[tillvz]

Trust & explainability is the biggest issue here.

We've been building natural language analytics at Veezoo (https://www.veezoo.com/) for 10 years, and what we find is that straight Text-to-SQL doesn't scale. If AI writes SQL directly, you're building on a probabilistic foundation. When a CFO asks for revenue the number can't just be correct 99% of times. Also you can't get the CFO to read SQL to verify.

We're solving that with an abstraction layer (Knowledge Graph) in between. AI translates natural language to a semantic query language, which then compiles to SQL deterministically.

At the same time you can translate the semantic query deterministically back into an explanation for the business user, so they can easily verify if the result matches their intent.

Business logic lives in the Knowledge Graph and the compiler ensures every query adheres to it 100%, every time. No AI is involved in that step.

Veezoo Architecture: https://docs.veezoo.com/veezoo/architecture-overview

[kburman]

Thanks for sharing the links, the architectural overview is very insightful.

I'm curious how this approach manages cardinality explosion? Also, how do you handle cases where a user asks for data that requires running multiple queries, specifically where each query depends on the results of the previous one?

[tillvz]

> I'm curious how this approach manages cardinality explosion?

The Knowledge Graph explicitly models cardinality and relationships between entities. The compiler uses that to generate SQL that handles it correctly, using e.g. DISTINCT

> Also, how do you handle cases where a user asks for data that requires running multiple queries, specifically where each query depends on the results of the previous one?

Veezoo can generate adaptive plans, so it can decide to wait for a database query to return results before continuing

[kburman]

Thanks for answering! Regarding cardinality, I was actually thinking more about high-cardinality dimensions on the NLU side, e.g., if a user asks for "Sales for [Obscure Company Name]," and you have 10M distinct customers. Does the Knowledge Graph have to index all those values for the mapping to work?

On the adaptive plans, Is that execution logic handled entirely by your deterministic compiler, or does it loop back to the LLM to interpret the intermediate results?

[tillvz]

>Does the Knowledge Graph have to index all those values for the mapping to work?

There are both options. You can index them as entities [1] within Veezoo and keep the mapping automatically synchronized with the database. Or decide to not index them, which will make Veezoo e.g. attempt answering the question using string search in SQL.

>On the adaptive plans, Is that execution logic handled entirely by your deterministic compiler, or does it loop back to the LLM to interpret the intermediate results?

The plan is done entirely by the LLM. The VQL steps (i.e. fetching answers from the database) within the plan is where the compiler kicks in.

[1] https://docs.veezoo.com/vkl/kb-layer/entity/

[Leynos]

Don't you still need to unit test and version control the SQL artefact that is produced? You need to be able to see which query was used on which date and how it was validated.

(Prompts need to be version controlled too, of course)

[tillvz]

Yes, every SQL query Veezoo runs is logged and visible to admins.

The fundamental artifact is VQL (Veezoo Query Language), which queries against a Knowledge Graph containing your business data model, things like your "Revenue" measure.

A query might look like this:

var order from kb.Order

date_in(order.Order_Date, date("#today"))

var retRevenue = kb.Order.Revenue(order)

select(retRevenue)

If the business decides to change how revenue is computed, the VQL stays valid but compiles to different SQL. At the same time Veezoo can test that with your knowledge graph change that you are not breaking anyones dashboard and even apply evolutions if needed

VQL: https://docs.veezoo.com/vkl/kb-layer/vql/

Evolutions: https://docs.veezoo.com/vkl/evolutions/

The Knowledge Graph itself is version controlled, so the data team can trace every change.