Embeddings and Vector Search

Embeddings are numerical representations of data — a sentence, a document, an image — expressed as a long list of floating-point numbers called a vector. The clever part is that an embedding model places semantically similar things close together in that vector space: “How do I reset my password?” and “I can’t log in” end up near each other even though they share almost no words. Vector search is simply the act of finding the stored vectors closest to a query vector.

This pairing is the retrieval backbone of modern AI applications. Large language models can only reason over what fits in their context window, so you need a way to fetch the right text before the model ever sees a prompt. Keyword search misses paraphrases and synonyms; vector search matches on meaning. That is what makes RAG, semantic search, memory for agents, and deduplication actually work in production.

In practice the loop is short: call an embeddings endpoint (such as OpenAI’s embeddings API) to convert your documents into vectors, store them in a vector database or index, then embed each incoming query with the same model and run a nearest-neighbor search using a metric like cosine similarity. The top results become context for your LLM call. You’ll build this pipeline repeatedly throughout this module, so make sure the mental model — text in, vector out, distance equals relevance — is solid before moving on.