Embedding model

An embedding model is the network that does the embedding. Given an input, it outputs a fixed-length vector, and its training objective shapes what the vector captures. Models trained contrastively on image-text pairs place images near the words that describe them, while models trained on visual similarity place lookalike images together regardless of language. The same image embedded by two different models can therefore be neighbors with very different company.

Popular families include CLIP-style vision-language models, self-supervised vision backbones, and multimodal models that embed several data types into one shared space.

Most embedding models are trained with contrastive or self-supervised objectives that pull related pairs together and push unrelated pairs apart. At inference, the model runs once per input and the vector is stored. Teams often evaluate several embedding models on their own data, since the best choice depends on the domain, and fine-tuning an embedding model on in-domain data can sharpen its notion of similarity further.

Every downstream capability, semantic search, deduplication, clustering, anomaly detection, inherits the strengths and blind spots of the embedding model that produced the vectors. Choosing and validating that model against your actual data is one of the highest-leverage decisions in an embedding-based workflow.

Test candidates on your own data with tasks you care about, such as retrieval quality or cluster coherence. Domain match usually matters more than leaderboard rank.

Yes. Fine-tuning on in-domain pairs adapts what the model considers similar, which often improves retrieval in specialized domains.

Embedding, Contrastive learning, Multimodal embedding, Feature extraction, Fine-tuning.

Learn: How image embeddings transform computer vision

Try it: Qwen3-VL embeddings for FiftyOne · Nomic Embed multimodal for FiftyOne