Pekka Enberg

When there was this first wave of ChatGPT 3.5, I think that was the sort of, at least for me, the sort of turning point. Then all of a sudden we were in this situation that everybody wanted to apply these large language models to their applications. And basically the models themselves are super useful, but there's this problem called hallucination because they just make up stuff.

When there was this first wave of ChatGPT 3.5, I think that was the sort of, at least for me, the sort of turning point. Then all of a sudden we were in this situation that everybody wanted to apply these large language models to their applications. And basically the models themselves are super useful, but there's this problem called hallucination because they just make up stuff.

When there was this first wave of ChatGPT 3.5, I think that was the sort of, at least for me, the sort of turning point. Then all of a sudden we were in this situation that everybody wanted to apply these large language models to their applications. And basically the models themselves are super useful, but there's this problem called hallucination because they just make up stuff.

So these large language models essentially... are limited to whatever they saw during training. And these things get trained by reading essentially through the whole internet. But there's always a cutoff date, right? So you train it and then after that, it doesn't really know about the new things that appear. But also for enterprises, these models don't really know your company specific data.

So these large language models essentially... are limited to whatever they saw during training. And these things get trained by reading essentially through the whole internet. But there's always a cutoff date, right? So you train it and then after that, it doesn't really know about the new things that appear. But also for enterprises, these models don't really know your company specific data.

So these large language models essentially... are limited to whatever they saw during training. And these things get trained by reading essentially through the whole internet. But there's always a cutoff date, right? So you train it and then after that, it doesn't really know about the new things that appear. But also for enterprises, these models don't really know your company specific data.

information. And that's why people came up with this retrieval augmented generation, which is essentially just retrieving data for the model. And this is where the vector search part comes in. Imagine an interface where you have a customer typing a question. So the way it essentially works is that you take that question, you run it through a

information. And that's why people came up with this retrieval augmented generation, which is essentially just retrieving data for the model. And this is where the vector search part comes in. Imagine an interface where you have a customer typing a question. So the way it essentially works is that you take that question, you run it through a

information. And that's why people came up with this retrieval augmented generation, which is essentially just retrieving data for the model. And this is where the vector search part comes in. Imagine an interface where you have a customer typing a question. So the way it essentially works is that you take that question, you run it through a

a large language model, generate an embedding, which is a vector. And then you use this vector or this embedding to find relevant information. And that relevant information is through vector search, which is managed in some database. For me, the really interesting thing is that initially what happened was that there was this like explosion of different special purpose databases, vector databases.

a large language model, generate an embedding, which is a vector. And then you use this vector or this embedding to find relevant information. And that relevant information is through vector search, which is managed in some database. For me, the really interesting thing is that initially what happened was that there was this like explosion of different special purpose databases, vector databases.

a large language model, generate an embedding, which is a vector. And then you use this vector or this embedding to find relevant information. And that relevant information is through vector search, which is managed in some database. For me, the really interesting thing is that initially what happened was that there was this like explosion of different special purpose databases, vector databases.

At some point they were embedding databases, but then I think the everybody's converged on vector databases. And these are special purpose thing to do just retrieval part. But quickly people also discovered that, hey, we still have this traditional data that we want to access, but also lots of different databases and data sources. So like, how can we simplify this thing?

At some point they were embedding databases, but then I think the everybody's converged on vector databases. And these are special purpose thing to do just retrieval part. But quickly people also discovered that, hey, we still have this traditional data that we want to access, but also lots of different databases and data sources. So like, how can we simplify this thing?

At some point they were embedding databases, but then I think the everybody's converged on vector databases. And these are special purpose thing to do just retrieval part. But quickly people also discovered that, hey, we still have this traditional data that we want to access, but also lots of different databases and data sources. So like, how can we simplify this thing?

And then you had a lot of Postgres was adding this extension and so forth. But with SQLite, what is really interesting is because it is such a lightweight thing and you can run it in mobile devices, for example, when you actually bring this vector capability to SQLite, you can do all of this model work. related processing and all of that searching within the device itself.

And then you had a lot of Postgres was adding this extension and so forth. But with SQLite, what is really interesting is because it is such a lightweight thing and you can run it in mobile devices, for example, when you actually bring this vector capability to SQLite, you can do all of this model work. related processing and all of that searching within the device itself.

And then you had a lot of Postgres was adding this extension and so forth. But with SQLite, what is really interesting is because it is such a lightweight thing and you can run it in mobile devices, for example, when you actually bring this vector capability to SQLite, you can do all of this model work. related processing and all of that searching within the device itself.

And you can imagine you have the latency advantage, but also increasingly people are super interested in the sort of privacy aspect, right? Because now you can have the private information on the device. It doesn't necessarily have to leave the device. So I think that's the cool part in using the old traditional SQL database and then vector search.

And you can imagine you have the latency advantage, but also increasingly people are super interested in the sort of privacy aspect, right? Because now you can have the private information on the device. It doesn't necessarily have to leave the device. So I think that's the cool part in using the old traditional SQL database and then vector search.