Skip to content

Chapter 0 — Introduction

Level: 🟢 Beginner · Status: 🟩 Complete Prerequisites: a little Python. No ML background needed. You'll build: a mental model of the whole search pipeline.

At a glance

This chapter sets the stage. By the end you'll understand what semantic search is, why LLMs changed search, and how the rest of the course fits together — so every later chapter has a place to land.

The motivation

Imagine you run a website with thousands of articles — picture a small Wikipedia, or a store with a big product catalog. People need to find things. For decades the answer was keyword search: match the words in the query to the words in the documents.

That works until someone searches for "how do I stop my laptop from dying so fast" and your best article is titled "Improving battery life on portable computers." Not a single important word matches — yet it's exactly the right answer. Keyword search matches words; people search by meaning. Closing that gap is what this course is about.

Intuition first: the four-step pipeline

Everything we build fits one simple pipeline. Keep this picture in your head:

            ┌──────────┐     ┌───────────┐     ┌────────────┐     ┌────────────┐
  query ──▶ │ RETRIEVE │ ──▶ │ RE-RANK   │ ──▶ │  GENERATE  │ ──▶ answer
            │ (find    │     │ (reorder  │     │ (LLM writes│
            │ candidates)    │  the best)│     │  from them)│
            └──────────┘     └───────────┘     └────────────┘
              Ch 2,5,6,8        Ch 7              Ch 10,11
  • Retrieve — quickly pull a handful of likely-relevant documents from the whole corpus.
  • Re-rank (optional) — carefully reorder those candidates so the best is on top.
  • Generate (optional) — hand the top results to an LLM so it can write a direct answer grounded in them (this is RAG).

Some systems stop after Retrieve (a classic search box). Others go all the way to Generate (an AI assistant that answers questions). You'll build each stage and understand the trade-offs.

What "semantic" adds

The single most important upgrade is in the Retrieve step. Instead of matching words, we convert text into embeddings — vectors of numbers that capture meaning — and find documents whose vectors are closest to the query's vector. "Battery life" and "stop my laptop dying" land near each other in this space even with no shared words. That's semantic search, and Chapters 3–6 build it up from scratch.

How to use this repo

  1. Do the setup once.
  2. Read chapters in order if you're new; jump around using the roadmap if you're experienced.
  3. Read each chapter's README, then run its notebooks/ to see the idea in code.
  4. Keep the glossary open in a tab.

Hands-on

Notebook: notebooks/00_pipeline_tour.ipynb — a 5-minute end-to-end demo: stand up a search system in three lines with the course's SearchStack, ask it questions in plain English, and watch it answer by meaning — so you see the destination before learning to build each part.

Slides

📊 Chapter 0 slide deck (PDF) — a visual intro: beyond keywords, to meaning.

Key terms

Large Language Model (LLM), embedding, retrieval, ranking, semantic search, RAG. (See GLOSSARY.)

Check your understanding

  1. In one sentence, why does keyword search miss relevant results?
  2. What does the "Retrieve" step do, and how does "semantic" change it?
  3. Which pipeline stages are optional, and when would you skip them?

💡 Try answering each question yourself first, then expand to check.

Show answer — 1 Because it matches literal *words*, not *meaning*: if the best document phrases the same idea with different words, keyword search sees no overlap and misses it.
Show answer — 2 **Retrieve** quickly pulls a small set of likely-relevant documents from the whole corpus. Making it *semantic* means comparing the *meaning* of the query and documents (via embeddings) instead of shared words, so paraphrases still match.
Show answer — 3 **Re-rank** and **Generate** are optional. Skip re-ranking when first-stage results are already good enough or latency is tight; skip generation when you just want to return passages (a search box) rather than a written answer.

References

  • DeepLearning.AI × Cohere — Large Language Models with Semantic Search (course inspiration).
  • Links to primary sources are added as we write later chapters.