Hi all,
How the predictive routing outcome probabilities actually update after a model re-train? We’ve just finished a new training run - added a bunch more agent skill data - but the probabilities aren’t changing in the Architect flow. It’s still using the old weights. I think.
The flow has a Predictive Route Step configured, using the “Lead Scoring” model. The model itself shows “Active” in the ML section. We ran an A/B test last month with similar model changes, and the lift was around 7% on contact conversion, so it should make a difference. The documentation mentions something about cache invalidation, but it doesn’t say how long that takes.
I checked a community post about this - someone had a similar thing with a custom model, but their solution was to restart the routing engine. That sounds… drastic. And it feels like there’s a better way.
Here’s what we’ve done so far:
- Model retrained with latest agent skill data (v2.0)
- Model status confirmed as “Active” in Genesys Cloud
- Flow published with the updated model
- Cache cleared in the Architect flow designer - the “Clear Cache” option
- Checked the
GET /api/v2/routing/predictors/{predictorId}/models endpoint and it shows the new version, but the response doesn’t show when it’s going to be applied.
- Waited 30 minutes.
The outcome probabilities stay the same - Agent 1 still has 0.8, Agent 2 has 0.2, even with the new skill weights. Is there something I’m missing? Or is that API endpoint just not showing the correct status? And does the ML model actually require a restart to apply the new weights?
i think it’s not just the model being active - sometimes the Architect flow doesn’t pick up the new model right away. it’s kinda weird.
we had this happen a couple weeks ago. the model showed active, but the predictive routing step was still using the old probabilities. i think it has to do with caching? or maybe something with the control engine? i’m not totally sure.
what we ended up doing was going into the Predictive Route step itself and just…re-saving it. like, open it up, don’t change anything, just click “Save”. it sounds dumb, but it forced it to re-load the latest model. it took maybe five minutes after that before the outcome probabilities updated.
quick one - are you sure you’re looking at the correct outcome probabilities in the flow? it’s easy to accidentally look at the wrong one if you have multiple predictive route steps. from what i’ve seen, the outcome probabilities displayed in the Architect flow are per-step, not globally.
also, do you have any data actions happening right before the predictive routing step that might be interfering? we had one that was setting a weird session variable and it was messing things up. just throwing that out there.
2 Likes
yes, the flow is in production - sorry i forget to mention. we have two flows, one for A/B test and one for live traffic. this is happening in the live flow, the test flow is ok. we are on Genesys Cloud, so maybe is something with the region?
GET /api/v2/routing/predictors/{predictorId}/models
The earlier reply is spot on. If the model is active but the flow’s still acting up, check the features list via the API to see if the weights actually shifted. It’s a total nightmare for the customer experience when the routing logic stays stale despite a re-train.
2 Likes
that’s a weird one. it sounds like the model is active but the cache in the live flow is just sticking to the old version. we’ve seen similar lag with other routing updates where the production version doesn’t refresh as fast as the test flow.
if the API call mentioned in the earlier reply doesn’t show the weight shifts, try hitting the features endpoint for that specific model id to see if the data actually updated on the backend.
GET /api/v2/routing/predictors/{predictorId}/models/{modelId}/features
if the features look right there but the flow is still acting funky, a common workaround is to just make a tiny, meaningless change to the Architect flow- like moving a block or changing a variable name- and then publish it again. that usually forces the flow to re-sync with the current model state and clears out whatever’s hanging on to those old probabilities.