Revolutionizing Text Transformation.
summary
Claims an LLM-based pipeline that goes beyond standard SQL to produce text-to-streaming-SQL and supports RisingWave for window joins, plus Python/dataframe and notebook SDKs. Metadata retrieval combines embedding+graph+LLM retrieval to pick the right tables. A compiler layer auto-patches and retries queries—benchmarks show gpt3.5 fails 54% without fixes, gpt4 out-of-box 22.8%, and the compiler-driven system drops failures to 2.8% on an internal benchmark; Spider tests drop from 18.4% to ~3.9%. An open-source evaluator called AERfish is available.
AI-written summary · about
chapters (11)
- 0:00Introduction and speaker intro
- 1:54History of text to SQL
- 3:03Building a basic text to SQL system
- 4:06Challenges with large language models
- 9:09Improving table retrieval with knowledge graphs
- 11:17Fixing compilation errors automatically
- 14:27Enforcing semantic correctness with rules
- 15:52Evaluation framework for SQL generation
- 18:04Live demo of the system
- 27:02Demo of SQL migration and notebook integration
- 28:29Q&A and closing
related talks
30mThe Future of Data is Words
24mHarnessing the Power of LLMs in Advanced Analytics
32mRevolutionizing Automation: Retrieval-Augmentation with LLMs
29mHow to roll your own LLM
26mReverse Engineering Using LLMs
15mFrom Words to Data: Using LLMs to Turn Questions into SQL
31mFrom Data to Precision: Fine-Tuning LLMs for Proprietary Query Languages
57mLLMs gone wild