embedcache

Models

22+ local embedding models

embedcache ships the FastEmbed catalogue — over 22 ONNX models that run on CPU. Choose a model per request, or restrict the set with ENABLED_MODELS.

A representative slice

These five are the models you will reach for most often. The full list of 22+ is in the FastEmbed catalogue and is reported by GET /v1/params for your build.

Model Dimensions Notes
AllMiniLML6V2 384 Fast, general purpose
BGESmallENV15 384 Best quality/speed balance
BGEBaseENV15 768 Higher quality
BGELargeENV15 1024 Highest quality among bundled English models
MultilingualE5Base 768 100+ languages

Choosing a model

Higher-dimension models (BGE base/large) generally capture more nuance at the cost of larger vectors and slower inference. BGESmallENV15 is a solid default for the quality/speed balance; AllMiniLML6V2 is the fast option.

Multilingual

MultilingualE5Base covers 100+ languages in 768 dimensions. Reach for it when your corpus is not English-only.

Because vectors are keyed by content and model id, mixing models is safe — each model's output is cached separately. See how it works.