Best practices

Short, opinionated guidance for getting the most out of the gateway. Each section links to the deeper docs.

Pick the right optimize mode

• Leave it on price (the default) for most traffic — the cheapest healthy route that can serve the request.
• Use latency for interactive, user-facing paths where responsiveness matters more than the per-token cost.
• Use uptime for critical jobs that must not fail — it favours the route with the lowest observed failure rate.
• Use auto when you don’t want to think about it: reliability comes first, then latency and price trade off sensibly.

Set a per-key default and override per request only where a specific call needs it. Full scoring is in Routing & residency.

Build a fallback chain before you need it

• Make chains cross-provider, not just cross-model — the point is to survive one provider going down. Put an India-resident step first if residency matters.
• Add a BYOE endpoint of your own as a ₹0 last resort so you always have somewhere to land.
• Save the chain as a Collection, version it, andfork before you use someone else’s so an upstream change can’t alter your routing silently.
• Watch it with reliability monitoring rather than assuming it works.

Enforce residency where it actually matters

• Send data_policy: "india_only" on regulated or sensitive workloads — it fails closed rather than leaving India, so it’s a guarantee, not a preference.
• Don’t set it blanket-wide if you don’t need it: it shrinks the route pool and can remove cheaper or faster global options.

Keep keys clean

• Use a separate key per environment or workspace (dev / staging / prod) so you can revoke one without disrupting the rest.
• Set a monthly ₹ budget, rate limit and expiry on every key — a leaked key then has a hard ceiling and a shelf life.
• For agents, mint ephemeral scoped keys rather than handing over a long-lived one — see MCP for agents.
• The full key is shown once; store it in a secret manager, never in source control.

Control cost

• Prefer the smallest model that clears your quality bar; reserve frontier models for the calls that need them.
• Use BYOK to ride your own provider rates where you already have an account — free during beta.
• Set per-key and (where available) per-workspace budgets so spend can’t run away, and reviewusage & activity regularly.

Before you call it production

• BharatRouter is a developer beta today — provided as-is with no SLA yet. Treat it accordingly for critical workloads and keep an eye onstatus.
• Handle errors idempotently and respect the error codes; a fallback chain plus monitoring is your safety net, not a reason to skip retries.

Questions? See the FAQ or email[email protected].