Embedding
vector embedding · vector representation
Kurz gesagt
An embedding is a fixed-length vector of numbers produced by a model to represent a piece of text, image or audio. Items with similar meaning end up near each other in that vector space, which is what makes semantic search, clustering and recommendation possible.
An embedding model reads text and emits a vector — commonly several hundred to a few thousand numbers. Distance in that space approximates similarity in meaning, so "cancel my subscription" lands near "how do I end my plan" despite sharing almost no words.
That property is the engine behind semantic search, deduplication, clustering, classification by nearest neighbour, and the retrieval half of RAG.
Four practical points:
- Vector spaces are model-specific. Embeddings from different models are not comparable, and changing model means re-embedding everything.
- Chunk size shapes quality. Chunks that are too large dilute the signal; chunks that are too small lose the surrounding context. Splitting on document structure usually beats a fixed character count.
- Semantic similarity is not relevance. Two passages can be about the same topic while only one answers the question. Reranking with a model that scores query-passage pairs directly fixes much of this.
- Exact matching needs keywords. Names, codes and identifiers are where vector search is weakest and keyword search is strongest.
Embedding calls are individually cheap and collectively significant: re-indexing a large corpus after a chunking change is a real line item.
Häufige Fragen
- Can I compare embeddings from two different models?
- No. Each model defines its own vector space. Switching embedding models means re-embedding the entire corpus.
- Why does vector search miss exact terms like invoice numbers?
- An identifier carries little semantic meaning, so its embedding is not distinctive. Keyword search finds it immediately, which is why hybrid retrieval combining both is the practical default.