How to Summarize YouTube Channel Videos with a Knowledge Graph and AI

In this video, I demonstrate how you can use infranodus.com to summarize KZbin videos of any channel using a knowledge graph and AI.
We use the example of the Y Combinator (‪@ycombinator‬) popular channel for startup founders. First, I create a knowledge graph representation of this content, which shows me the main ideas they're talking about in their videos and the clusters of key topics. Then I use the built-in AI to generate a general summary and a research question. Then I zoom onto specific topics to generate summaries of the content within.
I can even select the specific parts of the videos that are relevant to me and see where they're talking about particular concepts that I find interesting.
Try it out at infranodus.com
Timecodes:
00:00:58 - The KZbin channel Y Combinator focuses on business techniques for startup founders and startup culture.
00:02:23 - InfraNodus visualizes text as a Knowledge Graph, showing how words relate semantically and contextually.
00:03:56 - Exploring how software companies leverage time perception and product-market fit for high valuations.
00:05:17 - Filtering out common words like "people" and "things" to reveal more insightful content beneath.
00:07:34 - Analyzing the main ideas in clusters to understand the channel's key topics and discussion areas.
00:08:44 - Summarizing specific topics and highlighting the importance of feedback and prioritization in business.
00:09:53 - Identifying gaps in discourse to generate interesting questions and explore new ideas.
00:10:58 - How making decisions and building innovative structures can shape real-world change and enhance our lives.

[@noduslabs]

Timecodes: 00:00:58 - The KZbin channel Y Combinator focuses on business techniques for startup founders and startup culture. 00:02:23 - InfraNodus visualizes text as a Knowledge Graph, showing how words relate semantically and contextually. 00:03:56 - Exploring how software companies leverage time perception and product-market fit for high valuations. 00:05:17 - Filtering out common words like "people" and "things" to reveal more insightful content beneath. 00:07:34 - Analyzing the main ideas in clusters to understand the channel's key topics and discussion areas. 00:08:44 - Summarizing specific topics and highlighting the importance of feedback and prioritization in business. 00:09:53 - Identifying gaps in discourse to generate interesting questions and explore new ideas. 00:10:58 - How making decisions and building innovative structures can shape real-world change and enhance our lives.

[@Plush07]

Great traction of the product! Superb!

[@rationalityrules111]

Very good feature! Niceee!

[@mandelafoggie9359]

Thanks , I spent 30 min looking for this !

[@noduslabs]

Great! Have you tried using it now?

[@mandelafoggie9359]

@@noduslabs yes , need to watch the Intro videos! Great product

[@Redranddd]

I had this idea in the past good to know it's now possible

[@MartinBlaha]

What a tool! 🔥

[@JaMarrJ]

Great overview I will definitely test this out first thing in the morning reaching out to an email.

[@user-vh3vf7hc9s]

wow. This is soo cool. I can see this merging with generative ui apps to provide a seamless filtering system

[@noduslabs]

How would you see it working? Please, tell me more...

[@settingfeedback321]

this is really good

[@Salionca]

Good video

[@railzip]

With each video upload I get closer to paying for the product. I think I’ll be deciding this weekend.

[@thiagonamene]

Same man, same 😂😂😂

[@mariusorani]

I have a very useful usecase in my mind… i think people did not realize how valuable this thing is

[@noduslabs]

Interesting! What's your use case? Channel optimization? Regarding the people - yes, I agree with you :)

[@jarad4621]

Had the same thought, most won't see the real value, this is pretty amazing if you think a little

[@smnomad9276]

You need to dim out the colors of the branches between nodes a little bit, they are way too bright. Only keep the big dots of the nodes bright, but the connection branches need to be less visible and a little transparent.

[@noduslabs]

Ok, sir!

[@SevenOneTv.]

Very cool....the title of your video just brought up a good use case for me to use this app. I publish books and it makes sense that I analyse people's reviews and pain points to produce better tailored books. Can you do a video on that?

[@noduslabs]

Sure! InfraNodus can already analyze comments to KZbin videos and also Amazon reviews. You can also get your source data from anywhere else and upload it as a CSV file. Is that the data you'd like to use?

[@SevenOneTv.]

​@@noduslabsQuora and Reddit sources

[@desarrollojava]

👏👏👏👏👏👏👏👏👏

[@2010Sisko]

I am absolutely amazed by this tool that you’ve created. I definitely will be subscribing. I do have a question though. Would it be possible for me to analyze different videos from different channels that are similar in topic but from different creators for me that would be an amazing case.

[@noduslabs]

Yes, in fact we're working on a feature where you can analyze a playlist or just the Google search results for a certain topic. What would you prefer? Both?

[@2010Sisko]

@@noduslabs I think I’d prefer both but my first preference if possible would be the playlist option. It just seems to be more of a DIRECT curated connection that I could then mine and find interesting links with. I’m really enjoying your tool. I’ve already signed up and I’m gonna keep checking it out looks amazing.

[@noduslabs]

@@2010Sisko great! Will implement it soon! In the meanwhile try out other import tools also and the extension. May be very interesting for exploring other web content and own ideas. 🙏🏼

[@2010Sisko]

@@noduslabs Hello again. ABSOLUTELY loving your tool the further I go the more I enjoy it :-) Given yesterday's Openai recent announcement concerning GPT-4o (Omega) what are your plans in terms of GPT-3.5 vs GPT-4 for regular users vs pro users. Cheers.

[@herbiekruger]

Is there a way to compensate for the in-video ads made by the content creator, they are skewing the results in the graph. As an example [using just 1 video] I activated infranodus [ext] and of the 6 gaps it identified the first 5 referenced the in-video ad text against what the video was about. The 6th was fine but, by then, 5 calls to the server had been made. those 5 gaps effectively produced a nonsense result within the context of the actual video message

[@noduslabs]

Good point. You can exclude certain topics and words from analysis, so once you have it showing on the graph, just select the topic / terms you don't want and remove them and it should work!

[@SALAVEY13]

Very cool! I would love to use it for analyzing a playlist of videos that were saved during particular period of time for automatic creation of events history that happened during that time. like history of war - i have a playlist for every perior, for every day, from dozen experts - intersection of their viewpoints is awesome, it's currently only in my head - it really looks like graph))

[@noduslabs]

Great! Can introduce playlist analysis feature too!

[@SALAVEY13]

@@noduslabs Nice! Unfortunately i'm currently inside sunctioned russia and don't have valid bank card, so i can't use your awesome tool myself, yet, but i'm learning by watching, thank you! My idea is to make a questionnaire for a set of top topics from all discussions of dosen experts - to align youtubers themselves - some of then are obviously have some misconceptions themselves - i think it would be VISIBLE on a graph. It is useful tool to fight desinformation

[@desarrollojava]

Wow! I wonder if we can implement real time queries such as phrases.

[@noduslabs]

Sure you can. For now, you can visualize the titles and descriptions of the videos found so you understand the current informational supply. We're working on adding subtitles for search results as well.

[@jakobdango170]

Hey i have done some reasearch. I cannot access the repo u have used for the project named neo4j

[@noduslabs]

Sorry, but didn't really understand what you mean.

[@andrewjiang824]

Do the videos need to have close captions? Many vids don’t have transcripts

[@noduslabs]

It will use automatic captions but if there are no captions then your only way is to get it transcribed using Whisper for instance and then import the transcript into InfraNodus (as a file). That’d work.

[@mambetmambetov8531]

Will nodus have same features but in other languages for example Russian to summarize content in Russian KZbin channel ?

[@noduslabs]

Yes it does.

[@jarad4621]

Amazing could i run it on topics on yt or all post on a niche subreddit, if you cant scrape can you connect a db or upload data

[@noduslabs]

You could run it on KZbin comments or on a reddit, yes. You can scrape the comments or get the reddit's RSS feed and import it in InfraNodus.

[@OleksandrSelivansky]

Hi, I got error when try to get video analysis today Channel link just do nothing even after 10-15 minutes Please share solution

[@noduslabs]

It could be because of the language, I see from your username you might be searching for something other than English? Maybe that is the case?

[@samiennaciri8008]

Hello! Is there any AI tool to analyse movie sequences? Thank you

[@noduslabs]

Hey you could use this one for it

[@samiennaciri8008]

@@noduslabs Thank you !

[@gambaweb]

Mix in this graph with vectorDB for knowledge search

[@iva1389]

how dou you mean to load the data every time you click on new video lol easiest thing is to mention the vector db in yt comment section and feel like you'd thrown the epic solution nobody thought of before

[@noduslabs]

You think years of AI and there's no vector representations underneath all this?

[@noahverner5153]

Can it summarize my PRIVATE youtube videos? if not, can I upload a video file to your site and summarize what it contains?

[@noduslabs]

You can set your video to be invisible to public but accessible via a URL. Then you open that video, get the contents, and then make it private again. Then it'd work. Or you can of course upload the subtitles file and analyze it that way as well.

[@flor.7797]

Do one on Taxonomy creation where the resulting knowledge graph is not just irrelevant word fragments

[@noduslabs]

These are not irrelevant word fragments. Are the words that you used to write this comment also irrelevant fragments? The graph shows how they are connected, this helps you build a good visual representation of the main topics and use the underlying structure to query the AI. Btw you can also choose to perform entity detection if you want.

[@flor.7797]

@@noduslabs sorry this came off wrong. I’m still very interested in creating a solid taxonomy based on a knowledge graph. I guess human in the loop + more data could be the answer . Would love to hear your thoughts on this.

[@user-wr4yl7tx3w]

but so far, i haven't found a good way to leverage KG for RAG

[@noduslabs]

What is your use case? I find it works really well to generate ideas. Precise factual retrieval is of course more difficult.

[@Redranddd]

The channel I wanna use have 2.8k videos

[@noduslabs]

that will probably be a bit of an overkill at this point. what's the channel?

[@Redranddd]

@@noduslabs Thomas delauer

[@noduslabs]

@@Redranddd Would it be suitable for you if you can filter the videos by keywords first? For instance, only his videos on keto diet?

[@Redranddd]

@@noduslabs I want it to tell me what do you think about pineapple? What's do you think about eggs? Etc so I don't think keywords would help

[@iva1389]

that's far from knowledge graphs, what you are showing it's just visuals of most common words randomly represented in kind of vector space. it's not even the proper vector space because you have a single layer dimension with keywords that appear most often . thats far from any KG since it holds 0 informational value that is anyhow correlated in meaningful way.

[@noduslabs]

Your comment only shows that you don't know anything about the topic you're trying to discuss. I don't even know where to start, to be honest. First, the alignment is not random. We use force-layout algorithm, which is frequently used in network science to represent clusters of related nodes together. Then we use community detection algorithms to identify community clusters. Those are fed to an AI model to produce very precise topical cluster names that are better than any other tool that you can get out there. Then the underlying graph structure of text is also fed to the model to get much more precise prompts which take the relations between the words into account. According to you, however, anything that represents clustering would have 0 informational value, so I wonder how much you really know about the field to make claims like this one...

[@NashBrooklyn]

frankly what you are showing is nothing new - but to make it new is to show bulk percentage usage of every word you filtered out - cause once you finish filtering everything out - there will be nothing new left -

[@noduslabs]

I wonder where else you've seen something like this before? A live text network data visualization of the whole channel's content or a KZbin video? Oh, wait, you mean it's not new because you've seen it 5 years ago on Medium - towardsdatascience.com/capture-the-essence-of-any-video-visualize-its-subtitles-as-a-graph-5ce565d1820c - but surprise! It was done using this same tool :)

[@patrickbaumann8670]

Thank you, looks amazing! Is there currently a way to implement external LLMs? for example I have my own lama.ccp server that is reachable publicly, with REST-API and Authorization. If i could use my own AI Model, that would pretty much mean no cost for you(normally using openai tokens), and if possible unlimited chat interactions for the user! If that feature is available, I will become a User rn. if not, no worries, i love your work and your videos are exactly targeted for the right audience! Not long and your channel will go viral. I assure you. BR Patrick

[@noduslabs]

Hey thanks, but the main cost is not OpenAI but developing this tool. However yes, we will offer other models access as many people request that. At the moment, you can already get some of the graphs data using the API.