trying to catch these agent errors is such a mess because they look so much like actual good answers. it's not just one thing either. you've got factual stuff where the model makes up stats, then there's context drift and even self-contradictions when it forgets what it said two sentences ago.
>all hallucinations are hard to catch because they usually look like normal answersthe idea that we need a layered system is fine but it sounds like a massive headache for production. i don't see how checking for groundedness helps if the source material itself is garbage. feels like we're just adding more latency and more points of failure to the workflow.
good luck with the compute costs are you guys actually using multiple checks in your pipelines yet?
article:
https://blog.n8n.io/hallucination-detection-methods-for-ai-agent-workflows/