A semantic layer is more than just a BI query layer, and this distinction is even more important for the use of data with GenAI

#AgileData

Every data vendor is racing to add “GenAI” capabilities to their product, the “text to SQL” capability will be a tablestake feature for most vendors by the end of 2024, even if the the answers returned by the agent are often wrong.

I am seeing lots of webinars and posts talking about the need for a Semantic Layer to improve the LLM responses.

And the majority of these are focussed on the idea that a Semantic Layer is the same pattern as the BI Semantic Layer of old (BO Universe, Oracle End User Layer etc).

I believe it needs to be more than this. It needs to become a first class citizen in our archtecture, rather than something that is bolted on the end of our Information Factory.

It needs to be the place that we hold all semantics about the data, that can be used by any tool, including LLMs.

We need to follow the DORO principle, we need to Define these semantic objects Once, and we need to Reuse them Often.

We can define the name of a thing, we can hold common aliases, we can add some descriptive business context for that thing, we can hold statistics about the shape of the data, we can hold how often it was used and by who and when, we can determine if it is used often and likely to be valuable, we can hold qualitative feedback from user about the data.

And if we store this information as a first class citizen, we can then use this data for multiple use cases.

We can use it to create the BI Semantic Layer, to improve the response from the Text to SQL capability.

We can also use it to suggest data quality rules, we can use it personalise the data catalog, we can use it to point users at data that may answer their complex business questions.

We can use it for the hundreds of valuable GenAI use cases I have yet to discover.

But first we need to treat the semantics of our data as a first class citizen, not isolated in the BI Semantic layer of old (even if it is using new cool technology to do it).

Otherwise we are just creating yet another data silo, and we all know how that turns out.