Level 4 of 14 · about 5 weeks

Retrieval and RAG for agents

Ground a model in your own documents, measure how well retrieval works, then let an agent decide what to look up.

Modules

1. 1

   Parsing and chunking

   Turn messy documents into clean pieces a retriever can find and a model can use

   Core 5 parts · about 5 hours

2. 2

   Embeddings and vector indexes

   Search by meaning with vectors, and why large collections use approximate indexes

   Core 5 parts · about 5 hours

3. 3

   Hybrid search and reranking

   Combine keyword and vector search, then reorder the shortlist with a more careful scorer

   Core 4 parts · about 5 hours

4. 4

   Query transformations

   Rewrite, expand, split and route the question before it reaches the index

   Common 5 parts · about 4 hours

5. 5

   Grounding and citations

   Make the model answer from the sources, show where each claim came from, and check it

   Core 5 parts · about 5 hours

6. 6

   Reducing hallucination

   Why models make things up even with sources in front of them, and the controls that help

   Core 4 parts · about 5 hours

7. 7

   Retrieval metrics

   Precision, recall, MRR, MAP and nDCG, and how to build a test set you can trust

   Core 4 parts · about 5 hours

8. 8

   Agentic RAG and GraphRAG

   When one search is not enough: retrieve in a loop, or retrieve over a graph of connected facts

   Common 4 parts · about 5 hours

9. 9

   Keeping an index fresh

   Detect what changed, update only that, handle deletions safely, and survive an embedding-model change

   Common 5 parts · about 4 hours

10. 10

    Custom embedding models (optional)

    When the vocabulary gap survives everything else, train the embedding model on your own pairs, and check what it cost you

    Specialist 4 parts · about 5 hours