Instruction-Tuned Query Prefixes Lift Vector Recall 5-15%, and Price Badges Show Currency Symbols

An accuracy change under the hood, and a readability change on the surface: vector search now embeds queries with instruction-tuned prefixes, and price badges show a symbol rather than a currency code.

01 Instruction-tuned query prefixes

The embedding model now applies instruction-tuned prefixes when it builds a query vector, so the vector is optimised for retrieval rather than being a generic encoding of the text.

• 5–15% better recall. Measured on natural language queries.
• Stronger cross-language matching. Searching in Romanian, for example, now finds English results more reliably.

Nothing needs to be configured. The change applies to hybrid search as it already runs.

02 Currency symbols on price badges

Price badges in search results displayed the raw currency code. They now show the symbol instead — €, $, £, ¥ and the rest — which is quicker to read in a result list.

Prices also now display correctly for every index, regardless of how its schema is configured.