Click the image above to view video of this lesson
There's more to LLMs than chatbots and text generation. It's also possible to build search applications using Embeddings. Embeddings are numerical representations of data also known as vectors, and can be used for semantic search for data.
In this lesson, you are going to build a search application for our education startup. Our startup is a non-profit organization that provides free education to students in developing countries. Our startup has a large number of YouTube videos that students can use to learn about AI. Our startup wants to build a search application that allows students to search for a YouTube video by typing a question.
For example, a student might type in 'What are Jupyter Notebooks?' or 'What is Azure ML' and the search application will return a list of YouTube videos that are relevant to the question, and better still, the search application will return a link to the place in the video where the answer to the question is located.
In this lesson, we will cover:
- Semantic vs Keyword search.
- What are Text Embeddings.
- Creating a Text Embeddings Index.
- Searching a Text Embeddings Index.
After completing this lesson, you will be able to:
- Tell the difference between semantic and keyword search.
- Explain what Text Embeddings are.
- Create an application using Embeddings to search for data.
Creating a search application will help you understand how to use Embeddings to search for data. You will also learn how to build a search application that can be used by students to find information quickly.
The lesson includes an Embedding Index of the YouTube transcripts for the Microsoft AI Show YouTube channel. The AI Show is a YouTube channel that teaches you about AI and machine learning. The Embedding Index contains the Embeddings for each of the YouTube transcripts up until Oct 2023. You will use the Embedding Index to build a search application for our startup. The search application returns a link to the place in the video where the answer to the question is located. This is a great way for students to find the information they need quickly.
The following is an example of a semantic query for the question 'can you use rstudio with azure ml?'. Check out the YouTube url, you'll see the url contains a timestamp that takes you to the place in the video where the answer to the question is located.
Now you might be wondering, what is semantic search? Semantic search is a search technique that uses the semantics, or meaning, of the words in a query to return relevant results.
Here is an example of a semantic search. Let's say you were looking to buy a car, you might search for 'my dream car', semantic search understands that you are not dreaming about a car, but rather you are looking to buy your ideal car. Semantic search understands your intention and returns relevant results. The alternative is keyword search which would literally search for dreams about cars and often returns irrelevant results.
Text embeddings are a text representation technique used in natural language processing. Text embeddings are semantic numerical representations of text. Embeddings are used to represent data in a way that is easy for a machine to understand. There are many models for building text embeddings, in this lesson, we will focus on generating embeddings using the OpenAI Embedding Model.
Here's an example, imagine the following text is in a transcript from one of the episodes on the AI Show YouTube channel:
Today we are going to learn about Azure Machine Learning.
We'd pass the text to the OpenAI Embedding API and it would return the following embedding consisting of 1536 numbers aka a vector. Each number in the vector represents a different aspect of the text. For brevity, here are the first 10 numbers in the vector.
[-0.006655829958617687, 0.0026128944009542465, 0.008792596869170666, -0.02446001023054123, -0.008540431968867779, 0.022071078419685364, -0.010703742504119873, 0.003311325330287218, -0.011632772162556648, -0.02187200076878071, ...]The Embedding index for this lesson was created with a series of Python scripts. You'll find the scripts along with instructions in the README in the 'scripts` folder for this lesson. You don't need to run these scripts to complete this lesson as the Embedding Index is provided for you.
The scripts perform the following operations:
- The transcript for each YouTube video in the AI Show playlist is downloaded.
- Using OpenAI Functions, an attempt is made to extract the speaker name from the first 3 minutes of the YouTube transcript. The speaker name for each video is stored in the Embedding Index named
embedding_index_3m.json. - The transcript text is then chunked into 3 minute text segments. The segment includes about 20 words overlapping from the next segment to ensure that the Embedding for the segment is not cut off and to provide better search context.
- Each text segment is then passed to the OpenAI Chat API to summarize the text into 60 words. The summary is also stored in the Embedding Index
embedding_index_3m.json. - Finally, the segment text is passed to the OpenAI Embedding API. The Embedding API returns a vector of 1536 numbers that represent the semantic meaning of the segment. The segment along with the OpenAI Embedding vector is stored in an Embedding Index
embedding_index_3m.json.
For lesson simplicity, the Embedding Index is stored in a JSON file named embedding_index_3m.json and loaded into a Pandas DataFrame. However, in production, the Embedding Index would be stored in a vector database such as Azure Cognitive Search, Redis, Pinecone, Weaviate, to name but a few.
We've learned about text embeddings, the next step is to learn how to use text embeddings to search for data and in particular find the most similar embeddings to a given query using cosine similarity.
Cosine similarity is a measure of similarity between two vectors, you'll also hear this referred to as nearest neighbor search. To perform a cosine similarity search you need to vectorize for query text using the OpenAI Embedding API. Then calculate the cosine similarity between the query vector and each vector in the Embedding Index. Remember, the Embedding Index has a vector for each YouTube transcript text segment. Finally, sort the results by cosine similarity and the text segments with the highest cosine similarity are the most similar to the query.
From a mathematic perspective, cosine similarity measures the cosine of the angle between two vectors projected in a multidimensional space. This measurement is beneficial, because if two documents are far apart by Euclidean distance because of size, they could still have a smaller angle between them and therefore higher cosine similarity. For more information about cosine similarity equations, see Cosine similarity.
Next, we're going to learn how to build a search application using Embeddings. The search application will allow students to search for a video by typing a question. The search application will return a list of videos that are relevant to the question. The search application will also return a link to the place in the video where the answer to the question is located.
This solution was built and tested on Windows 11, macOS, and Ubuntu 22.04 using Python 3.10 or later. You can download Python from python.org.
We introduced our startup at the beginning of this lesson. Now it's time to enable the students to build a search application for their assessments.
In this assignment, you will create the Azure OpenAI Services that will be used to build the search application. You will create the following Azure OpenAI Services. You'll need an Azure subscription to complete this assignment.
- Sign in to the Azure portal.
- Select the Cloud Shell icon in the upper-right corner of the Azure portal.
- Select Bash for the environment type.
For these instructions, we're using the resource group named "semantic-video-search" in East US. You can change the name of the resource group, but when changing the location for the resources, check the model availability table.
az group create --name semantic-video-search --location eastusFrom the Azure Cloud Shell, run the following command to create an Azure OpenAI Service resource.
az cognitiveservices account create --name semantic-video-openai --resource-group semantic-video-search \
--location eastus --kind OpenAI --sku s0From the Azure Cloud Shell, run the following commands to get the endpoint and keys for the Azure OpenAI Service resource.
az cognitiveservices account show --name semantic-video-openai \
--resource-group semantic-video-search | jq -r .properties.endpoint
az cognitiveservices account keys list --name semantic-video-openai \
--resource-group semantic-video-search | jq -r .key1From the Azure Cloud Shell, run the following command to deploy the OpenAI Embedding model.
az cognitiveservices account deployment create \
--name semantic-video-openai \
--resource-group semantic-video-search \
--deployment-name text-embedding-ada-002 \
--model-name text-embedding-ada-002 \
--model-version "2" \
--model-format OpenAI \
--sku-capacity 100 --sku-name "Standard"Open the solution notebook in GitHub Codespaces and follow the instructions in the Jupyter Notebook.
When you run the notebook, you'll be prompted to enter a query. The input box will look like this:
After completing this lesson, check out our Generative AI Learning collection to continue leveling up your Generative AI knowledge!
Head over to Lesson 9 where we will look at how to build image generation applications!