Editorial AI in AEM: Giving Editors Control Over LLM-Generated Image Descriptions
Most enterprise DAMs contain thousands of images with missing or broken dc:description metadata—an SEO and accessibility liability that no team can fix manually. This session shows how to solve it properly: an open-source AEM workflow that integrates a connection to a remote LLMs (via plain HTTP, no SDK, no bundle conflicts) to generate contextually accurate alt text at scale.
The intelligence isn't hardcoded into the system but stored as plain-text prompt files inside the DAM itself. That single design decision makes this architecture worth understanding: editors iterate on prompts like any other content. Consequently, a product image gets packaging-focused descriptions, a campaign asset gets CTA-aware copy, and an app screenshot gets feature-level descriptions, all from the same workflow without a single redeployment.
wolf
Why go to a custom LLM at all if you have started with Azure AI? It does have Image Analysis as one of the pre-built services (and it does captioning, tagging and all the rest that you need here).
wolf
Is the LLM there just to allow custom prompting?
Jose Maria
Azure Vision services is traditional machine learning, like classification, clustering... but not natural language generation.
maroof
How are you validating the LLM generated content for the images?
minich
Trust and Hope
Jose Maria
The prompt includes guardrails and guidelines, but the last line is human validation.
Yegor Kozlov
Adobe has recently released Brand Aware Metadata which allows mapping metadata fields to AI prompts. Did you evaluate it?
Jose Maria
Not yet. We need to get access first, but our customer is still on-prem.
Simon
How many hours did it take to implement this solution? ;)
Jose Maria
I think was around 40
Robert
Are you taking the surounding content based on the actual image usage into account? So e.g. if the image is used on a page with a headline. Do you add this headline to the context handed over to the LLM and, if so, what is the mechanic behind?
Jose Maria
Short answer is no. Images are properly stored in the dam according to their purpose, so that's all the context we need.
Georg Henzler
You used language-specific prompts - could you have used one english version that asks to detect the language from the path (or something else) and then to output in the respective language?
Jose Maria
We thought of that, but many times it's not just a matter of translation. Sometimes the customer needs to adapt the message to the language, so we thought that adding this flexibility was a good idea.
KatieP
If there are dozens of campaigns, do I need a workflow per campaign?
Jose Maria
Each campaign has a different focus. We could come with a single workflow that takes the campaign as a parameter, but how would that be different?
Michael
The 100$ quote does not include the cost of the development of this solution. Is it really cheaper if this is accounted for?
Jose Maria
Good question. Maybe not for the first use, but this approach scales. Manual doesn't.