Gemini Integration

Evaluate Google Gemini model inputs and outputs against your governance policies.

Overview

The Gemini integration lets you evaluate Google AI model inputs and outputs against your governance policies using the Aguardic SDK. Your application calls Gemini directly — Aguardic never touches your LLM traffic. Policy evaluation happens in your code before and after each LLM call.

Setup

Create integration

Navigate to Integrations in the Aguardic dashboard, click Add Integration, and select Gemini. Give it a name and copy the API key.

Bind policies

Go to Policy Bindings and bind your governance policies to the Gemini integration.

Install the SDK

npm install @aguardic/sdk

Add evaluation calls

Use the SDK to evaluate inputs before sending them to Gemini, and evaluate outputs after receiving them.

Store your Aguardic API key securely. It is shown only once. If you lose it, regenerate it from the integration settings.

Code Example

import Aguardic from "@aguardic/sdk";
import { GoogleGenerativeAI } from "@google/generative-ai";
 
const aguardic = new Aguardic(process.env.AGUARDIC_API_KEY);
const genAI = new GoogleGenerativeAI(process.env.GEMINI_API_KEY);
 
const prompt = "Summarize our compliance requirements";
 
// 1. Evaluate input before sending to Gemini
const inputCheck = await aguardic.evaluate({
  input: {
    provider: "gemini",
    model: "gemini-2.0-flash",
    prompt,
  },
  targetKey: "chat-completion",
});
 
if (inputCheck.enforcementAction === "BLOCK") {
  console.log("Blocked:", inputCheck.violations);
} else {
  // 2. Call Gemini directly
  const model = genAI.getGenerativeModel({ model: "gemini-2.0-flash" });
  const result = await model.generateContent(prompt);
  const responseText = result.response.text();
 
  // 3. Evaluate output
  const outputCheck = await aguardic.evaluate({
    input: {
      provider: "gemini",
      model: "gemini-2.0-flash",
      response: responseText,
    },
    targetKey: "chat-completion-output",
  });
 
  if (outputCheck.enforcementAction === "BLOCK") {
    console.log("Output blocked:", outputCheck.violations);
  } else {
    console.log(responseText);
  }
}

What Gets Evaluated

Input Evaluation

PromptUser prompt and conversation history

System instructionThe system instruction if provided

ModelThe model name for model-specific policies

Multimodal contentText parts from content arrays

Output Evaluation

Response textGenerated text from the model response

Function callsAny function calls the model produced

Enforcement Actions

The enforcementAction field in the evaluation response tells you what to do:

BLOCKDo not send the request to Gemini (input) or do not show the response to the user (output).

APPROVAL_REQUIREDHold the request until a reviewer approves it. Poll the reviewRequestId for a decision.

WARNAllow the request but log the violation for review.

ALLOWNo violations or policy allows continuation.

Why Sidecar Over Proxy

Unlike proxy-based approaches that route all LLM traffic through a third party, Aguardic's sidecar model means:

No added latency — Evaluation runs in parallel, not in the request path
No single point of failure — Your LLM calls go direct; if Aguardic is down, you decide the fallback
Your keys stay with you — Aguardic never sees your Gemini API key
Full control — You choose exactly when and what to evaluate

Next Steps

OpenAI Integration — Same pattern for GPT models
Anthropic Integration — Same pattern for Claude models
JavaScript SDK — Full SDK reference
Your First Policy — Create policies to evaluate against

Anthropic GitHub