4 ms·
> There's no way for a model to correctly interpret the meaning of every column in a real world database using the `information_schema` alone. Why would text-t
by dragonwriter 3y ago
> There's no way for a model to correctly interpret the meaning of every column in a real world database using the `information_schema` alone.
Why would text-to-sql be limited to information_schema alone? Human analysts would use additional documentation, why wouldn't an LLM-based text-to-sql system?
- mritchie712 3y agoI should have clarified. There's a large number of apps that are: 1. taking info strictly from SQL (e.g. information_schema, query history) 2. taking a user input / question 3. writing SQL to answer that question An app like this is what I call "text-to-sql". Totally agree a better system would pull in additional documentation (which is what we're doing), but I'd no longer consider it "text-to-sql". In our case, we're not even directly writing SQL, but rather generating semantic layer queries (i.e. https://cube.dev/ https://cube.dev/).