How to Summarize YouTube Videos in ChatGPT using Mantium Plugin Wizard

Introduction

Do you have long YouTube videos you need to watch, but would you prefer interacting with the text to ask questions about key highlights or summaries for a comprehensive overview of the content? Or perhaps you want to use the content to create new material, and ChatGPT doesn't have the necessary context?

Well, this tutorial will help you solve that by showing you how to build plugins that you can use to interact with transcripts from YouTube videos and save time from watching the video.

For this tutorial, we are going to interact with the Microsoft Build 2023 Keynote Speech . This video is over 2 hours long, and might take sometime to consume, hence why our solution is timely.

Prerequisites

• API Keys for OpenAI.
• Sign up for ChatGPT Developer Access, if you don't have access.

Create Datasets in Mantium

While we haven't officially launched a Mantium YouTube connector yet, we will work with the transcripts of the YouTube video.

If you don't have access to the transcript of the video, another approach will be to use the Mantium File connector to import the mp3 file, and use the Transcribe to Audio transform to generate the transcripts of the video. Read this tutorial here to learn more.

For this tutorial, I have prepared a word document with the full transcripts. Download the file, and let's use it to create a dataset in Mantium.

Upload the Word Document to Mantium

1. Click Data Source on the left navigation bar to go to the Data Sources section.
2. On the top right corner, select Add Data Source.
3. From the list of Data Sources, select the Word Documents Data Connector.
4. Provide the necessary details to label the Data Source and wait for the job process to complete.
5. Complete the data upload process by uploading the Word document (.docx).
6. Click the Finish and Sync button to finalize the setup and synchronize the data.

Create a Dataset in Mantium

1. After the job is done, click on Create Custom Dataset.
2. Provide a name for the dataset and click on Save.

Once the dataset is created, navigate to the Transform section to add transformations to the dataset.

Add Transformations

Split Text Transformation

After creating the custom dataset, navigate to the "Transform" section to add a couple of transformations. For this tutorial, we're going to be working with the "Split Text" transformation.

The "Split Text" transformation is used to split the content to get it ready to generate embeddings on that so that it doesn't go above the OpenAI recommendation.

Here's how to set it up:

1. Click on Transform and select Split Text from the dropdown menu.
2. In the Source Column field, add the content column. This is the column that contains the text data to be split.
3. In the Destination Column field, type "segmented_text". This will be the new column that will hold the segmented text.
4. Set "Split By" to "word". This means the text will be split at every word.
5. In the "Split Length" field, type 500. This means the text will be split every 500 words, creating an additional row.
6. Leave the remaining configuration as default.
7. Click on the Plus sign(+) to add the next transform. The job will run and split the text in the "content" column into smaller chunks of 500 words each, storing the result in a new column called "segmented_text".

Generate Embeddings

We are going to generate numerical representations(embeddings) of the segmented_text column. This embeddings is what we are going to ship to Redis (the Vector Database) for ChatGPT to query for response.

1. Select the Generate Embeddings from the list of transforms.
2. Configure the Source Column parameter by selecting the column containing the segmented_text .
3. Configure the Destination Column parameter by specifying the new name for the column that will hold the embeddings.
4. Configure the Embedding model parameter by selecting the LLM model to use for the embeddings.
5. Enter the Max Token Length, you can leave the default value - 8191
6. Configure the Credential ID parameter by selecting the appropriate credential from the list of available credentials in your Mantium account.

Now we have processed dataset that is ready for use in ChatGPT!

Create a Plugin

🚧

Quick Warning

• If you select the Standard option and have previously created a split_content column, ensure to pick this same split_content column in subsequent steps rather than the original text column. This will prevent the unnecessary expansion of your dataset, ultimately keeping your OpenAI usage costs in check.
• To prepare your dataset, Select the Advanced option if you have Embeddings already or you have completed the steps above.

1. Navigate to the Plugin section and click on New Plugin.
2. Select Existing Dataset and choose the dataset you just created.
3. Select an existing OpenAI credential or create a new one.
4. Select the Advanced, because we do have the embeddings column.

5. Provide the name of the columns in the field:
   • Content Column - segmented_text
   • Embedding Column - embeddings

6. Click on Create Destination, and Mantium will automatically set up a redis destination for you.
7. Provide a name and description for the plugin (see image below), then click on Create.

Congratulations, you have completed the process of creating your plugin in Mantium. Now, the next step is to add the newly created plugin in ChatGPT using the credentials.

Set Up the Plugin in ChatGPT

1. Navigate to your ChatGPT account and click on the Plugins dropdown, then select Plugin Store.
2. Click on Develop Your Own Plugin.

3. Paste the Plugin URL from Mantium into the Enter your website domain field and click on Find Manifest File.
4. Click on Next, then click on Install for Me.
5. Enter your access token from Mantium to install the plugin.

Test the Plugin

Now, let's interact with the Plugin in ChatGPT. Here are some example prompts:

Prompt 1

Tell me about Microsoft Fabric, and how can I use it to streamline my work as a data analyst ?

Result from Prompt 1

Prompt 2

Using Microsoft Build QA plugin, how can I use GitHub Copilot Chat as described by Thomas Dohmke?

Result from Prompt 2

That's it! You've successfully set up a plugin to to interact with the Microsoft Build 2023 Keynote Speech, which is over 2 hours long.