Ask HN: Integrating SQL and LLM what is you experience

1 pointsyonisto1y ago3 comments

I was tasked with building a "native language interface" to our database. After developing it for a month, it seems to me that I hit a wall with the correctness of the resulting SQL. 80-20 would be a good estimate. My feeling is that it something that cannot be used a regular user.

Did you do better and you have tips?

NOTE: I'm not talking about performant queries only about producing a correct result. I would say that only about 50% cannot be better optimized, but I don't care about that.

3 comments

3 comments · 1 top-level

lunarcave1y ago· 2 in thread

Does this [1] look like the thing you're looking for?

We built a basic Re-Act agent that can dynamically inspect the database context and generate the queries iteratively.

Our experience was that it was hard to do single-shot.

Happy to answer any questions.

[1] https://github.com/inferablehq/inferable/tree/main/adapters/...

yonistoOP1y ago

Thanks

I did build something similar (although in python and with some other hard requirements).

My problem, and seems that your experience it too, is that the users will never know that the query is invalid. They will get a result for sure, but the SQL could be totally wrong.

lunarcave1y ago

Oh, you mean in terms of semantics? Yeah, that's a bit tricky.

I guess that's true for any LLM related result though.

j / k navigate · click thread line to collapse