Evals & Observability¶
Production AI fails silently without measurement. This hub ties together evaluation, tracing, monitoring, and safety across the handbook.
The quality loop¶
flowchart LR
Build[Build / Prompt] --> Offline[Offline Evals]
Offline --> Gate{Quality Gate}
Gate -->|Pass| Deploy[Deploy]
Gate -->|Fail| Build
Deploy --> Online[Online Monitoring]
Online --> Alert[Alerts & Feedback]
Alert --> Build
Learning path¶
| Step | Course | Focus |
|---|---|---|
| 1 | Course 13 · LLM Evals | Golden sets, LLM-as-judge, CI/CD gates |
| 2 | Course 12 · LLMOps | Observability, caching, A/B tests |
| 3 | Course 08 · Harness L6 | Agent step tracing |
| 4 | Course 14 · Safety | Red teaming, adversarial evals |
| 5 | Course 06 · RAG L8 | RAG-specific metrics |
Eval types¶
| Type | When | Handbook |
|---|---|---|
| Unit / regression | Every PR | Course 13 |
| RAG retrieval | Chunking / index changes | Course 06 · RAG eval metrics |
| Agent trajectory | Tool-use agents | Course 13 · Agent trajectory evals |
| Safety / red team | Pre-release | Course 14 · Red teaming |
| Online / drift | Production | Course 13 · Production monitoring |
Observability pillars¶
| Pillar | What to capture | Tools (OSS) |
|---|---|---|
| Traces | Full request path, agent steps | Langfuse, OpenTelemetry, LangSmith |
| Logs | Prompts, tool I/O (redacted) | structlog, CloudWatch |
| Metrics | Latency, tokens, cost, error rate | Prometheus, Grafana |
| Feedback | Thumbs, edits, escalations | Custom + Course 13 monitoring |
OSS hubs¶
- Awesome Agent Evals — eval frameworks catalog
- DeepEval — pytest-style LLM tests
- Promptfoo — red team + CI evals
- RAGAS — RAG metrics
Related¶
- Agentic AI — what you're measuring
- Essential Papers — HELM, MT-Bench, etc.
- Roadmap — upcoming observability labs