What are embeddings?

Embeddings are lists of numbers that represent the meaning of a piece of text, an image or another item, produced so that items with similar meaning end up with similar numbers. They let a system find related content by meaning rather than exact wording. AI agents use them mainly to retrieve the documents and records most relevant to a task.

Why do embeddings matter once agents act?

Because they decide what the agent gets to read. When an agent asks "what is our policy on refunds for annual plans?", an embedding search returns the passages closest in meaning. If it returns a superseded version, or a forum post someone saved into the same store, the agent works from the wrong source. The quality of its action rests on a similarity score it never sees.

Similarity is not authority. Two passages can be close in meaning while one is the approved policy and the other is a draft that was never approved.

How do embeddings work in practice?

A model converts each chunk of text into a vector: a long list of numbers. The vectors are stored, usually in a vector database. When a query arrives, it is converted the same way, and the system returns the stored chunks whose vectors are nearest. Those chunks go into the agent's context.

Good practice keeps the metadata with every vector: which source the chunk came from, which version, who owns it, its provenance class and who may see it. Search results can then be filtered before the agent sees them, so it retrieves only from sources its brief allows. That filtering should happen in one central place, for the reasons set out in whether agents should talk to each other directly or through a central layer.

What does that look like in an example?

A pension administrator's helpdesk agent receives a question about transferring out. The embedding search finds five close passages: the current transfer procedure, a superseded one, a training note, an internal email and a member's old complaint. Without metadata, all five look equally relevant. With it, the system keeps the current procedure, which is owned and approved, drops the superseded version, and marks the training note as background. The agent answers from the right document.

How are embeddings different from keyword search?

Keyword search matches words: a query for "transfer out" finds documents containing those words. Embedding search matches meaning: it also finds "moving your pension to another provider". That makes it better at finding what people mean, and worse at guaranteeing that an exact term appears. Many systems use both and combine the results.

What can embeddings not do?

They cannot prove identity or origin. A near-perfect match between a supplier's message and a known supplier's past messages says the text is similar; it says nothing about who sent it. That needs authentication, covered in how AI agents authenticate to other systems. Embeddings also do not make retrieval repeatable to the letter: re-indexing can change which chunk comes first, which matters if an action is retried. See why agent actions must be idempotent.