Andrew Ng and Nikita Namjishi introduce text embeddings, vectors that capture semantic meaning. Learn how embeddings power text search, clustering, and retrieval-augmented generation, which lets LLMs answer questions using external documents without fine-tuning. The course includes hands-on code to build these systems.
Hi, I'm excited to introduce Understanding and Applying Text Embeddings with Vertex AI built in partnership with Google Cloud. Taught by Nikita Namjishi and me, this short course shows you how to use text embeddings that is given a sentence, paragraph or other arbitrary length piece of text, how to compute a feature vector for it that tries to capture the meaning or the semantics of the text, and it turns out this helps you to figure out what pieces of text are similar. If you've heard of word embeddings like word2vec, it's totally fine if you haven't. The short course doesn't assume this as prior knowledge, but these sentence level embeddings that you learn about in this course are significantly more useful and powerful. I'm delighted to introduce Nikita, who is a developer advocate for Generative AI at Google Cloud. Her full time job is to help developers build generative AI applications with LMS large language models. So so I'm looking forward to her sharing her deep experience on building many practical applications of us in this course.
Thanks Andrew. I've had the pleasure of supporting a lot of large companies as well as startups building exciting applications using embedding algorithms. These give us a way to quickly build text search, clustering, keyword extraction and other applications and are a key part of the toolbox for developers using generative AI. I really look forward to sharing what I'm seeing out in the field with you. Specifically, embeddings are a way that we can represent data as points in space where the locations are semantically meaningful. That is, the locations capture something about the meaning of a piece of text. By using a large pre trained embeddings model, you can prototype many text applications in minutes that frankly used to take most teams months, and this is unlocking a lot of exciting and creative new applications.
So as one example, embeddings are an important part of using LLMs for question answering tasks when you want your answers to be based on an external knowledge base that wasn't included in the LLM's original training data. Say you want the LLM to answer questions about an organization you work for or on a specialized domain. The key technique for this, called retrieval augmented generation, requires giving the LLM access information that wasn't included in the training data, and this often is too much information to simply fit into the prompt.
By allowing the large language model to retrieve information from a specific knowledge base, you can also base the response on a particular document and figure out where the answer came from. This is called grounding an LLM where you can get the LLM to basically cite a specific source and significantly reduce the chance of hallucinations, which is when a large language model produces text that seems plausible but isn't factually accurate or grounded in reality. And we can do all of this without specialized model fine tuning, but just using embeddings and a bit of prompting. So that makes the whole development process a lot faster.
After finishing this course, you understand and also end up with code that you can use to compute embeddings, and also be able to use it in your own question answering system or other LLM applications that might refer to a specific document or set of documents. But more broadly, these embeddings are a key tool for generative AI developers, and having this tool in your pocket will be useful for many things you might want to build in the future. So I hope you enjoy the course.
