We also have an eval space on huggingface where you can evaluate different prompts and models (for DuckDB) -->
huggingface.co/spaces/duckd...
And a leaderboard -->
huggingface.co/spaces/duckd...

It's an interesting question. Thinking that pay-as-you-go endpoints with general purpose models + rag/prompt eng. may be the more practicable and economical choice here. You could do 10k requests/month with a few $. I'm also working on a paper on how to do that well for DuckDB Text2SQL.

Just realized the video was too long. Here with 3x speedup 🏃

Updated our MCP connector and docs recently. It's super easy to install and gives your Claude Desktop, DuckDB superpowers!

🔗 pypi.org/project/mcp-...

Added support for local DuckDB pypi.org/project/mcp-... and a new system prompt with lots of knowledge about DuckDB syntax and functions. Feel free to give it spin - curious how it works for you @davidgasquez.com

Created a experimental MCP server for MotherDuck earlier today, happy to share the code or add local duckdb support pypi.org/project/mcp-...

3 - I'm not sure how to feel about text-only interfaces in that context, esp. with tool-to-tool communication in mind. Maybe I'm missing something, but wish I could just pass an arrow table from the MotherDuck connector to Claude's Data Anlayst. @anthropic.com :)

2 - I believe the true power of this comes from being able to use different tools in the same context. E.g. "Here are some spreadsheets, a presentation, a few docs, and a s3 bucket with some sales data. Tell me what happened in the last 2 months, and put it into context of recent political events"

1 - I like the idea of extending Claude's capablities with e.g. large-scale data analytics, knowledge retrieval etc., all within Claude's familiar user interface.