Reference · Glossary
Embedding
Last updated
An **embedding** is a list of numbers that captures the meaning of text (or other media) so computers can compare how similar two ideas are — not just whether they share keywords.
#Intuition
Map sentences into a high-dimensional space where “refund policy” sits near “money-back guarantee,” even if the words differ. Distance metrics like cosine similarity then power search and clustering.
#When to use
- Semantic search over docs or tickets
- RAG retrieval before generation
- Recommendations and near-duplicate detection
- Routing support questions to the right queue
#When not to
Exact string matching — looking up an order ID or error code is usually faster with keyword or database lookup than vectors. Also avoid treating embeddings as a fact store; they encode similarity, not truth.
#Practical tips
- Prefer the same embedding model at index time and query time.
- Chunk size and overlap strongly affect retrieval quality.
- Hybrid search (keyword + vector) often beats either alone on IDs and rare proper nouns.
- Re-embed when you change models; old vectors are not compatible across every provider.
#Example
vec = embed("laptop won't turn on")#Learn next
- Lesson: `what-are-embeddings`
- How-to: Pick an embedding model
- Related: RAG, vector database