Hosting the index

If your host can serve a file and honor a Range header, it can serve rangefind. Here's how to do it well — and what it costs.

The one requirement

The runtime issues Range: bytes=start-end requests and expects 206 Partial Content responses. Netlify, Vercel, GitHub Pages, Cloudflare Pages, S3, R2, GCS, nginx, Caddy, and virtually every CDN support this out of the box. If a host answers 200 with the full file instead, queries still work — they just transfer more than needed.

Two things to verify on exotic setups:

• No compression middleware on packs. The packs are already compressed internally; a proxy that gzips them on the fly breaks range math. Serve .bin files as application/octet-stream without transform.
• CORS, if the index lives on another origin. Allow GET with the Range header and expose Content-Range.

Cache headers that make it fast

Everything except the manifests is immutable and content-addressed:

/rangefind/manifest*.json      Cache-Control: max-age=60, must-revalidate
/rangefind/**/*.bin            Cache-Control: public, max-age=31536000, immutable
/rangefind/**/*.bin.gz         Cache-Control: public, max-age=31536000, immutable

With those headers, incremental updates are CDN-friendly by construction: unchanged packs keep their filenames, so a site update invalidates only the small manifests.

Small sites

For a docs or marketing site the index is usually a few megabytes — host it with the site itself and stop thinking about it. This site does exactly that on GitHub Pages.

Large indexes on object storage

Multi-gigabyte indexes (Wikipedia-scale corpora, geo datasets) belong on object storage with a CDN in front:

• Cloudflare R2 — native range support, zero egress fees. A ~750 GB planet-scale OSM index costs about $11/month in storage; a country- scale index (~10 GB) fits in the free tier. Front it with a custom domain and the cache rules above.
• Backblaze B2 + Cloudflare CDN — ~$6/TB-month with free egress through the bandwidth alliance; the cheapest legitimate setup at scale.
• Avoid standard S3/GCS egress pricing (~$0.09/GB) — per-query transfer is small, but bulk downloads and misses add up; eliminating egress cost is the point of this architecture.

For comparison: hosted search APIs charge per request (commonly ~$5 per 1,000). A million queries a month against a static index is a few dollars of CDN traffic — or zero on free tiers.

Offline and embedded

Because the index is plain files, it also works where servers can't: download a region's index directory and query it locally with rangefind/node, bundle one into an Electron or mobile webview app, or mirror it wherever your users are. No key management, no rate limits, no phoning home.