If I may suggest one thing... instead of range requests on a single huge file how about splitting the file in 1-page fragments in separate files and fetching them individually?

Edgesearch does that though with Cloudflare Workers mediating searches: https://github.com/wilsonzlin/edgesearch

Uses roaring-bitmaps to index, but could also use Stavros' trick (bloom/cuckoo-filters) to further gate false-positives on-the-client: https://news.ycombinator.com/item?id=23473365