Module 3 of How AI works. About 15 minutes.

Retrieval

A model answers from what it learned in training, unless you give it the right text when it answers. That is retrieval-augmented generation, or RAG, shown here step by step on a clinical record.

A model does not know your patient. Retrieval finds the parts of a record that bear on a question and hands them to the model with the question, so the answer can come from the record and point back to it. This module follows one clinical note through the three steps.

Chunks and meaning

First the note is split into chunks: short passages that each make sense on their own. Each chunk is turned into a list of numbers, called an embedding, that places it on a map by meaning.

Step through it and watch where each chunk lands. Chunks that mean similar things sit close together, even when they share few words.

A simulation: the positions are placed by hand to show the idea, and the clinical note is illustrative.

Finding the right chunks

A question is turned into an embedding the same way and lands on the same map. The chunks nearest to it are pulled out and passed to the model.

Try each question, then try the third: "Any history of kidney disease?" The closest chunk is a two-year-old line that rules kidney disease out. Similar is not the same as relevant, and the search cannot tell the difference.

A simulation with hand-set similarity scores. The clinical note is illustrative.

Answering with sources

Now compare the answer with and without retrieval. Without it, the model writes a fluent, plausible answer that is wrong for this patient. With it, the answer is built from the retrieved chunks and cites them. Tap a citation to check the chunk it came from.

A simulation with scripted answers, and no model is called. The clinical note is illustrative.

What to take away

Retrieval gives a model the right text to work from, and citations let you check its answer against the record. Retrieval can still pull the wrong chunk, so a citation is where your checking starts. To build one yourself, the Building a RAG system course does it step by step in notebooks.

Where next

How AI works: Module 4: Agents

Back to your lesson: Building a RAG system, Module 2: Documents, chunking and embeddings

Back to your lesson: Building a RAG system, Module 3: Retrieval

Back to your lesson: Building a RAG system, Module 4: Generation with citations

A shared question channel is on the way. When it opens, each answer will be written once and shared with everyone taking the course.