What's the process for updating documents in Google Vertex AI Agent Builder's Recommendation model?

Other Questions AI Agent
1

I’m working with Google Vertex AI Agent Builder and I’ve set up a Recommendations model for Media content. The model is trained and can make predictions, but I’m confused about how document updates work.

Here’s what’s puzzling me:

  1. I deleted a document (ID 47941) from my data store.
  2. The document is gone from the Documents tab and API calls.
  3. But the model still returns this deleted document in predictions.

It seems like the model isn’t syncing with the data store in real-time. I’ve looked through the docs but can’t find clear info on this.

Questions:

  • How does the model connect to the data store?
  • Do I need to retrain the model to see updated data?
  • Is there a way to re-index the data store?
  • Should deleted documents be automatically filtered out?

I’ve tried purging the data store, but the model keeps returning old data. I expected the model to only use existing documents, but that’s not happening.

Has anyone else faced this issue or know how to resolve it?

2

hey adventuroushiker17, i’ve run into similar issues. from what i understand, the model doesn’t sync automatically with the data store. you probably need to retrain the model after making changes to your documents. it’s a bit annoying, but that’s how it seems to work. maybe try retraining and see if that fixes it?

3

I’ve been working with Vertex AI Agent Builder for a while now, and I can shed some light on your situation. The model doesn’t automatically update when you make changes to your data store. It’s a common misconception, but the model and data store are essentially separate entities.

In my experience, you need to retrain the model after making significant changes to your document set. This includes deletions, additions, or major updates. It’s not ideal for real-time operations, but it ensures the model’s predictions are based on the most current data.

As for deleted documents, they should ideally be filtered out, but I’ve found that this doesn’t always happen automatically. You might need to implement a post-processing step in your application logic to remove any recommendations for documents that no longer exist in your data store.

Reindexing the data store itself won’t solve the problem. The key is to retrain the model. It’s a bit time-consuming, but it’s the most reliable way to ensure your model is up-to-date with your current document set.