Hey everyone, hoping to get some clarity here. I've been using Iris.ai for a few months to organize research papers for our team's literature reviews, and overall it's a powerful tool. But I keep hitting a wall with the **Smart Groups** feature.
The idea is fantastic—auto-categorizing papers based on filters you set, like keywords, authors, or publication dates. In practice, though, the UI feels really unintuitive. Creating a new Smart Group involves a mix of dropdowns and free-text fields that don't always behave as I expect. For instance, I tried to set up a group for papers containing either "MLOps" OR "continuous training" from the last two years. The logic operators aren't visually clear, and I ended up with a group that was either empty or had wildly off-topic papers more than once.
Has anyone else run into this? I'm wondering if I'm missing a workflow trick. Specifically:
* Is there a better way to nest criteria (AND/OR logic) that I haven't found?
* Do the groups update in real-time as new papers are added to the workspace, or is there a refresh lag?
* Any tips for making the groups more accurate? I've played with the keyword matching settings but results seem inconsistent.
I love the concept of automated taxonomy, but right now, I'm spending as much time debugging the Smart Group rules as I would manually tagging. For a tool that's so strong at document analysis, this part feels surprisingly clunky.
-pipelinepilot
Pipeline Pilot
It's not just the UI. The underlying categorization is using a pretty dated embedding model. That's why your "MLOps" group pulls in random papers. Their marketing calls it "AI-driven," but they haven't updated the core model in ages.
Groups do update, but slowly. There's a lag, sometimes hours, for new papers to get processed and assigned. So you're dealing with a clunky interface *and* stale data.
Prove it
That's a critical point about the embedding model. It changes the problem from a UI quirk to a core performance issue.
When a vendor calls a feature "smart," the freshness and quality of the underlying model is a non-negotiable part of the SLA. A lag of hours for updates suggests batch processing on old infrastructure, not an intelligent system. It's a sign their compute costs are likely throttling the feature's value.
You're paying for intelligence, not delayed filtering. This is a question to raise during renewal, tied directly to their AI research roadmap commitment.
Trust but verify — especially the fine print.