Teaching databases to speak human with LLMs and MCP
summary
Demo connects an LLM through Model Context Protocol (MCP) to a live Kafka stream, reads minute‑level crypto data from Apache Kafka into Apache Iceberg, and queries it via Trino. The system lists catalogs and schemas, runs small subqueries to probe table structure, retries on parser errors, then generates efficient SQL and returns the answer; benchmarks cited (Spider, BIRD) show humans ~93% vs top models ~77% and the demo succeeds after a couple of retries. Recommendations: start curated data, use read‑only roles and strict MCP permissions.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 3:01Natural language queries use case
- 8:28Natural language to SQL challenges
- 15:17Model Context Protocol (MCP)
- 20:04Apache Kafka streaming
- 23:35Apache Iceberg lakehouses
- 26:45Evaluating and securing system
- 28:59Adoption best practices
- 30:01Q&A
related talks
35mThe Missing Protocol: How MCP Bridges LLMs and Data Streams
53mAgents, Tools, and Mcp, Oh My! Next Level AI Concepts for Developers
32mMCP for dummies: Feed more context to LLM
28mTalk to your data without coding agentic AI and Claude with the MCP protocol
38mLLMs for Self-serve Analytics - Lessons Learned
46mMCP for LLM Agents, Infrastructure APIs & Data Services
24mAn MCP for your Postgres DB
16mKeynote: DBRX: Training Modern LLMs on Databricks