An embedding model transforms unstructured input into a fixed-length vector in a high-dimensional space. Semantically similar items cluster together; nearest-neighbor search becomes the foundation for RAG, semantic search, recommendations, classification, and clustering.
Batch embedding pipeline with deduplication and quality gates.
MTEB leaderboard ranks models by retrieval, classification, clustering. BGE-large and E5-mistral lead open-source; text-embedding-3-large leads closed.
1536-dim vectors give the best recall but cost 2x storage vs 768-dim. Matryoshka embeddings let you truncate to any size with minimal quality loss.
Use models trained on multilingual data (E5-multilingual, BGE-M3). English-only models will cluster all non-English text together.
For L2-normalized vectors, cosine sim = dot product. Use dot product (faster). Never mix metrics across an index.
E5/BGE require prefixes: 'query:' for queries, 'passage:' for documents. Skipping prefixes silently degrades retrieval by 5-15%.
A model topping MTEB may underperform on your data. Always run a domain-specific eval before committing to one for re-indexing.
Sign in to share your feedback and join the discussion.