Verification Engine

See how the verification engine fact-checks AI content against your knowledge base — through a real scenario of a consultant who uses an AI assistant for email, tickets, and client communications.

The Problem

The Solution: GroundTruth as a Safety Layer

Instead of letting her assistant call external APIs directly, Sarah routes every action through GroundTruth's execute endpoint:

Step 1: Upload Her Knowledge Base

Sarah uploads her key business documents to the Library:

Now when her assistant drafts an email mentioning "$75/hr", GroundTruth checks it against the Rate Card and flags the discrepancy.

Step 2: Create Safety Policies

Sarah creates five policies that match her risk tolerance. Here's how they look in the Policies dashboard:

Step 3: Wire Up the Assistant

Sarah adds a wrapper so every assistant action goes through GroundTruth. Here's the key integration code:

import requests

GROUNDTRUTH_URL = "https://app.groundtruth.dev/api/execute"
GROUNDTRUTH_KEY = "hg_sk_sarah_key"

def safe_execute(action: str, params: dict, content: str = None,
                 channel: str = None):
    payload = {
        "action": action,
        "channel": channel,
        "params": params,
        "context": { "agent": "sarah-assistant", "userId": "sarah" }
    }
    if content:
        payload["content"] = content

    response = requests.post(
        GROUNDTRUTH_URL,
        headers={"Authorization": f"Bearer {GROUNDTRUTH_KEY}"},
        json=payload
    )
    return response.json()

# Example: send a client email
result = safe_execute(
    action="send_email",
    channel="email",
    params={
        "to": "client@acme.com",
        "subject": "Project Update - March",
        "body": "Hi Tom, our rate remains $150/hr as agreed."
    },
    content="Our rate remains $150/hr as agreed."
)
# result["decision"] → "approved" | "escalated" | "blocked"

Step 4: A Day in Sarah's Life

Here's a typical day with the safety layer active. Each scenario shows exactly which policy fired and the specific condition that triggered it.

Step 5: End-of-Day Review

Sarah opens the Execution Log to see everything her assistant did today:

7 actions attempted, 3 approved, 2 blocked, 2 escalated (then approved after review). Zero wrong information sent to clients.

Results

Build Your Own

You can replicate Sarah's setup in about 15 minutes:

Related Docs