coreason-sentinel 0.2.0

alerts you when things go wrong in production

coreason-sentinel

The "Watchtower" & Circuit Breaker for Production AI Agents

coreason-sentinel is an automated monitoring service for deployed LLM agents. It goes beyond simple "System Error" tracking to detect "Cognitive Errors" (Hallucinations) and "Business Errors" (Budget Spikes). Acting as a pharmacovigilance unit for AI, it monitors the agent's behavior in the real world, detects side effects like drift or toxicity, and triggers a Circuit Breaker to protect users and budgets if safety limits are breached.

Features

• Deep Tracing & Observability: visualize the entire cognitive chain (Retrieve -> Rerank -> Scout -> Generate) using OpenTelemetry and Arize Phoenix.
• Holistic Signal Detection: Monitors three key signal planes:
  • Cognitive: Hallucination rates, RAG relevance.
  • Business: Token costs, Latency spikes.
  • User: Sentiment analysis, refusal rates, and frustration signals.
• The Drift Engine: Advanced statistical detection for:
  • Content Drift: Changes in retrieved document vectors.
  • Style Drift: "Lazy Agent" syndrome detection using Kullback-Leibler (KL) Divergence.
  • Relevance Drift: Semantic distance between queries and responses.
• Automated Circuit Breaker: A "Dead Man's Switch" that transitions to OPEN (blocking traffic) if quality or safety drops below configured thresholds. Includes HALF-OPEN self-healing to test recovery.
• The Spot Checker: Automated QA that randomly samples live traffic (or focuses on negative sentiment) and loops it back for grading.

For a full breakdown of requirements and philosophy, see the Product Requirements Document.

Installation

pip install coreason-sentinel

Usage

Here is how to initialize the Circuit Breaker to protect your agent:

import time
from redis import Redis
from coreason_sentinel.circuit_breaker import CircuitBreaker
from coreason_sentinel.models import SentinelConfig, CircuitBreakerTrigger

# 1. Configure the Sentinel
config = SentinelConfig(
    agent_id="agent-alpha",
    owner_email="[e-mail]",
    phoenix_endpoint="http://localhost:6006",
    triggers=[
        # Trip if Faithfulness drops below 0.7 in the last hour
        CircuitBreakerTrigger(metric="faithfulness", threshold=0.7, window_seconds=3600, operator="<"),
        # Trip if Latency exceeds 10s in the last minute
        CircuitBreakerTrigger(metric="latency", threshold=10.0, window_seconds=60, operator=">"),
    ]
)

# 2. Initialize the Circuit Breaker
# Note: You must provide a valid Redis client and a NotificationService implementation.
# (Assuming a mock notification service for this example)
class MockNotificationService:
    def notify(self, message):
        print(f"NOTIFICATION: {message}")

breaker = CircuitBreaker(
    redis_client=Redis(),
    config=config,
    notification_service=MockNotificationService()
)

# 3. Protect your Agent
if not breaker.allow_request():
    # Return a maintenance message or failover
    print("Circuit Breaker OPEN: Traffic blocked due to safety violation.")
else:
    # Process your agent request...
    start_time = time.time()
    try:
        # ... agent logic ...
        print("Agent processing request...")
        pass
    finally:
        # Record metrics for the breaker to monitor
        # In a real app, you would calculate actual latency
        breaker.record_metric("latency", time.time() - start_time)