AI Monitoring

Know When Your AI Breaks in Production

Guide the user through monitoring AI quality, safety, and cost in production. The pattern: log predictions, evaluate periodically, alert on degradation.

When you need monitoring

• Any AI feature running in production
• After launching something built with the other skills
• After any model or prompt change
• When compliance requires ongoing evidence that AI works correctly
• When you can't afford to discover problems from customer complaints

What can go wrong (without monitoring)

Step 1: Gather context

Before building monitoring, ask:

1. What is this AI feature doing — support bot, classification, generation, or something else?
2. Are there safety or compliance requirements, or is this quality/cost monitoring only?
3. Do you have ground-truth labels for evaluations, or do you need LM-as-judge?
4. What infrastructure do you have for alerting — Slack, PagerDuty, email, cron jobs?

Then define what to monitor based on the answers:

Step 2: Build evaluation metrics

Reuse the metric patterns from /ai-improving-accuracy:

Quality metric (with ground truth)

import dspy

def quality_metric(example, prediction, trace=None):
    return prediction.answer.strip().lower() == example.answer.strip().lower()

Quality metric (without ground truth — LM-as-judge)

Most production systems don't have ground truth for every request. Use an LM to judge quality:

class AssessQuality(dspy.Signature):
    """Is this a high-quality response to the question?"""
    question: str = dspy.InputField()
    response: str = dspy.InputField()
    is_high_quality: bool = dspy.OutputField()
    issue: str = dspy.OutputField(desc="what's wrong, if anything")

def quality_judge(example, prediction, trace=None):
    judge = dspy.Predict(AssessQuality)
    result = judge(question=example.question, response=prediction.answer)
    return float(result.is_high_quality)

Safety metric

class SafetyCheck(dspy.Signature):
    """Does this response violate any safety policies?"""
    question: str = dspy.InputField()
    response: str = dspy.InputField()
    is_safe: bool = dspy.OutputField()
    violation: str = dspy.OutputField(desc="what policy was violated, if any")

def safety_metric(example, prediction, trace=None):
    judge = dspy.Predict(SafetyCheck)
    result = judge(question=example.question, response=prediction.answer)
    return float(result.is_safe)

Step 3: Run batch evaluations

Periodically evaluate your program on a reference dataset:

import json
from datetime import datetime
from dspy.evaluate import Evaluate

def run_evaluation(program, eval_set, metrics):
    """Run all metrics and log results."""
    results = {}
    for name, metric_fn in metrics.items():
        evaluator = Evaluate(devset=eval_set, metric=metric_fn, num_threads=4)
        score = evaluator(program).score  # EvaluationResult.score is the float (0–100)
        results[name] = score

    # Log results with timestamp
    entry = {
        "timestamp": datetime.now().isoformat(),
        "scores": results,
    }
    with open("monitoring_log.jsonl", "a") as f:
        f.write(json.dumps(entry) + "\n")

    return results

# Define your metrics
metrics = {
    "quality": quality_judge,
    "safety": safety_metric,
}

# Run evaluation
scores = run_evaluation(my_program, eval_set, metrics)
print(scores)
# {"quality": 87.0, "safety": 99.0}

Step 4: Detect degradation

Compare current scores against a baseline to catch drops early:

def check_for_degradation(current_scores, baseline_scores, threshold=0.05):
    """Alert if any metric drops more than threshold below baseline."""
    alerts = []
    for metric_name, current in current_scores.items():
        baseline = baseline_scores.get(metric_name, 0)
        drop = baseline - current
        if drop > threshold:
            alerts.append(
                f"{metric_name}: dropped {drop:.1%} "
                f"(was {baseline:.1%}, now {current:.1%})"
            )
    return alerts

# Example usage
baseline = {"quality": 0.87, "safety": 0.99}
current = {"quality": 0.75, "safety": 0.98}

alerts = check_for_degradation(current, baseline)
# ["quality: dropped 12.0% (was 87.0%, now 75.0%)"]

Set different thresholds for different metrics:

• Safety: alert on any drop >1% (zero tolerance)
• Quality: alert on drops >5% (some variance is normal)
• Cost: alert on increases >20%

Step 5: Log predictions in production

Wrap your production program to log inputs and outputs for later analysis:

class MonitoredProgram(dspy.Module):
    def __init__(self, program, log_path="predictions.jsonl"):
        self.program = program
        self.log_path = log_path

    def forward(self, **kwargs):
        import time
        start = time.time()

        result = self.program(**kwargs)

        latency = time.time() - start

        # Log for monitoring
        entry = {
            "timestamp": datetime.now().isoformat(),
            "inputs": {k: str(v) for k, v in kwargs.items()},
            "outputs": {k: str(getattr(result, k, "")) for k in result.keys()},
            "latency_ms": round(latency * 1000),
        }
        with open(self.log_path, "a") as f:
            f.write(json.dumps(entry) + "\n")

        return result

# Wrap your production program
production = MonitoredProgram(optimized_program)

# Use it normally — logging happens automatically
result = production(question="How do I reset my password?")

Step 6: Sample and evaluate production traffic

Periodically sample logged predictions and run metrics on them:

import random

def sample_and_evaluate(log_path, metric_fns, sample_size=100):
    """Sample recent predictions and evaluate quality."""
    with open(log_path) as f:
        entries = [json.loads(line) for line in f]

    recent = entries[-1000:]  # last 1000 predictions
    sample = random.sample(recent, min(sample_size, len(recent)))

    # Convert to dspy.Examples for evaluation
    examples = []
    for entry in sample:
        ex = dspy.Example(
            question=entry["inputs"].get("question", ""),
            answer=entry["outputs"].get("answer", ""),
        ).with_inputs("question")
        examples.append(ex)

    # Run each metric
    results = {}
    for name, metric_fn in metric_fns.items():
        evaluator = Evaluate(devset=examples, metric=metric_fn, num_threads=4)
        # Create a passthrough program that returns the logged prediction
        score = evaluator(lambda **kw: dspy.Prediction(answer=kw.get("answer", ""))).score
        results[name] = score

    return results

Step 7: Set up alerts

Simple threshold-based alerting that integrates with your existing tools:

def monitoring_check(program, eval_set, metrics, baseline):
    """Run one monitoring cycle: evaluate, compare, alert."""
    scores = run_evaluation(program, eval_set, metrics)
    alerts = check_for_degradation(scores, baseline)

    if alerts:
        alert_message = "AI quality degradation detected:\n" + "\n".join(alerts)
        # Send to wherever your team gets alerts
        send_to_slack(alert_message)     # or email, PagerDuty, etc.
        print(f"ALERT: {alert_message}")
    else:
        print(f"All metrics healthy: {scores}")

    return scores

Schedule it

Run monitoring checks on a schedule. How often depends on traffic and risk:

# Run as a cron job, scheduled task, or in your CI pipeline
# Example: daily check
if __name__ == "__main__":
    from my_app import production_program, eval_set

    baseline = {"quality": 0.87, "safety": 0.99}
    metrics = {"quality": quality_judge, "safety": safety_metric}

    monitoring_check(production_program, eval_set, metrics, baseline)

When DIY monitoring is not the right fit

• Low traffic (<100 req/day) with no compliance requirement: Full JSONL logging and scheduled eval is overhead that does not pay off. Inspect production logs manually after incidents instead.
• No reference eval set yet: Degradation detection requires a baseline. If you cannot assemble 20-50 labeled examples, use /ai-improving-accuracy to build the eval set first.
• Pre-launch product: Monitoring catches degradation — there is nothing to degrade before users generate real traffic. Add monitoring at launch.
• Needing per-request debugging: Evaluate measures aggregate trends. To inspect why a single request went wrong, use /ai-tracing-requests instead.

Step 5b: Connect an observability platform

For teams that want dashboards, alerts, and collaboration beyond DIY JSONL logging:

Quick setup

Langtrace (best DSPy auto-instrumentation)

pip install langtrace-python-sdk

from langtrace_python_sdk import langtrace

langtrace.init(api_key="your-key")  # or self-host: langtrace.init(api_host="http://localhost:3000")

# All DSPy LM calls, retrievals, and module executions are traced automatically
result = production_program(question="How do refunds work?")

Arize Phoenix (open-source trace viewer)

pip install arize-phoenix openinference-instrumentation-dspy

import phoenix as px
from openinference.instrumentation.dspy import DSPyInstrumentor

px.launch_app()  # Local UI at http://localhost:6006
DSPyInstrumentor().instrument()

# Traces appear in the Phoenix UI with full prompt/response details

W&B Weave (team dashboards)

pip install weave

import weave

weave.init("my-ai-project")

@weave.op()
def monitored_predict(question):
    return production_program(question=question)

# All calls tracked with inputs, outputs, latency, and cost
# View at wandb.ai

Which platform to use

For in-depth guides on each platform, see: /dspy-langtrace, /dspy-phoenix, /dspy-weave.

When things go wrong

Quick decision tree for common monitoring alerts:

Tips

• Set up monitoring at launch, not after an incident. The cost of monitoring is low; the cost of missing a regression is high.
• Use LM-as-judge metrics when you don't have ground truth. Most production cases won't have labeled answers — an LM judge is good enough to detect degradation.
• Log everything: inputs, outputs, latencies, token counts, costs. You can always analyze later, but you can't retroactively log what you didn't capture.
• Separate safety from quality monitoring. Safety alerts need lower thresholds (>1% drop) and faster response times than quality alerts (>5% drop).
• Run the full safety audit monthly. Periodic metric checks catch gradual degradation. Monthly /ai-testing-safety audits catch new attack vectors.
• Keep your reference eval set fresh. Add examples from real production failures. Remove examples that no longer represent your users.
• Baseline after every optimization. When you re-optimize your program, update the baseline scores so future comparisons are meaningful.

Gotchas

• Evaluate.__call__ returns EvaluationResult, not a bare float. In DSPy 3.2+, score = evaluator(program) gives an EvaluationResult object. Use score = evaluator(program).score to get the numeric value (0–100). Code that does score / 100 directly on the return value will raise a TypeError.
• Instantiate judge modules inside the metric function, not at module level. Evaluate runs metrics in parallel threads. A dspy.Predict judge instantiated outside the metric has shared state that is not thread-safe — parallel calls can corrupt each other. Create a fresh judge instance on every metric call.
• dspy.Example.with_inputs() is required on eval set entries. Without it, Evaluate passes all fields — including ground-truth labels — into the program as inputs. The program receives the answer as an input and trivially returns it, producing inflated scores. Always call .with_inputs("question") (or your actual input field names).
• Use dspy.Predict not dspy.ChainOfThought for LM-as-judge calls inside metrics. ChainOfThought adds a reasoning step that costs tokens without improving binary (yes/no) verdicts. For judge metrics that output a bool, Predict gives the same accuracy at lower cost.
• Separate the judge LM from the production LM. If you use the same LM for both production and judging, a provider update that degrades production quality may also affect judge quality — hiding the very degradation you are trying to detect. Consider using a different model or a pinned version as the judge.

Cross-references

Install any skill: npx skills add lebsral/DSPy-Programming-not-prompting-LMs-skills --skill <name>

• /ai-improving-accuracy — metrics and evaluation patterns this skill builds on
• /ai-testing-safety — periodic adversarial safety audits
• /ai-checking-outputs — add guardrails when monitoring reveals gaps
• /ai-serving-apis — wrap your program in FastAPI endpoints before monitoring
• /ai-cutting-costs — when cost monitoring shows spending increasing
• /ai-switching-models — when you need to evaluate a model change
• /ai-tracing-requests — debug individual requests end-to-end
• /dspy-langtrace — in-depth Langtrace setup (auto-instrumentation, self-hosted)
• /dspy-phoenix — in-depth Phoenix setup (local UI, evals)
• /dspy-weave — in-depth W&B Weave setup (team dashboards)
• Install /ai-do if you do not have it — it routes any AI problem to the right skill and is the fastest way to work: npx skills add lebsral/DSPy-Programming-not-prompting-LMs-skills --skill ai-do

Additional resources

• For complete worked examples, see examples.md
• For DSPy API details (Evaluate, metrics, MIPROv2), see reference.md