What is an embedding?

An embedding is a numerical vector representation of content designed so useful relationships, such as semantic similarity, can be compared mathematically.

RAG retrieves relevant passages from an approved knowledge source and places them in a language model's context before the model generates an answer.

Does RAG guarantee a correct answer?

No. Retrieval and generation can both fail. Teams should evaluate source retrieval, answer support, citations, permissions and freshness separately.

Business documents mapped into a semantic vector space and retrieved into an AI response

Back to Blog

18 September 2026Posted by Mahdi

AI Concepts in a Nutshell · Part 5 of 8

Embeddings, Vector Search and RAG Explained

Understand embeddings, vector search and retrieval-augmented generation, including how they ground AI responses in approved business information.

Embeddings turn content into numerical vectors that capture useful relationships. Vector search compares those representations to find content with similar meaning. Retrieval-augmented generation, or RAG, uses that search step to place relevant information into a model's context before it answers.

Embeddings make meaning searchable

A traditional keyword search looks for matching terms. An embedding can place semantically related content near each other even when the wording differs. A query about “ending an agreement” may retrieve a passage headed “contract termination” because the representations are similar.

Embeddings are not readable facts and proximity is not proof. They are mathematical representations optimised for a task. Search quality still depends on the source content, chunking, metadata, model and ranking method.

The basic RAG flow

Collect approved documents or records.
Split them into useful passages and preserve metadata such as source, owner, access level and review date.
Create embeddings and store the vectors with their source references.
Embed the user's question and retrieve likely relevant passages.
Give those passages to the language model with instructions to answer from the evidence.
Return citations or source links and record the retrieval result for evaluation.

Diagram showing approved sources being embedded in a vector index, retrieved for a question and supplied to a model for a cited answer — RAG retrieves relevant passages from an approved index and places them beside the question so the model can answer from evidence and return citations.

RAG does not fix bad knowledge

If documents are outdated, duplicated or contradictory, retrieval can surface the wrong passage efficiently. Establish content owners, review dates, version rules and a deletion process. Preserve access controls during indexing and retrieval; a model must not receive a document merely because it is semantically relevant.

Measure retrieval separately from generation

First ask whether the correct source appeared in the retrieved set. Then ask whether the final answer used it accurately. This separation shows whether to improve the search index, ranking, prompt or model. A single thumbs-up score cannot locate the fault.

RAG is useful for policies, product documentation, service knowledge and other changing information. It is not a substitute for a transactional query when the answer must come from a live balance, booking, permission or order state.

Next: Chatbots, Copilots and AI Agents: Key Differences.

Frequently Asked Questions