Skip to main content

Week 03 — Hands-on Labs

Two labs this week. Together they cover the full LLM engineering cycle: extract real-world data → process with LLMs → structure the output → measure cost → benchmark strategies.

Labs Overview​

LabTitleTimeDifficulty
Lab 3.1YouTube → Subtitles → Topics → Timestamps JSON Pipeline~3 hrs⭐⭐⭐
Lab 3.2Cost-Tracking Dashboard via LangSmith + Prompt Strategy Benchmarks~2 hrs⭐⭐⭐

What You'll Have After Both Labs​

code
Lab 3.1 output:
  ✅ CLI tool that accepts any YouTube URL
  ✅ Downloads subtitles with yt-dlp (no API key needed)
  ✅ Extracts topics and answer timestamps using Claude
  ✅ Produces a structured summary.json with chapters, topics, Q&A pairs

Lab 3.2 output:
  ✅ LangSmith project with 50+ traced LLM calls
  ✅ Cost breakdown by model, prompt strategy, and task type
  ✅ Benchmark comparing zero-shot vs few-shot vs CoT vs Self-Consistency
  ✅ A FastAPI dashboard endpoint that serves live cost analytics

Prerequisites​

Before starting, make sure you have:

  • Anthropic API key set in .env
  • OpenAI API key set in .env (for Lab 3.2 comparisons)
  • LangSmith account at smith.langchain.com (free tier)
  • LangSmith API key set in .env
  • yt-dlp installed: uv tool install yt-dlp
  • ffmpeg installed (required by yt-dlp for some formats)

Environment Setup​

Create one .env file for both labs:

bash
.env
ANTHROPIC_API_KEY=sk-ant-...
OPENAI_API_KEY=sk-...
LANGCHAIN_TRACING_V2=true
LANGCHAIN_API_KEY=ls__...
LANGCHAIN_PROJECT=tds-week-3-labs
Project 1 Notice

These labs contribute to Project 1 (due end of Week 3). Build them carefully — the pipeline patterns from Lab 3.1 and the benchmarking mindset from Lab 3.2 are directly applicable to the project.