Embeddings
Dense vector representation of text that captures semantic meaning, used for similarity search.
Dense vector representation of text that captures semantic meaning, used for similarity search.
Définition détaillée
Embeddings are numerical representations of textual data as vectors of real numbers (typically 384 to 3072 dimensions). Specialized models (text-embedding-3, BGE, E5) transform any text into a vector encoding its semantic meaning. Semantically similar texts have close vectors in the vector space, measured by cosine similarity. Embeddings are the fundamental building block of vector databases and RAG systems.
Cas d'usage
Internal semantic search engine: employees instantly find relevant documents by meaning similarity, without exact keywords. Search works even with different phrasings.
Termes associés
En savoir plus
Questions fréquentes
What's the difference between embeddings and keywords?
Keywords search for exact matches. Embeddings search for meaning: 'car' and 'automobile' have similar vectors even without shared letters.
Which embedding model should I choose?
text-embedding-3-small for cost/quality, BGE-M3 for multilingual (including French), E5-mistral for best overall quality.