Indexing could be done on the client and sorted on there?

I would recommend relays to dump to parquet files on cloudflare R2 with 0 egress charges and clients use DuckDB to query.

Indexes plus an LLM that was built on realtime data like Grok is.