQL models that clean, rename and normalize product data into a staging layer
- Vertex AI / Gemini — model called from BigQuery through
ML.GENERATE_TEXT()to write keywords - Algolia BigQuery Connector — publishes the enriched table to the search index
- Prompted keyword generation — one fixed prompt per product record, output stored as a column
Google or a one-time email code. No password.