Module 1The search problem
3 lessonsThe search problem. See Larkfield's search fail and price each failure, name the kinds of search, and set up the project with the catalogue and the query log.
Make search find the best product first. Measure relevance with graded judgments, then compare BM25, query understanding, dense and hybrid retrieval, re-ranking and learning to rank on the same queries, and choose a design by relevance, speed and cost.
Read every lesson as a guest, or sign in to enroll and track your progress.
Enroll to save the lessons that you complete. It is free.
You are enrolled. From now on, the lessons that you complete are saved.
This course teaches ranking for people: relevance judgments and ranking measures with confidence intervals, lexical search with an inverted index and BM25, query understanding in English and French, dense and hybrid retrieval, re-ranking with a cross-encoder, learning to rank from judgments and clicks, autocomplete, facets, zero results and business rules, and latency, cost and online tests. It does not teach retrieval as a step before a language model, gradient boosting from the start, or how to operate a search server.
Take another course if you want to:
By the end of this course, you will be able to:
Larkfield, an online shop for home and garden products, has a search box that many people use and few people like. Someone types “grey outdor cushion” and sees an empty page. Someone types “sofa” and sees sofa covers first. A French customer types “coussin gris” and gets nothing useful. Each failed search is a lost basket or a message to support.
In this course, you follow Priya, Larkfield’s data scientist, Amira, the developer who owns the search service, and Omar, the product manager, as they fix it. You start by measuring: you use graded relevance judgments, compute NDCG and recall by hand, and build a test set with common and rare queries and confidence intervals. Then you change one thing at a time and measure it, on the same queries, in one comparison table.
You go from a classic inverted index and BM25 to query understanding in English and French, search by meaning with embeddings, hybrid search, re-ranking with a cross-encoder and a learning-to-rank model with LightGBM. You learn why clicks are not relevance labels, and how a click model corrects for position.
You build one project through the course: larkfield-search, a small search engine with a search command and a ready-made search page. You finish with the parts people see, such as autocomplete, facets and business rules, and with the production work: latency and cost per thousand queries, and online tests. The catalogue and the judgments come from a real, open furniture dataset; the query log and the clicks are synthetic, made by our code.
You completed Machine Learning, From Problem to Reliable Model, or you have the same skills (train, validation and test sets, gradient boosting, metrics). You can read and change Python and pandas code, use a notebook and a terminal. The RAG and mathematics courses help but are not required; lessons that use their ideas explain them briefly and link to them. You need Python 3.12 on your computer, or a free Google or Kaggle account for the notebooks. You need no key, no graphics card and no money.
Courses to take first: Machine Learning: From Problem to Reliable Model.
AI engineers, backend developers and data scientists who must make search work for people, such as product search in a shop, help-centre search or internal search, and who want to change it by measurement, not by feel.
Experienced AI practitioners designed this course. AI helped us draft the content; people chose the teaching approach and the examples, and reviewed every lesson. The outputs that you see are the actual results of running the code. How we make our courses
The search problem. See Larkfield's search fail and price each failure, name the kinds of search, and set up the project with the catalogue and the query log.
Measure relevance before you change anything. Graded judgments and agreement, precision, recall, MRR and NDCG by hand, a test set with head, torso and tail queries and confidence intervals, and labels written by a language model.
Lexical search, the baseline to beat. An inverted index by hand and in SQLite, BM25 with each part of the score visible and field weights, and exact-match rules for codes and numbers.
Understand the query. Normalisation and stemming in English and French, spelling correction, synonyms and query rewriting tested against the test set, and filters from what the query asks for.
Dense and hybrid retrieval. Search by meaning with embeddings, the queries where it fails, reciprocal rank fusion and score blending, and French queries against an English catalogue.
Re-rank the top results. Two-stage ranking, a cross-encoder and its extra milliseconds, learning to rank with LightGBM on judgments, and why clicks need a click model before you learn from them.
The parts of search that people see. Autocomplete, facets that never return zero, fallbacks and business rules with their cost in relevance, and new products with no clicks.
Run search in production. Latency and cost per thousand queries, A/B tests and interleaving, a fresh index without downtime, and the same ideas in a search server.
Three end-to-end case studies on real open data. Product search for a furniture shop, shopping queries with hard negatives, and search on an open retrieval benchmark. Complete at least one for the certificate; each one is a portfolio piece.
Complete at least one case study for the certificate. The others are optional. What each one covers
Use every module in one practice project. You collect evidence, and check it against the course outcomes.
Realistic projects from start to finish, on real data, with every output explained. Each one makes a portfolio piece. Complete at least one for the certificate; the others are optional.
Product search for a furniture shop2 hr 30 min
On the full WANDS dataset: a BM25 baseline, query understanding, hybrid search, a cross-encoder and LambdaMART, compared on NDCG@10 and latency; recommend a design for a 200 ms budget.
Shopping queries with hard negatives2 hr 30 min
On the Amazon Shopping Queries dataset: rank products with exact, substitute, complement and irrelevant labels; measure the cost of showing substitutes; compare English and Spanish slices.
Search a help centre or a scientific corpus2 hr 30 min
On an open retrieval benchmark with a permissive licence: lexical, dense and hybrid retrieval, re-ranking, and when the cheap method is good enough.
Loading the reviews…
Took this course? Sign in to review it. Learners can write a review after they complete 25% of the lessons.