Question about UID assignment logic in rollout generation (do_search vs. not do_search)

[passionate11]

Hi, thank you for your very interesting work!
While reading your code, I came across something that confused me and I’d like to ask for clarification.

In the rollout generation logic, when do_search is False, the code assigns UIDs as:

batch.non_tensor_batch["uid"] = np.array(
    [str(uuid.uuid4()) for _ in range(len(batch.batch))],
    dtype=object,
)

However, when do_search is True, it instead uses:

batch.non_tensor_batch["uid"] = batch.non_tensor_batch["index"].copy()

I’m curious about the design rationale behind using two different UID assignment strategies here.
Why do we generate random UUIDs in the non-search case, but reuse the existing index field when do_search is enabled?

Any insight into this difference would be greatly appreciated — thank you again for sharing your work!