do-schema | DataObjects Schema Reflection | Reflection library

The Elasticsearch connector doesn't support this. You can update documents in-place but need to send the full document, not a delta for appending which I think it what you're after.

You should be building a reference to the document, then creating its contents using set() on that reference instead of add() (since add always creates a new random document id). This is detailed in the documentation, which I strongly recommend studuing in order to understand how Firestore works. You will use that information a lot.

To create a document with a known ID (in the variable thread_id) in a top level document:

Your dictionary is that complex then you're doing this all wrong. Consider creating classes and store an object of that class in the dictionary. Those classes can also hold other objects of other classes. This way you'll avoid the nesting of dictionaries. Create functions within classes to validate its data.

I'd use EXISTS for that.

Is this the right approach to take or is there a better way to relate the tables here?

Your approach seems OK at the DB layer, except that if users_x_conversations serves only as a join table (i.e. if there are no extra properties associated with the (user, conversation) associations represented within), then I would use (conversation_id, user_id) as its PK instead of giving it a surrogate key. If you don't do that, then you should at least put a uniqueness constraint on that pair.

What I'm stuck on now is: how should I model my [users_x_conversations] table at the JPA layer?

I take you to be asking whether you should model that table as an entity. If you insist on giving it a surrogate key as you have done, then that implies "yes". But as I already discussed, I don't think that's needful. Nor much useful, for that matter. I would recommend instead modeling a direct many-to-many relationship between Conversation and User entities, with this table (less its id column) serving as the join table: