Vectors of numbers encoding what text means, placing related content close together as judged by cosine similarity. Switching model or vector size requires embedding all content again.
Also called vector embeddings.
Read more: Microsoft Learn
In the Ultra Transcenders books
Each book explains Embeddings in context, with comparison tables and the common traps.
Terms in this definition
- ENCODING
A COPY INTO option for CSV sources that states whether the files use UTF8, the default, or UTF16 character encoding.
- RELATED
Fetches a column value from the lookup table, that is the one side of a many-to-one relationship, for whichever row is being evaluated. It therefore needs to run inside an iterator or a calculated column, where row context exists.
- Cosine similarity
Scores how semantically close two embedding vectors are by the angle between them; vector search relies on it.
- Embedding
A list of numbers, spanning many dimensions, that captures meaning; items with similar meaning end up near one another, as cosine similarity measures.
- ALL
A DAX function that ignores any filters and gives back every row of a table or every value of the named columns. Used within CALCULATE, it works as a modifier that clears filters, although REMOVEFILTERS states that intent more clearly where it is available.
Related terms
- Azure OpenAI
Use of OpenAI models in Azure, including GPT, o-series, embeddings, image and Whisper, through either a Foundry resource or an Azure OpenAI resource. Content Understanding isn't available from a standalone Azure OpenAI resource.
- Azure OpenAI in Foundry Models
OpenAI models offered by Azure as a fully managed service with guardrails built in, covering GPT-4.1, GPT-5, o-series, image, audio and embeddings. Rather than running infrastructure, you pick a deployment type.
- External model
A database object, made with CREATE EXTERNAL MODEL, that stores the details AI_GENERATE_EMBEDDINGS needs to call an AI model: endpoint, API format, model type (EMBEDDINGS), model name and credential. It is changed with ALTER, and dropping it leaves the credential in place.
- GIN
PostgreSQL index that the book places on jsonb metadata columns; scalar filtering relies on B-tree and embeddings on vector indexes.
- Integrated vectorization
Azure AI Search capability that lets an indexer split documents into chunks using the Text Split skill and produce embeddings via a skill like Azure OpenAI Embedding; a corresponding vectorizer on the index turns search queries into vectors.
- ONNX Runtime
An open-source engine for running ONNX models on the machine itself. When installed on Windows, SQL Server 2025 points an external model at it with LOCAL_RUNTIME_PATH, so embeddings are created with nothing remote involved.
- Semantic ranker
Feature of Azure AI Search that uses Bing-derived language models to re-order the leading BM25 or hybrid results, optionally adding answers and captions. Relevance improves with no change to embeddings.
- SQL Server 2025
This SQL Server version brings AI into the engine itself: a vector type with vector indexes, external models, and AI_GENERATE_EMBEDDINGS. Native json and regular expression functions arrived with it too.