Why Retrieval Augmented Generation Outperforms Fine Tuning

Key Takeaways

[Music] When it comes to implementing enterprise application, do you see fine-tuning being anywhere in this situation? Fine tuning is a useful tool in the toolbox of it. But let's say injecting information isn't the first thing I would reach for finetuning uh to do. So definitely including that information at query time with rag light systems is the best way to start with injecting let's say or grounding a model in a specific data set that we want which tends to be the use case that is in high demand and it's where you know a lot of people in companies require. Finetuning comes when for example you have a very deep domain that the model was not exposed to. You know, you have a a company with millions of documents with product names and jargon and let's say abbreviations that are not commonly used outside a system like this would benefit somewhat from being fine. Or if you want a very specific format that you're not able to get from just prompting, that's another high level sort of way to use fine tuning to do. But if you want to instill as you know information into a model, rag is better for a lot of reasons. It's cheaper. You can update it at any time without the need to continue you know retraining the model. So that's why I would definitely reach towards raglike systems for a lot of these approaches and leave fine tuning for other more advanced needs that after experimenting with solving it with prompt engineering then I think about things like