I have been working on similar stuff. I started with a graphdb in pure Python file-based then switched to EAV and bsddb. Nowdays I work mainly on https://github.com/amirouche/hoply/ and https://github.com/awesome-data-distribution/datae.
I have not touched yet the NL to database query part because there is no silver bullet. At the end of the day, one need a use case, existing problems to come up with a solution. I learned the hard way that coming up with an elegant solution is not enough it must help with an actual need. For that, there is no silver bullet either. Experience helps. Being curious helps too.
I thought that the chatbot mania would stir new tools. Chatbots are narrow.
See also:
- https://en.wikipedia.org/wiki/Concept_search
- legaltech
I know there isn't a silver bullet, and NL queries will never be as versatile as a real query language, but from testing with various people I know it _seems_ to be an OK on-ramp to more advanced stuff.