I think, within the context of fine tuning, one area that I'm very eager to explore is this idea of fine-tuning the embedding models or the embeddings, as opposed to the LLM itself, to optimize the retrieval mechanism within RAG.
Time to meet another face from Bytewax! This time we met with Laura Gutierrez Funderburk, Senior Developer Advocate. We talk about streaming analytics and interesting use cases such as RAG, continue reading or watch the episode to learn more.
My name is Laura Gutierrez Funderburk. I am based in Vancouver, Canada, but I’m originally from Mexico. I got into data science through mathematics, initially planning to become an applied mathematician. Along the way, I learned about programming, particularly Python, and how it was used to solve problems in bioinformatics. I became hooked on using Python and programming to tackle real-life problems. Fast forward a few years, and I now have the opportunity to apply both my mathematical skills and programming skills to solve problems in data science. I am also involved in community outreach as part of Bytewax.
One of the things that we're very excited about at Bytewax, being an open source community, is letting people leverage the Python ecosystem while bringing in real time analytics. For us, the purpose is to make it as easy as possible for people to build efficient pipelines. So one of the things that I think is quite nice and that the community likes about us, is we're leveraging the Rust package Timely behind the scenes. And then we have exposed the Python API that allows people to simply import and pip install as you would any other Python package. And then from there you can combine other Python packages with our functionality to build some of these pipelines for data that changes in real time.
So one of the use cases is Retrieval Augmented Generation (RAG). We've learned about fine-tuning as one of the mechanisms to ground an LLM. But the problem with something like fine-tuning is that if you have data that's constantly changing or the fine tuning process is not necessarily successful. You can then turn to RAG as a processor or mechanism to ground the LLM. When you're dealing with curating a database that's changing on the regular, say for example, financial data. You have data coming in every single day, either from the market prices, but also in the form of news, articles or any kind of unstructured information. Leveraging Bytewax and Unstructured, or Bytewax together with Haystack, Langchain, or Hopsworks, allows you to ground the LLM with the latest data.
I think, within the context of fine-tuning, one area that I'm very eager to explore is this idea of fine-tuning the embedding models or the embeddings, as opposed to the LLM itself, to optimize the retrieval mechanism within RAG. So, for instance, if I think about retrieving information from social media platforms where I want to optimize my search for specific hashtags or specific formats that the community is using online, leveraging platforms like Hopsworks or Bytewax is one of those compelling use cases. One of the tricky things about social media, sometimes you have these phenomenons of things blowing up and you want to be in it. You want to leverage it, you want to capture it as it's happening, not tomorrow or the week after, you want to catch it now. I think that's one of the compelling use cases that I see for sure.
One community that I found was really helpful in getting started with working with AI and LLMs was the AI Makerspace community. They're an online based community, they offer workshops and free seminars and webinars. Once a week they have these community sessions where the members of the community can come in and talk about a topic. It’s very broad in terms of the community, but they have a very strong focus on building, shipping and sharing.
Listen to the full episode: