LLMs hallucinate.Your job is to ensure they don’t embarrass you, your company, or your brand.

Table of contents

1. Introduction
2. Choosing the right LLM
3. Human Evaluation – HHH Framework
4. Scaling Evaluation – LLM as Judges
5. Hands-On exercise to run evals

Introduction

LLMs are powerful but unpredictable. Unlike traditional software, they don’t always give the same answer to the same input. Without structured evaluation, you’re flying blind — you can’t guarantee quality, safety, or trust. Evals are how you make sure your AI is reliable enough to put in front of customers. Enter LLLm Evals .

• AI evals are like unit tests for agents.
• Difference between software testing vs AI evals:
  • Software testing & unit tests are deterministic. LLM agents are non-deterministic, with multiple possible paths.
  • Integration tests rely on code/docs, but improving agents relies on data.

Part 1 - Choosing the right LLM

• Start with requirements along these dimensions
  • Accuracy: example >90% ideal in legal contexts.
  • Latency: Medium–low is acceptable for offline jobs.
  • Cost: Less sensitive initially since manual legal review is expensive.
  • Context length: for example Must handle long docs (≈20K–1M tokens).
  • Throughput (QPM): Size for expected query volume.
  • Grounding: Model should cite source contract text.
• Some benchmarksyou could use , typically published by models -
  • Language understanding
  • Q&A
  • Document classification
  • Reasoning (planning, chain-of-thought)
  • Tool usage (email, APIs, CRM)
• Model Selection Use published evals/model cards (e.g., a “model matrix”) to compare options. Example: O3 High mini may offer a strong balance of accuracy, latency, and cost.

Part 2 Human Evaluation – HHH Framework

• Helpful → Solves problem, complete, succinct.
• Honest → Factual, validated, clickable links.
• Harmless → Ethical, legal, policy-aligned, guardrails applied.
• How to use:
  • Define north star metrics: Job completion, customer satisfaction.
  • Create yes/no evaluation questions.
  • Human evaluators score against ground truth.
  • Define launch thresholds (e.g., 60% helpfulness).
  • Update criteria with customer feedback.

Start with this simple framework

Part 3 Scaling Evaluation – LLM as Judges

• Challenge: Human eval doesn’t scale.
• Solutions:
  • Cloud APIs for moderation (toxicity, hate speech, frustration).
  • Train an LLM judge using good conversation examples.
  • Use SDKs or pre-built judges to measure: correctness, relevance, guideline adherence, context sufficiency, RAG chunk relevance.
• Process:
  • Collect requests + responses + ground truth → apply judges → score → dashboard metrics.
  • Continuous improvement: Judges improve over time but still an evolving field.

## Part 4 Hands-On exercise to run a few evals

• PromptFoo is a test-driven development framework for LLMs
• Install Node.js & npm (nodejs.org)
• Set up your OPENAI_API_KEY environment variable
• Configure promptfooconfig.yaml. Mine Here
• npm install -g promptfoo
• run this to run the evaluations npx -y promptfoo eval -c «config».yaml
• run this to see the results in the browser like below - npx -y promptfoo view -p 8080