AI Builders: Building a text-to-SQL agent

Writing SQL has long been one of the biggest barriers standing between business users and the data they need. In this episode of AI Builders, we build a text-to-SQL analytics agent inside a Marimo notebook using DuckDB, Plotly, and an LLM-powered SQL generation workflow, then layer in observability and tracing and discuss what it takes to evolve the prototype into a production-ready reporting application. If you're a developer interested in building practical AI-powered analytics and reporting applications, this video is for you. Use the blog post and GitHub repo below to follow along and start experimenting. 💻 GitHub repo: https://github.com/rratshin-wandb/ai-builders 📘 Blog post: https://wandb.ai/wandb_fc/ai-builders/reports/Building-a-text-to-SQL-agent--VmlldzoxNzAyMDIwNQ NOTE from Russ: Second episode of the AI Builders series of YouTube videos "for developers who love to build and are having a great time bringing the wonders of AI into our applications." While this video does include mention of Weights & Biases products, the intro sequence should not be branded. Video will be linked to from a W&B blog post and GitHub repo. Thank you! ⏳Timestamps: 0:00 — Introduction and motivation for a text-to-SQL agent 1:39 — Diving into the agent notebook 2:08 — Overview of the project star schema and data model 2:23 — System prompt design and JSON output format 2:58 — Agent class design and function walkthrough 3:56 — Building Plotly charts from LLM-returned chart config 4:32 — Invoking the agent and inspecting results in Marimo 5:36 — Tracing and observability 6:28 — Monitoring and the path to production