What Are Embeddings? A Non-Mathematical Guide
Embeddings are the quiet workhorse behind AI search, recommendations and RAG. The idea is simpler than the name.
- Published
The idea
An embedding model reads a piece of text and outputs a list of numbers — its coordinates in a "meaning space". Texts about similar things land close together. "How do I reset my password?" and "I forgot my login" end up near each other even though they share no words.
What they power
- Semantic search: find documents by meaning, not exact words.
- RAG: pick the right passages to show a language model.
- Recommendations: "customers who read this also read".
- Clustering: group thousands of tickets or reviews into themes automatically.
- Deduplication: spot near-identical records.
Their limits
- They blur exact details: product codes, numbers and names can be matched poorly — pair them with keyword search.
- They are model-specific: vectors from different embedding models are not comparable; switching models means re-embedding.
- They carry the model’s language coverage: check quality for Hindi, Marathi or mixed-language text.
Choosing an embedding model
Test two or three candidates on your own data with a small labelled set of queries and expected results. The best model on a public leaderboard is often not the best on your documents, and a smaller, cheaper model sometimes wins.
Frequently asked questions
Are embeddings expensive?
Embedding text is cheap per document and done once per change; the ongoing cost is usually small.
Can embeddings leak my data?
Embeddings can partially reveal source text, so protect them like the documents themselves.
Do images have embeddings too?
Yes — multimodal models embed images and text in the same space, enabling search across both.
