> For the complete documentation index, see [llms.txt](https://mariadb.com/docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://mariadb.com/docs/ai.md). # AI Keep vectors next to the data they describe. Add semantic search to MariaDB in plain SQL, with no separate vector store to keep in sync. The shop sells the three products from the landing page: a desk lamp, a standing desk, and a notebook. A customer opens the shop and types "something to light my desk" into the search box. The search runs the query that every shop starts with, a `LIKE` match on the product name: ```sql SELECT name FROM shop.products WHERE name LIKE '%light%'; ``` It returns an empty set. The shop sells exactly what the customer wants, but the product is called "Desk lamp", and the word "light" appears nowhere in it. Keyword matching compares letters, so it can't see that a lamp is a light. Semantic search compares meanings instead. An embedding model, which is an AI model that turns text into a list of numbers called a vector, places "something to light my desk" and "Desk lamp" close together because they mean similar things. MariaDB doesn't include an embedding model. Your application calls one, from a provider or running locally, and MariaDB Server stores the vectors in the same database as the `products` table and finds the nearest ones in SQL.

Semantic search: your application calls an embedding model, from a provider or running locally, to turn the search phrase into a vector, then sends one SQL query to MariaDB Server. The query walks the vector index on the product_embeddings table, joins the products table for names, and returns the two nearest products with their distances. — Your application calls the embedding model, so the database only ever compares numbers, and one query returns both the matches and their names.

Vector search is built into MariaDB Community Server, introduced in 11.7 and generally available from the 11.8 long-term release, and into MariaDB Enterprise Server from 11.4.5-3, so there's no plugin or extension to install. ## Add Semantic Search {% stepper %} {% step %} ### Create a Vector Table Create a table that holds one `VECTOR` value per product, with a `VECTOR INDEX` built for cosine distance, which measures how closely two vectors point in the same direction. ```sql CREATE TABLE shop.product_embeddings ( product_id INT PRIMARY KEY, embedding VECTOR(4) NOT NULL, VECTOR INDEX (embedding) DISTANCE=cosine ); ``` The indexed column must be `NOT NULL`, and a table can have only one vector index. [CREATE TABLE with Vectors](/docs/server/reference/sql-structure/vectors/create-table-with-vectors.md) covers the `M` and `DISTANCE` options. {% endstep %} {% step %} ### Insert the Vectors The `VEC_FromText` function turns a JSON array of numbers into a vector value. Every row needs the same number of dimensions as the column, which is four here so the numbers stay readable. A real embedding model produces hundreds or thousands, for example `VECTOR(768)`, and your application code writes them. The ids 1 to 3 are the desk lamp, the standing desk, and the notebook from [step 2 on the landing page](/docs/readme.md#create-some-data). ```sql INSERT INTO shop.product_embeddings (product_id, embedding) VALUES (1, VEC_FromText('[0.91, 0.12, 0.30, 0.05]')), (2, VEC_FromText('[0.20, 0.88, 0.35, 0.10]')), (3, VEC_FromText('[0.15, 0.10, 0.20, 0.95]')); ``` The [VEC\_FromText reference](/docs/server/reference/sql-functions/vector-functions/vec_fromtext.md) shows the binary format it produces, which is four bytes per dimension. {% endstep %} {% step %} ### Find the Nearest Products The vector in this query stands in for the customer's phrase after the same model embeds it. Sort by the `VEC_DISTANCE_COSINE` function and stop at a `LIMIT`, because that `ORDER BY ... LIMIT` shape is what lets the optimizer walk the vector index. The join to the `products` table brings back names in the same query. ```sql SELECT p.name, VEC_DISTANCE_COSINE(e.embedding, VEC_FromText('[0.85, 0.20, 0.25, 0.10]')) AS distance FROM shop.product_embeddings AS e JOIN shop.products AS p ON p.id = e.product_id ORDER BY distance LIMIT 2; ``` ``` +---------------+---------------------+ | name | distance | +---------------+---------------------+ | Desk lamp | 0.00688833819260537 | | Standing desk | 0.5012322990863076 | +---------------+---------------------+ ``` A smaller distance means a closer match. [Using the vector index efficiently](/docs/server/reference/sql-structure/vectors/vector-overview.md#using-the-vector-index-efficiently) lists the query patterns that fall back to a full table scan. {% endstep %} {% endstepper %} The customer who typed "something to light my desk" now sees the desk lamp first, and the notebook doesn't show up at all. ## Build With a Coding Agent The steps above use made-up vectors. Wiring in a real embedding model, backfilling every product, and keeping the vectors current when a description changes is application code. MariaDB doesn't include a coding agent, but the one you already use can write most of that code with you if it knows MariaDB. The [MariaDB AI plugins](https://github.com/mariadb/ai-plugins) give Claude Code, Codex, Pi, Hermes Agent, and OpenCode a set of MariaDB skills, including one for vectors, plus a local sandbox server where the agent can try its SQL before you run it. Add the docs MCP server from the [landing page](/docs/readme.md#keep-your-agent-in-the-docs) as well, so the agent reads the current `VECTOR` syntax instead of guessing from older training data. Ask it to add an embedding column for product descriptions, and it can check its own work against the reference. [AI Tools](https://mariadb.com/docs/ai-tools/) covers both kinds of agent work: coding agents that build with you, and production agents that reach live data over MCP. ## Keep Going


AI Tools	Set up a coding agent that knows MariaDB, and choose an MCP server for agents that work on production data.	AI Tools
Vector Overview	Read about the `VECTOR` data type, the HNSW-based vector index, and the distance functions before you design a production schema.	Vector Overview
Hybrid Search	Merge vector matches and full-text keyword matches into one ranking with reciprocal rank fusion.	Optimizing Hybrid Search Query with Reciprocal Rank Fusion (RRF)
Framework Integrations	Use MariaDB as the vector store for LangChain, LlamaIndex, Spring AI, and other frameworks.	Vector Framework Integrations
MariaDB Enterprise MCP Server	Give agents access to production databases with the authentication and permissions you set.	Broken link
Vector Search Workshop	Build semantic search step by step in Python with LangChain and MariaDB.	https://github.com/mariadb-developers/vector-search-workshop
RAG Demo in Java	Run retrieval augmented generation on MariaDB vector search with the OpenAI API.	https://github.com/mariadb-developers/mariadb-rag-demo-java