Human Expertise in Text to SQL (Databricks, Snowflake, and Numbers Station)
Recent work in Text-to-SQL shows that once you get past demo datasets, the performance drops. By incorporating human expertise, you can build better Generative AI systems for Text to SQL.
Let me explain and give you more resources:
🔬 Databricks researchers released a paper with a more realistic Text-to-SQL benchmark (a lot of existing datasets were way too easy). They found:
• Using existing Text-to-SQL models provides pretty poor performance (something I hear about all the time)
• They were able to markedly improve the performance with the addition of hand-written pipelines that leverage expert knowledge of the table schema
Source: https://arxiv.org/pdf/2408.14717
❄️ Snowflake has shared their agentic workflow for their Text-to-SQL service. Key points:
• A crucial part is a semantic layer for enrichment
• This semantic layer often works best when experts add their knowledge
• (Disclaimer: I work for Snowflake and demo this product)
Source: https://www.snowflake.com/engineering...
🔢 Numbers Station AI, a startup that cares about Text-to-SQL, shared their learnings on the difficulties of building enterprise-grade Text-to-SQL solutions
Source: https://www.youtube.com/live/xmzda44hUgk
━━━━━━━━━━━━━━━━━━━━━━━━━
★ Rajistics Social Media »
● Home Page: http://www.rajivshah.com
● LinkedIn: / rajistics
━━━━━━━━━━━━━━━━━━━━━━━━━
Смотрите видео Human Expertise in Text to SQL (Databricks, Snowflake, and Numbers Station) онлайн без регистрации, длительностью часов минут секунд в хорошем качестве. Это видео добавил пользователь Rajistics - data science, AI, and machine learning 01 Сентябрь 2024, не забудьте поделиться им ссылкой с друзьями и знакомыми, на нашем сайте его посмотрели 43 раз и оно понравилось 6 людям.