Module 05 · Agent

Memory and RAG

Give an agent knowledge it was never trained on. Chunking, embeddings, a retriever allowed to return nothing, the pipeline that grounds an answer in what it found, and the separate problem of remembering the conversation.

5 lessons · 4 graded exercises·5.5 hours·intermediate·Prerequisite: Agent frameworks

Why retrieval

25 min

The model knows its training data and nothing about your documents. Retrieval is one of three ways to fix that, and the only one that hands you a citation and a prompt whose size you control.

Chunking

40 min · exercise

The chunk is the unit you retrieve and the unit you paste into a prompt, so where the boundaries land decides what the model gets to read. Strategies, overlap, and the argument about size.

Embeddings and similarity

40 min · exercise

What a vector actually is, why cosine similarity is a dot product, what the hashed embeddings in this course can and cannot do, and why a retriever has to be allowed to return nothing.

The RAG pipeline

50 min · exercise

Ingest, retrieve, assemble, generate. The assembly step in the detail it deserves, why a grounding instruction needs a score floor behind it, and why retrieval quality and answer quality are two different measurements.

Conversation memory

40 min · exercise

Retrieval is memory of documents; this is memory of the conversation. Four strategies and what each loses, the two things that must never be trimmed, and the corruption a naive trim causes.