What Are Embeddings? AI's Secret Language Explained Simply (2026)

An embedding represents an input as a vector, or list of numbers, for tasks such as finding related text. A retrieval system compares compatible vectors to identify candidates. A close match is evidence of similarity under that model, not proof that a passage is true or answers the question.

What vector similarity can and cannot tell you

An embedding model can place related phrases near one another even when they use different words. That can help connect a question about “reimbursement” with a policy discussing “repayment of expenses.”

It can also retrieve an irrelevant passage with similar vocabulary. A policy saying an expense is allowed and a policy saying it is prohibited may be close in topic. Dates, exceptions, identifiers and access permissions require additional handling.

Search does not always require embeddings. Keyword search, exact identifiers, metadata filters and structured queries can be the right tools. A hybrid system can combine lexical and vector retrieval rather than making one universal choice.

A worked retrieval test

Consider these fictional records:

For “What receipt do I need for a meal claim?”, an ordinary employee should receive record A. A result from B is outdated even if it has a strong similarity score. Returning C to that employee is an access-control failure even if it is highly relevant.

Add an exact-reference test such as “Find expense policy EXP-204.” Compare keyword or identifier lookup with semantic retrieval. Include a question that no record answers, so the system must report missing evidence rather than return the nearest unrelated passage.

The test defines expected behavior before a vendor or model is selected. No vector coordinates or scores in this example are measured model output.

Build the retrieval path

First identify the source records and their owners. Store a stable document identifier, version, date, source location and access metadata alongside each searchable passage.

Then divide documents into passages that preserve useful context. An isolated sentence can lose an exception or heading. A passage that is too broad can mix unrelated topics. Compare candidate chunking methods on your own questions instead of assuming one fixed size works for every document.

Encode passages and queries using the model's intended retrieval methods. Sentence Transformers' semantic-search documentation distinguishes symmetric and asymmetric tasks and recommends the appropriate query and document encoding methods. Follow the specific model's configuration. Arbitrarily mixing vectors from unrelated models does not create a compatible search space.

Apply access and version constraints as part of retrieval, before exposing results to a user or answer model. Keep the source passage available for inspection. A generated answer should identify its supporting records and leave unsupported questions unresolved.

Compare models and retrieval settings

Use a labelled set of representative queries with acceptable sources. Record which relevant passages appear, which irrelevant ones appear, whether the user can access them and whether the correct version is returned.

Compare quality, latency, storage, operating cost and deployment requirements on the same test. Model dimensions and input limits differ. Read current model documentation rather than relying on one price table or a claim that all models truncate at the same length.

A reranker can change the order of candidate passages, but it does not establish truth or authorization. Test the final result rather than treating an extra retrieval stage as an automatic improvement. The Sentence Transformers retrieve-and-rerank example explains that pattern.

Plan for changes and failures

When source material changes, update or remove the corresponding indexed passages. Record the embedding model and configuration used. Changing the representation can require rebuilding the affected index and rerunning the query tests.

Inspect failures by category: missing source, extraction error, bad passage boundary, retrieval miss, wrong version, permission error or unsupported generated answer. Those require different repairs. Increasing the number of returned passages does not resolve every category.

Use the after-action review worksheet to record a failed query, its cause, the proposed change and the test that will confirm the repair.