Embeddings and semantic similarity

Ordinary search matches words. Search for "cancel my subscription" and a help article titled "How to end your plan" may never show up, because no words match. Embeddings let a computer compare meaning instead.

An embedding is a long list of numbers (often hundreds or a few thousand) that a special embedding model produces for a piece of text. The numbers themselves mean nothing to a human. What matters is that texts with similar meaning get similar lists, and unrelated texts get very different ones. You can picture each text as a point on a map where "cancel my subscription" and "end your plan" land close together, and "banana bread recipe" lands far away.

To compare two embeddings you compute a similarity score, most often cosine similarity: roughly 1 means very similar, around 0 means unrelated.

What people use them for:

• Semantic search: embed every document once, embed the question, return the closest documents.
• Grouping and de-duplication: find support tickets that describe the same problem in different words.
• Recommendations: "articles similar to this one".
• Classification: compare a text with example texts for each category.

Practical points:

• Embedding models are separate, cheaper models (for example OpenAI's text-embedding models or Google's Gemini embedding models; open models exist too). They have their own token limits.
• Only compare embeddings made by the same model; numbers from different models are not compatible. Changing the model means re-embedding everything.
• Embeddings capture topic and meaning well, but can miss exact details like a specific product code; combining them with keyword search often works best.