Why Search Needed to Change

Two kinds of data, one search problem

Every organization sits on two very different kinds of information.

Structured data lives in tidy rows and columns — PLM (Product Lifecycle Management) and ERP (Enterprise Resource Planning) systems, data lakes, web analytics tables. It was built to be queried: filter by date, sort by price, join on a customer ID.

Unstructured data is everything else, and there's far more of it — wikis, internal docs, emails, slide decks, images, videos, meeting recordings, social posts. It has no rows or columns. A traditional database has no idea what a paragraph of text means, only what characters it contains.

Keyword search was the first attempt to make unstructured data searchable: index the words, match the words in the query. It works until it doesn't — ask "how do I get my money back?" and a document titled "Refund Policy" won't match if it never uses the word "money." The words are different even though the meaning is identical.

The unsung hero: embedding models

Generative AI gets the headlines — large language models writing summaries, translations, code. But underneath the search systems this course is about sits a quieter, less flashy kind of model: the embedding model.

An embedding model doesn't generate text. Its one job is to convert the meaning of a word, sentence, or document into a list of numbers — a vector — such that pieces of text with similar meaning end up with similar vectors. That single capability is what powers classification, clustering, recommendation systems, and — the focus of this course — semantic search.

If you've already seen the beginner-level tour of RAG in Introduction to Generative AI, this course picks up right where that one left off and goes deeper into the mechanics: how embeddings actually work, how documents get chunked and indexed, how results get reranked, and how you'd measure whether any of it is actually working.

Why this matters

Nearly every technique in this course — dense retrieval, RAG, chunking, reranking — is built on the same underlying idea: turn meaning into numbers, then use math on those numbers to find what's relevant. Understanding embeddings first makes everything that follows click into place instead of feeling like a pile of separate tools.

Key takeaway

Structured data was already searchable; unstructured data wasn't, until embedding models made it possible to search by meaning instead of by matching exact words.

What's next?

There isn't just one way to use language models for search. The next lesson lays out the three major approaches — dense retrieval, reranking, and RAG — and how they relate to each other.