Strategies for embedding json observations?

The observations for an environment I’m working with are large, nested json objects. Right now it’s infeasible to flatten them into consistent vectors. My initial thought is to use a text embedding model to convert them to vectors. What other approaches have people used when they encounter problems like this?