AI/ML Update - OpenHouse Amsterdam 2025

https://www.youtube.com/watch?v=wmbCqXvyynI

Transcript

1. 1 ClickHouse for AI / ML Pete Hampton Principal Software Engineer
3. ©2025 CLICKHOUSE INC., CONFIDENTIAL & PROPRIETARY 3 UDFs in Cloud now in private preview • Communicate with LLMs (generate embeddings, enrich data) • Perform complex financial risk modeling • …anything else that isnʼt easily expressed in SQL
4. ©2025 CLICKHOUSE INC., CONFIDENTIAL & PROPRIETARY 4 Vector similarity index now generally available Also supports BFloat16 (default) and int8 quantization
5. 5 + UDFs Vector Search Generate & store vector embeddings for your analytical data Retrieve & analyze data based on vector similarity (...and everything else) SELECT embed('ClickHouse is cool') [-0.022583008,0.003112793,0.025024414,0.013427734,0.02 331543,-0.014587402,-0.010803223,0.008117676,0.0123901 37,-0.067871094,0.012451172,-0.04272461,-0.013061523,- 0.03881836,0.0043945312,0.04638672,-0.006134033,-0.006 164551,-0.03173828,0.06738281,-0.032714844,0.03955078, 0.001335144,0.013427734,-0.012634277,0.03955078,-0.051 02539,0.0069885254,0.012878418,-0.049804688,0.08300781 , ||.] SELECT user, uniq(commentId), avg(score), FROM user_comments WHERE cosineDistance(embed(title || caption), targetVector) > 0.2) GROUP BY user ORDER BY uniq(commentId) DESC
6. 6 ClickHouse for AI/ML • clickhouse-local • clickhouse-server • chDB (in-process) • ClickHouse as an exploratory environment and data transformation tool Exploration / Preparation • ClickHouse as a • - Feature store • - Vector store • ClickHouse + UDFs Training / Inference
7. 7 Human-machine interface
10. 10 Circa 2022
11. 11 Classic real-time analytics Middleware layer SQL queries
12. 12 Real-time analytics needs… Always fresh data blazing fast queries scalability to thousands of users and with
13. 13 Real-time analytics needs… Always fresh data blazing fast queries scalability to thousands of users and with millions of agents
14. 14 > Agentic analytics
16. 16 Agentic analytics SQL queries
17. 17 What do you need for high performing AI Agents? Supports rapid query bursts Maintains seamless UX Delivers results in ms Analytics in context Complex, large-scale data aggregations Real-time business pattern insights Agentic observability Low latency data and cost efficiency Immediate access to live data Minimal hardware required, lower AI spend Unified data access Consolidated multi-system domain data Reliable analytics source-of-truth Process structured & unstructured data High performance querying
18. 18 DB Analysts
20. 20 Model Context Protocol
21. ©2025 CLICKHOUSE INC., CONFIDENTIAL & PROPRIETARY 21 Model Context Protocol MCP LLMs mostly rely on training data MCP links AI apps with external tools and private data EXAMPLES • Company DB queries • Local file access • Smart-home control • Live service data Fully managed remote MCP server • Turn-key experience, no infra to setup or manage • Built into ClickHouse Cloud • Leverage your data with external agents and MCP-compatible clients • Secured with OAuth for authentication • Bring your own MCP-compatible client ◦ Claude, Cursor, Windsurf, 25 others
22. 22 Developers DB Analysts
24. 24 Literally anyone else. Developers DB Analysts
25. ©2024 CLICKHOUSE INC., CONFIDENTIAL & PROPRIETARY 25 The database for the agentic era Tesla OpenAI