RETRIEVAL_QUERY

RETRIEVAL_QUERY is an embedding task type that optimizes vectors for representing search queries — short, intent-driven user phrases. It's used on the query side in RAG systems: when a user enters a question, their text gets vectorized using this task type.

It works in tandem with RETRIEVAL_DOCUMENT (on the document side), forming an asymmetric retrieval system. Separating these two types lets the model better match a short, intent-driven question to a long, informational document — without this separation, the model struggles to match short questions with long documents effectively. For SEO professionals, this enables building semantic search engines and content-based chatbots for websites.

For example, when building a chatbot for law firms, use RETRIEVAL_QUERY to vectorize client questions and RETRIEVAL_DOCUMENT to index the firm's blog articles.