Embeddings and Retrieval
Learn how embeddings let a system find content by meaning: the ingestion and query paths, approximate nearest-neighbor search, choosing keyword, vector, or hybrid search, placing filters and rerankers, measuring retrieval with recall@k and precision@k, and migrating embedding models safely.
Embeddings are numeric representations of text that place content with similar meaning close together in a shared vector space. They let a search system match a question such as "Why did my subscription stop?" to an article titled "Renewal failures due to expired cards," even though the two share almost no words.
Similar meaning suggests relevance, not correctness. Retrieval therefore focuses on finding strong candidates, and later stages check whether a result is current, accurate, and safe to show.
In this lesson, we will explore retrieval workflows, search strategies, and ways to maintain reliable retrieval as embedding models evolve.
From query to candidates
Semantic retrieval has two paths. During ingestion, documents are split into chunks and converted into document embeddings that are stored for search. At request time, the user's query is converted into a ...