Glossary

A dense vector representation of an image produced by a vision model, typically a vision transformer or CNN. Modern image embeddings capture both low-level visual features, such as colors and textures, and high-level semantic content, such as objects and scenes. When produced by jointly trained multimodal models like CLIP, image and text embeddings share a common vector space, enabling similarity search across both modalities.

A training efficiency technique where positive pairs from other examples in the same batch are reused as negatives. If a batch contains 64 query-passage pairs, each query treats the other 63 passages as negatives. This dramatically increases the number of comparisons without requiring additional data.

Data or evaluation scenarios that closely match the distribution a model was trained on. In-domain performance measures how well a model has learned its training task but is a weak test of generalization, as inputs in production rarely match training data precisely.

Adding content to a search system so it can be found later. For semantic search, this means generating embeddings for all documents and storing them in a vector index. Indexing is done ahead of time so searches can be answered quickly.

Running a trained model on new inputs to produce outputs, as opposed to training the model. When you send text to an embedding API and get back a vector, the model is performing inference. Inference speed and cost are the main considerations for production deployment.

A contrastive loss function that frames the task as classification: given an anchor, identify the correct positive from a set of candidates (one positive and many negatives). It is the foundation of training approaches used in CLIP and many modern embedding models.

The practical constraint on how much text you can send to a model in a single request. This is determined by the model's context window and may be further limited by API settings. Exceeding it typically results in truncation or an error.

A training approach in which the model learns to follow natural-language instructions describing the desired task. For embedding models, this means the model can adjust its behavior based on a text prefix — generating different embeddings for the same text depending on whether the instruction says "retrieve a relevant passage" or "classify the topic."

A data structure used in lexical search that maps each word to the list of documents containing it. When a query arrives, the inverted index quickly identifies all documents with the query terms. This is the backbone of traditional search engines and is used alongside vector indexes in hybrid search.