The following step can also be being safe. The problem is, if the LLM can run any potential question towards the database, then how do you ensure you don’t exfiltrate and leak info? We’ve constructed this know-how that we name parameterized safe views within the database itself, that allows you to outline the precise safe obstacles and encodes the safety insurance policies that you simply want, in order that the LLM can generate any question it needs, however with respect to the logged-in person we won’t allow them to see any info that they don’t seem to be imagined to see. We’ll additionally, on an information-theoretical foundation, not leak info that they need to not have entry to.
Heller: I do know you’ve spent plenty of time desirous about the way forward for databases and generative AI. The place are we headed?
Krishnamurthy: A part of my considering right here has advanced during the last couple of years, however for 50 years the world of databases has been no less than SQL databases the place it was all about producing actual outcomes. I prefer to say databases had one job: retailer the info, don’t lose the info, after which once you ask a query, give the precise outcome. OK, possibly two jobs. It was all about actual outcomes as a result of we’re coping with structured information. I feel the largest change that’s occurring proper now could be that we’re not simply coping with structured information. We’re additionally coping with unstructured information. Once you mix structured and unstructured information, the following step is that it’s not nearly actual outcomes however about essentially the most related outcomes. On this sense databases begin to have among the capabilities of engines like google, which is about relevance and rating, and what turns into vital is sort of like precision versus recall for info retrieval methods. However how do you make all of this occur? One key piece is vector indexing. In different phrases, you’ve structured information, which is within the database, however now we have different kinds of knowledge, unstructured information, semi-structured information.
