An End to Dumb Asset Titles

Introducing an effortless way to create better content titles

We get it. Part of the hard part of creating and maintaining an up-to-date content library is making sure that the content is easily findable and readable.

With the emphasis on readable, let’s break down the anatomy of a typical asset and what we’ve done to make sure content is the most easily accessed for you and your team as you organize and share those go to market assets that drive sales forward.

We do all of this with AI – and we’ve experimented with different models and discovered some surprising issues along the way that we’ve had to sort out. The end result is like magic, works seamlessly, and even some of our team didn’t even realize we had released it, so we consider that a job well done. All the benefits of an update with zero distractions.

Adding assets to a content library and a typical challenge

For this example, we’re going to zoom in on the addition of a single asset to the content library. Sure, you have tons - maybe even thousands of assets - but, yeah, we’ll zoom in on just a single marketing asset for now.

Let’s say it’s a datasheet or one-pager or a case study. Maybe you used to upload that directly to Wordpress or your site, but now you’re going to use Content Camel to organize it, tag it, and share it with the team and prospects.

Let’s say the file is called resources_customerStory_sonos_brandUpdate_jan23_letter_webReady.pdf. Because that makes total sense to the internal marketing team, right?

In the past, if we added that PDF, it would just be called “resources_customerStory_sonos_brandUpdate_jan23_letter_webReady.pdf” in the content library. Not very helpful, right?

Or, we’d extract the title if we could, but that’s based on the team manually marking that up inside the file at creation time.

But, more often, it would just be the filename. And that’s still the case if you upload it today to other tools like Google Drive, or Sharepoint / Onedrive, or Dropbox.

Introducing smart asset titles

All those underscores, version numbers, “final XXX”, “v2”, “v3”, “final final”, “final final final” – all of that is just noise and needs to go away.

So, we’ve deployed a model that takes the filename and intelligently extracts the most important parts of the filename and creates a title for you. And behind the scenes we automatically generate some tags from just the filename too.

It works really well, and it works for assets that are added one at a time, in a group, imported from Google Drive, Dropbox, etc, and those assets imported via CSV. Anytime that we generate a title now for a saved asset, it goes through our AI model and comes out the other side with a title that makes sense.

This space is moving quickly, but as of writing this, we settled on the Microsoft Wizard LLM 8 x 22B model, which is a mouthful, but it’s been working well for us. We tried out other models (including GTP-4 / GPT-4o and Llama), but the Wizard LM 2 model was the best at keeping things real when the going got touch and minimized hallucinations – “making stuff up”.

The tricky parts of AI and asset titles

Creating meaningful titles from unstructured filenames is easy when there’s some meaning baked into the titles in the first place. The models are generally pretty good at detecting date formats, including quarters and abbreviations, but they can get really tripped up when the filenames are just unique IDs or more random text. This is where most of the models that we tested fell down and invented titles or tags with labels that were simply made up. And they did that despite our best efforts to remove that behavior via our prompt and instructions.

Image if your filenames look like this, headed into your content library:

• ‘_4f7c305f-99eb-4fe2-a0ff-c1b4096becef.jpeg’
• ‘IMG_3406.JPG’
• ‘52948698_351812775674711_9218550532338089984_n.csv’

Or this surprise, when the AI wanted to break this up into multiple filenames:

• ‘Marketing Activities Directory.xlsx - Blogs Processed.csv’

Given the non-determintistic nature of the AI (that the results aren’t entirely the same every time or predictable), we’ve had to add some manual review and override capabilities to the system. We’re working on improving the model and the training data, but it’s a tough problem to solve permanently. If you run into any issues – and we’re sure that there are some out there – drop us a line, because we’d love to hear more about it.

A better content library

All together, this is an incremental and mostly invisible update that builds upon our existing best-in-class AI search capabilities and layers in something that’s immediately useful for everyone including all of your users that are searching for and deploying content.

It’s an easy win and no more dumb asset titles!

We’ll be expanding the suggested tagging capabilities and AI-powered similarity search soon as well – something we’re all excited about and would love your feedback on.