Artem Keydunov: recommendation

Transcript context

…Can you run through the steps of how that works today? So the initial part is like the natural language query, like what are the steps that happen in between to do model, to semantic layer, semantic layer, to SQL and all that flow? The first key step is to do some sort of indexing. That's what I was referring to, like write a book about your data, right? Describe in a text format what your data is about, right? Like what metrics it has, dimensions, what is the structures of that, what a relationship between those metrics, what are potential values of the dimensions. So sort of, you know, like build a really good index as a text representation and then turn it into embeddings into your, you know, like a vector storage. Once you have that, then you can provide that as a context to the model. I mean, there are like a lot of options, like either fine tune or, you know, like sort of in context learning, but somehow kind of give that as a context to the model, right? And then once this model has this context, it can create a query. Now the query I believe should be created against semantic layer because it reduces the room for the error. Because what usually happens is that your query to semantic layer would be very simple. It would be like, give me that metric group by that dimension and maybe that filter should be applied. And then your real query for the warehouse, it might have like a five joins, a lot of different techniques, like how to avoid fan out, fan traps, chasm traps, all of that stuff. And the bigger query, the more room that the model can make an error, right? Like even sometimes it could be a small error and then, you know, like your numbers is going to be off. But making a query against semantic layer, that sort of reduces the error. So the model generates a SQL query and then it executes us again, semantic layer. And semantic layer executes us against your warehouse and then sends result all the way back to the application. And then can be done multiple times because what we were missing was both this ability to have a conversation, right? With the model. You can ask question and then system can do a follow-up questions, you know, like then do a query to get some additional information based on this information, do a query again. And sort of, you know, like it can keep doing this stuff and then eventually maybe give you a big report that consists of a lot of like data points. But the whole flow is that it knows the system, it knows your data because you already kind of did the indexing and then it queries semantic layer instead of a data warehouse directly. Maybe just to make it a little clearer for people that haven't used a semantic layer before, you can add definitions like revenue, where revenue is like select from customers and like join orders and then sum of the amount of orders. But in the semantic layer, you're kind of hiding all of that away. So when you do natural language to queue, it just select revenue from last week and then it turns into a bigger query.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.