Be able to design a retrieval system end to end, decide how to cut documents up, choose between exact terms and meaning, pick an index and a shortlist size, and measure whether the whole thing is actually finding what people asked for

Finding Things by Meaning

Retrieval is the part of most systems that quietly decides whether an answer is any good. This course covers cutting documents up, what gets compared, searching without looking at everything, and how to tell whether it works.

8 lessons, written and corrected before you arrived. Reading them here needs no account. The first reads the whole way through; the others open and then stop, because a page nobody owns cannot tell who is reading it. Starting the course gives you your own copy, where every idea has problems standing under it and you can ask about any sentence.

Start reading

01The Step That Decides Everything After ItA retrieval system does not answer anything. It hands over a short list of candidates, and the quality of that list is a ceiling on everything that happens next.
02The Decision Made Before Anything Is Searchedopening onlyNothing is ever retrieved. Pieces of documents are, and the choice of where to cut sets a limit on what can be found that no later stage can lift.
03Neither Side of the Comparison Is Textopening onlyA question is never matched against a document. Both are turned into lists of numbers by a model trained to put related things near each other, and everything follows from that.
04The Two Methods Fail on Opposite Questionsopening onlyMatching on meaning finds paraphrases and loses part numbers. Matching on words does the reverse. Running both and merging the lists is close to free and fixes most of each.
05You Cannot Compare the Question Against Everythingopening onlyComparing a question against every piece is simple, correct and impossible at scale. The structures that avoid it have build times, memory costs and rules about changing them.
06One Knob, and What It Is Really Selling Youopening onlyEvery approximate index has a single dial trading accuracy against time. Where to set it is an arithmetic question, and the answer is usually lower than instinct suggests.
07The Stage That Can Afford to Be Slowopening onlyThe first stage compared two representations made in ignorance of each other. A second pass that reads the question and the piece together is far better, and only affordable on a short list.
08A Hundred Questions You Wrote Yourselfopening onlyWithout a set of questions whose answers you know, every opinion about a retrieval system is an anecdote. Building that set is a day of work and it changes what you do next.