Build an AI API with Google Gemini

Wrapping Gemini in an endpoint takes ten lines. Doing it without leaving your API key open to the internet takes a few more — here are both.

Step 1 — Key and model, as variables

Create a key at aistudio.google.com/app/apikey, then set two managed variables:

Step 2 — The endpoint

->json_escape is not optional. The request body is JSON being assembled from user input, so a prompt containing a quote character produces malformed JSON and a 400 from Google — and a prompt built to close the string early could append fields to the payload you never intended to send. The escape runs before substitution, which turns that from a live problem into a non-issue. Any user text going into a JSON body needs it.

Step 3 — Return an answer, not a payload

Gemini's response is deeply nested — candidates, parts, safety ratings, token counts. Useful when you want it, noise when you don't:

Step 4 — Constrain the answer

The cheapest way to control output is to prepend an instruction to every prompt:

Step 5 — Don't ship it open

Here's where the tutorial version stops and the real one starts. What you have so far is a public URL that spends your money. Two things fix it.

Step 6 — Watch what it costs you

LLM endpoints fail differently to normal ones. They rarely go down; they go slow, and the slowness is the model's, not yours. Every request produces a trace showing the Gemini call as its own span, so "is it us or them" is a question you answer by looking rather than guessing. Prometheus metrics per route give you the request rate to compare against your quota before you hit it.

The gotcha that looks like a bug

A 429 with limit: 0 in the error body is not rate limiting. It means the model you asked for isn't on your key's free tier at all — gemini-2.0-flash often isn't. Switch GEMINI_MODEL to gemini-2.5-flash or gemini-flash-latest, or enable billing. Because the model is a managed variable, that's a one-field change with no redeploy.

The shortcut

All three variants ship as one pack — full payload, text-only, and the instruction-prefixed version. Set two variables and deploy.

Related packs

• Gemini AI — Query Google's Gemini models over a simple HTTP API. Three routes: the full generateContent payload, just the answer text, or an instruction-prefixed answer. The API key and model are managed variables, so you can swap models without redeploying; the prompt is JSON-escaped before it hits the Gemini API.
• LLM Gateway — A configurable LLM completion and content-moderation endpoint. System prompt, model, and API key live in managed variables — swap providers or update prompts without touching the config or redeploying. Exposes POST /ai/complete (chat completion) and POST /ai/moderate (OpenAI moderation, free) with request validation and per-model metrics.
• RAG Chatbot — Retrieval-Augmented Generation chatbot backed by Postgres + pgvector. Ingest your own documents, store OpenAI embeddings alongside them, and answer questions using only your data — no third-party vector DB, no proprietary cloud, nothing leaves your infrastructure. Includes seed, ingest, chat, history, and delete endpoints.
• Rate Limited API — A public endpoint rate-limited to 10 requests/min per IP. Demonstrates the native network rate_limit policy.