Vector embeddings in ClickHouse with Ollama and OpenAI

Mark Needham

ClickHouse's new aiEmbed function generates embeddings directly in SQL — no external pipeline needed. In this video, Mark Needham shows how to use it with both a local Ollama model and OpenAI.

https://clickhouse.com/blog/clickhouse-release-26-06#aiembed

• Point aiEmbed at a local Ollama model (Qwen3 Embedding) via a named collection with an OpenAI-compatible endpoint
• Auto-generate embeddings on insert using a DEFAULT aiEmbed(text) column, backed by an HNSW vector similarity index
• Run approximate nearest neighbor search with L2Distance to find phrases closest to a lookup embedding
• Query a 1M-row Hugging Face dbpedia dataset pre-embedded with OpenAI's text-embedding-3-large
• Switch aiEmbed to OpenAI, match embedding dimensions to the dataset, and search for "the best footballer ever"