YouTube Video and Audio Search

This repository performs searches over content that appears in YouTube videos. This demo uses Marqo's YouTube channel to search for relevant video clips. It will then direct you to that specific timestamp on YouTube.

Note, this demo uses already chunked data that is hosted online. If you want to do this for your own YouTube videos, we have provided the code that downloads a YouTube playlist and chunks the data for you. It is then up to you to host these files somewhere. We recommend using AWS S3 buckets.

Step 1: Clone this Repository

You can do this with:

git clone https://github.com/marqo-ai/youtube-search

Then, load all the dependencies needed for this project.

python3 -m venv venv
source venv/bin/activate   # For Windows, run  venv\Scripts\activate
pip3 install -r requirements.txt

Step 2: (Optional) Download Your YouTube Videos

This step is optional. If you are happy using Marqo's youtube videos to build this project then you can jump straight to Step 4.

If you want to download your own YouTube videos, keep reading. We encourage you to create a small playlist of your videos to begin with, depending on their size. For this demo, we have 10 videos that vary in size from 3 minutes to 24 minutes. Note, if you want to test this out with just one YouTube video, you can also just specify the individual YouTube link. To download your YouTube playlist, run the following code:

python3 preprocessing/download_playlist.py

Enter your YouTube playlist when prompted to do so.

This will first begin downloading your YouTube video(s) and placing them into the folder processing/raw_youtube_videos. Once all videos have been downloaded to this folder, it will begin chunking them into 20 second splits and placing these into the processing/video_chunks. This is needed when loading the video files into Marqo ready to perform video and audio searches.

Step 3: (Optional) Hosting Your YouTube Videos

This step is optional. If you are happy using Marqo's youtube videos to build this project then you can jump straight to Step 4.

To load these video files into Marqo, they need to be hosted at a URL address so Marqo can successfully retrieve them. For this demo, we hosted our YouTuube videos using an AWS S3 Bucket. Feel free to use whichever hosting platform you'd like, just ensure the URL ends in .mp4 and is publicly accessible so Marqo can access and retrieve it.

Once you have hosted your YouTube videos, we need to create the data CSV file that will be fed into Marqo. This keeps track of all URLs. First, obtain the starting URL to your videos. In our example, our videos are hosted at https://marqo-tutorial-public.s3.us-west-2.amazonaws.com/youtube-search-demo/marqo-youtube-videos-chunked/ and so, we append each video title in video_chunks to this URL and store them in a CSV, for each video.

To create this data CSV, run the following:

python3 preprocessing/generate_urls.py

This will create a new CSV in the data folder called video_urls.csv which will contain all of the URLs for your chunked videos. It will also create another file called youtube_ids.csv which will contain the corresponding YouTube video IDs. Note: if you do this for Marqo's YouTube playlist, the IDs will be slightly different as the files were preprocessed and loaded using a different playlist.

For peace of mind, we encourage you to check that all of the URLs in the video_urls.csv file are valid. This will avoid Marqo throwing any errors if it cannot access them. Run this script to check your URLs:

python3 data/test_urls.py

The terminal will only populate with URLs if an error has occurred.

Step 4: Set up Marqo Index

Now we have all of our data, we can set up a Marqo index and begin adding these video files to our index.

Obtain API Key

First, you will need a Marqo Cloud API Key. To obtain this, visit this article. Once you have your API Key, place it inside a .env file such that:

MARQO_API_KEY = "XXXXXXXX"   # Visit https://www.marqo.ai/blog/finding-my-marqo-api-key 

Create Marqo Index

Next, we can create our index:

python3 marqo/create_index.py

This may take a few minutes to create. You can see the status of your index in the Marqo Cloud Console:

Your terminal will also begin populating:

2024-12-18 16:13:44,437 logger:'marqo' INFO Current index status: IndexStatus.CREATING
2024-12-18 16:13:55,621 logger:'marqo' INFO Current index status: IndexStatus.CREATING
2024-12-18 16:14:06,816 logger:'marqo' INFO Current index status: IndexStatus.CREATING
...

Add Documents to Marqo Index

Once your index is successfully created, you can begin adding documents to your index. If you index creation failed for whatever reason, reach out to us on our Slack Community where a member of our team can help.

To add documents to your index, run:

python3 marqo/add_documents.py

This will read all of the video URLs in data/video_urls.csv and begin uploading them to Marqo Cloud. We have chosen a batch size of 1 so that we can see logs for each video URL added. Your terminal will look like:

2024-12-18 17:16:57,063 logger:'marqo' INFO     add_documents batch 0: took 39.488s for Marqo to process & index 1 docs. Roundtrip time: 39.698s.
2024-12-18 17:17:27,544 logger:'marqo' INFO     add_documents batch 1: took 30.276s for Marqo to process & index 1 docs. Roundtrip time: 30.480s.
2024-12-18 17:18:00,310 logger:'marqo' INFO     add_documents batch 2: took 32.615s for Marqo to process & index 1 docs. Roundtrip time: 32.766s.

(Optional) List All Documents / Get Stats

While documents are being added to your index, I always find it helpful to track the statistics of the index and to list all the documents with their IDs (these IDs are automatically generated by Marqo).

To list all the documents currently in your index, run:

python3 marqo/list_all_documents.py

This will show you all the documents in your index. This will return the video_field and ID of the document. Note, this ID is created automatically by Marqo. It'll look similar to the following:

{'video_field': 'https://marqo-tutorial-public.s3.us-west-2.amazonaws.com/youtube-search-demo/marqo-youtube-videos-chunked/video2_1.mp4', 'ID': '0aa1f2b8-27ee-4222-83e2-4f0881a35a2b'}
{'video_field': 'https://marqo-tutorial-public.s3.us-west-2.amazonaws.com/youtube-search-demo/marqo-youtube-videos-chunked/video2_2.mp4', 'ID': '792e7c0b-c8c9-4203-99b9-9915b711ef8d'}

To obtain index statistics, run:

python3 marqo/get_stats.py

This will show you a similar output to:

{'numberOfDocuments': 19, 'numberOfVectors': 37, 'backend': {'memoryUsedPercentage': 2.13577924416, 'storageUsedPercentage': 1.24770521372}}

If you wish to delete documents from your index, inspect the marqo/delete_documents.py script to do so.

Step 5: Begin Searching

You can begin searching even while your documents are being added. Note, the more documents in your index, the better the search results will be.

For searching, we use Streamlit as the UI and use the search feature in Marqo. If you want to experiment with just Marqo search without the UI, you can run python3 marqo/test_search.py which will perform a simple search.

To run the UI,

streamlit run app.py

This will launch the user interface ready for you to perform searches!

Step 6: Clean Up

If you follow the steps in this guide, you will create an index with GPU inference and a basic storage shard. This index will cost $1.03 per hour. When you are done with the index you can delete it with the following code:

python3 marqo/delete_index.py

If you do not delete your index you will continue to be charged for it.

Questions? Contact Us!

If you have any questions about this search demo or about Marqo's capabilities, you can:

• Join Our Slack Community
• Book a Demo