Retry Without Paying Twice: Idempotency Keys for Safer AI API Calls



Your app sends a request to an AI model. Ten seconds pass. The connection drops, your code times out, and it retries. A moment later, the user taps "Send" again for good measure. Was that one AI call, two, or three? And how many did you pay for?

Retries are a normal part of running AI features in production. The trick is making them safe. This post explains, in plain terms, how idempotency keys help, and how OROO Blue uses them so a retried request isn't charged twice.

Why retries quietly cost you money

A timeout on your side does not always mean the request failed on the other side. The model may have finished the work, but the response never reached you. If your code simply sends the same request again, the provider sees a brand-new call. You get billed again for tokens, and any side effects in your own system (saving a reply, sending a notification, deducting a user's credits) can happen twice too.

At small scale this is a rounding error. At production scale, with flaky mobile networks and impatient users, it adds up to wasted spend and confusing bugs.

What an idempotency key actually does

An idempotency key is a unique ID you attach to a request. It tells the server: "If you've already seen this exact operation, don't do it again; treat this as a repeat." The same key sent twice means one operation, not two.

The important rule is simple: one key per user action, reused on every retry of that action. If a chat feature, a support bot or an internal tool triggers one AI request, generate one key for it. If you retry because of a timeout, send the same key. If the user starts a genuinely new action, generate a new key.

Here is a generic, vendor-neutral sketch of the pattern (not any specific API):

key = new_unique_id()          // once per user action
for attempt in 1..max_attempts:
    result = call_ai(request, idempotency_key = key, timeout = T)
    if result.ok: return result
    if not retryable(result.error): raise result.error
    sleep(backoff_with_jitter(attempt))
raise "gave up after retries"

Retry smart, not just often

Idempotency makes retries safe; good retry habits make them sensible. A few practical tips:

  • Set clear timeouts. Long AI responses need generous timeouts, but not infinite ones.
  • Use exponential backoff with jitter. Wait a little longer after each failure, with some randomness, so many clients don't retry at the same instant.
  • Retry only what's worth retrying. Network errors, timeouts and temporary server or rate-limit errors are usually retryable. Bad requests or authentication errors won't fix themselves, so fail fast.
  • Cap the attempts. Three to five tries is a common starting point. After that, show the user a friendly message.
  • Log the key. Record the idempotency key with each attempt so you can trace exactly what happened later.

How OROO Blue fits in

OROO Blue is a secure AI gateway for Indian teams and developers. With one API and one OROO key, you can reach OpenAI, Google and Anthropic models using the model names oroo-chatgpt, oroo-gemini and oroo-claude. Send an idempotency key with your request, and a retried request won't be charged twice.

Billing is prepaid in rupees through OROO Mint, a wallet you top up via Razorpay, and you pay only for tokens used. That makes double charges easy to spot and, with idempotency keys, easy to avoid.

A few other controls help keep retry logic from turning into a runaway loop:

  • Per-key model, token and rate limits, so one buggy retry loop can't burn through your wallet.
  • Freeze, rotate or revoke keys when something looks wrong.
  • IP allowlists, so a key only works from the places you expect.
  • Key secrets shown once and stored hashed.
  • No storage of prompts or completions by OROO.

Build retry-safe AI features with OROO Blue

Retries will happen. Paying twice for them doesn't have to. Generate one idempotency key per user action, retry with backoff, and let your gateway do the rest. Explore OROO Blue and start building at orooblue.com.

Comments

Post a Comment

Popular posts from this blog

The Same-Day Recall Habit: How Students Can Remember More with OROO GRIXI

Image
Most students in Classes 6–12 know this feeling: you sat through the lesson, you took notes, and a week later the chapter feels brand new again. The notes didn't fail you. The problem is what usually happens to them after the bell rings: nothing, until the night before the test. In this post we'll look at a simple habit called same-day recall and how OROO GRIXI , an AI study companion for students, makes it quick to do every day. Why re-reading your notes isn't enough Re-reading feels productive because the words look familiar. But recognising something on a page is very different from being able to recall it in an exam hall with a blank answer sheet in front of you. Learning research has long pointed to two habits that work better: retrieval practice (pulling information out of your memory by testing yourself) and spacing (coming back to it over several days instead of cramming once). The good news is that neither habit needs hours. A short, focused session on the s...
Read more

Morning Attendance in One Place: A Practical Guide for School Offices

Image
For many schools, the morning starts with a familiar scramble: paper registers on desks, a quick message in a staff chat, a call from a parent asking whether their child was marked present, and an office still waiting for every class list to come in. Attendance should be the simplest routine of the day. When it lives in too many places, it becomes one of the noisiest. This guide shares practical habits for principals, school offices and teachers — and shows how OROO Campus , a free school ERP and AI operating system for schools and coaching institutes, keeps morning attendance in one shared record alongside the rest of school life. Why morning attendance gets messy Attendance problems rarely begin with a dramatic failure. They start with small gaps. One teacher marks on paper; another sends a list on chat; a substitute uses yesterday's class sheet; the office updates a spreadsheet later. By mid-morning, three people can give three different answers about the same child. The root is...
Read more

GPT, Gemini or Claude? How to Pick the Right AI Model by Testing on Your Own Data with OROO Blue

Image
"Which model should we use: GPT, Gemini or Claude?" It's one of the first questions every team asks when building an AI feature, and the internet is full of confident answers. The honest one is simpler: the best model is the one that does best on your own data . In this post we'll walk through a practical way to run that comparison, and how OROO Blue , a secure AI gateway for Indian teams and developers, makes testing and switching models a small change instead of a big project. Start by defining what "good" means Before you call a single model, write down what a good answer looks like for your feature. A support reply might need to be accurate, polite and under 120 words. A data extraction task might need valid JSON with every field filled. A summary might need to keep key numbers and names intact. Practical tip: turn these into a short scoring checklist of three to five points. If you can't describe "good", no benchmark will tell you whi...
Read more