Your Cheat Sheet
The whole course on one page.
Here's the whole course, as short as we can make it. Pause and take a screenshot of each screen. These will save you time on every future project.
- The whole course, compressed
- Pause and screenshot
- Use on every project
Which tool for which job?
Fixed patterns, use regex. Sort text at volume, use TF-IDF and logistic regression. Few labels, use embeddings with a simple model. Find facts, use spaCy or a transformer. Search by meaning, use embeddings. Answer from documents, use RAG. Change style, fine tune.
| Job | Tool |
|---|---|
| Fixed patterns | Regex |
| Classify at volume | TF-IDF + logistic |
| Few labels | Embeddings + logistic |
| Extract entities | spaCy, transformer NER |
| Search | Hybrid + reranker |
| Answer from docs | RAG |
| Change style | Fine-tune |
Which metric?
Classification, read precision and recall per class, plus macro F one. Retrieval, use hit rate and MRR. RAG answers, use faithfulness and relevance. Generated text, use human review or a model judge. And for labels, use kappa.
| Task | Metric |
|---|---|
| Classification | Per-class F1, macro F1 |
| Retrieval | Hit rate, MRR |
| RAG answers | Faithfulness, relevance |
| Generation | Human review, judge |
| Labelling | Cohen's kappa |
The recipe
And here's the recipe. Name the job. Build a cheap baseline. Make a golden set. Try prompts, then RAG, then fine tuning. Add guardrails. Estimate your costs. Launch in stages. And keep measuring.
Summary
Three screens, and you've got the whole course. One last lesson. Let's talk about where you go next.
Key takeaways- Pick the right tool
- Measure the right thing
- Follow the recipe