Embedding
also called Vector
A list of numbers representing meaning, so similar things land close together.
A dense numeric vector produced by a model to represent a piece of text, an image or a record. Semantically similar inputs produce vectors that are close in space, which makes similarity search possible.
Embeddings are how you search by meaning rather than by keyword. "How do I cancel?" and "end my subscription" share almost no words, but their vectors are neighbours.
Embeddings are a compression of meaning into geometry, and the geometry is what you actually rely on. That means two things to check before trusting a retrieval system: whether the embedding model was trained on text like yours, and whether its distance metric matches how the index searches. Mismatches produce confidently irrelevant results.
The model that did not know your jargon
Semantic search works on general queries and fails on internal acronyms.
A general-purpose embedding model has never seen your product names, so they have no meaningful position in its space. Fixes, in order of cost: enrich chunks with definitions, add keyword search alongside the vectors, or fine-tune an embedding model on your query-document pairs.
Assuming embeddings capture facts. They capture similarity of meaning, not truth.
- docsGet text embeddings ↗
Embeddings over text with semantic search, classification and clustering as the canonical uses — plus the vector dimensionality that decides your index cost.
- paperDense passage retrieval for open-domain question answering ↗
Dual-encoder dense retrieval trained on questions and passages, and the comparison to BM25 that started the dense-versus-sparse debate.
Vector Database · Vector store
Storage built for finding the nearest vectors, fast.
Semantic Search
Search by meaning instead of matching words.
RAG · Retrieval-Augmented Generation
Look things up, then put what you found in the prompt.
Multimodal
A model that handles more than text — images, audio, video, screen pixels.