New UserQueryLLM for german Label Picking

Today, we’re excited to announce that we have quietly launched our new a model we have been working on for the past 12 months.

In ecommerce, product data usually follows a clear structure. Whether for SEO, GEO, PIM systems, or catalog management, products are organized, enriched, and standardized. User queries, however, are not.

This gap creates a major opportunity: aligning user queries with a structured representation can unlock significant value for ecommerce and retail players and their users.

Structured queries are easier and faster for humans to scan. This is especially visible in autosuggestions: when suggestions follow a recognizable structure, users understand them more quickly and learn the underlying query patterns over time. As a result, they can formulate, refine, and expand their searches more effectively. This led to significant uplifts across the board in user engagement.

Ecommerce engagement data is often extremely sparse, making it difficult to identify meaningful trends, behavioral shifts, and intent patterns. By structuring queries and especially by normalizing their formulation we can dramatically reduce the long tail of events and shift engagement mass toward the short head and mid-tail.

This significantly strengthens behavioral feedback loops such as learning-to-rank, embedding-model fine-tuning, and causal analytics.

We continuously strive for higher computational efficiency. Query understanding is a key part of this.

Query Understanding combined with hybrid search based on embeddings has become the de-facto standard for product discovery, but most base models are still primarily trained on structured content. This content structure often does not align with the language and structure of real user queries, which makes it harder for models to learn, adapt, and generalize.

By aligning user-query structure with product-content structure, we can substantially reduce computational effort through better caching and more reusable representations. Having more structurally aligned labels reduced our query-understanding API calls by roughly 20% whilst increasing recall over 10%.

To our knowledge, we are the first to have successfully achieved this kind of task at this scale. Therefore a big thank you to the whole team.

We’re excited to bring even more downstream capabilities to our customers, powered by our new .