Embedding — AI Glossary

Reference details

Related terms

• Vector Database
• Semantic Search
• RAG (Retrieval-Augmented Generation)
• Reranker (Cross-Encoder)

How it works

1. The text is tokenised and run through an encoder, producing one vector per token, which is not yet an embedding of the passage.
2. Those token vectors are pooled into a single fixed-length vector, usually by averaging them or by taking the representation at a designated position, and that pooling choice is part of the model.
3. The result is commonly normalised to unit length, which makes the dot product equal the cosine and is why the two similarity measures are interchangeable for most models.
4. The reason similar texts land close together is the training objective: pairs known to match are pulled together and mismatched pairs pushed apart, so proximity is a learned property rather than a geometric accident.
5. Those negatives are usually other items in the same training batch, which is why embedding models are trained at large batch sizes — the batch is the difficulty of the task.
6. Some models are asymmetric and expect a prefix marking a text as a query or as a document, and omitting it places the two sides of a comparison in regions the model was trained to keep apart.

Example: Why cosine similarity, not keyword overlap — and what it costs on identifiers

“How do I reset my password?” and “I forgot my login credentials” share almost no words — keyword search scores them near zero. Their embeddings sit close together, so cosine similarity is high and retrieval finds the right document. The same property is the failure mode: “cancel my subscription” and “renew my subscription” are also close, because they are about the same topic while meaning opposite things. The identifier case is where that costs measurable recall. Suppose 12% of queries contain a SKU, an error code or an exact product name, and the embedding path retrieves the right document for 40% of them while a lexical index manages 95%. The gap is 0.12 x (0.95 - 0.40) = 6.6% of all traffic that a hybrid retriever answers and a dense-only one does not. That figure is what decides whether a second index is worth running — and it is a property of your query mix rather than of the embedding model.

Commonly confused with

An embedding is not a summary and cannot be read back. It is a fixed-length vector, and nothing in it is human-inspectable. Similarity in embedding space means topical relatedness, which is not the same as relevance and definitely not the same as agreement. It is also not a fixed representation of meaning: two models place the same sentence in unrelated coordinate systems, so a vector is only interpretable against the model that produced it and an index cannot mix them.

When to use it

Reach for it when:

• Users phrase the same intent in many different ways
• You need clustering, deduplication or classification over free text
• Recall matters more than exact term matching
• Where the identifier share of queries is small enough that 6.6% of traffic is an acceptable loss

Reach for something else when:

• The query is an identifier — a SKU, error code or name — where lexical search wins outright
• Negation or opposites must be distinguished; embeddings are weak at this
• You need to explain to a user or auditor why a document was returned
• Mixing vectors from two models in one index, which places the same text in unrelated spaces

Referenced by

This term is used by the following ModelRefs references:

• rag grounded answer
• hyde
• long term memory rag
• vector search tool
• naive rag stack
• advanced rag stack
• learn rag
• become ai engineer
• build enterprise search
• build coding copilots
• openai builder path

Where this appears on ModelRefs

• BGE-M3

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Embedding — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path