RAG Course
Chunking, embeddings, hybrid search, reranking and evaluation for retrieval-augmented generation, with the failure modes that show up once real users arrive.
Chunking, embeddings, hybrid search, reranking and evaluation for retrieval-augmented generation, with the failure modes that show up once real users arrive.
- Hybrid search, not embeddings alone
- Reranking as a first-class step
- Retrieval evaluated on its own
- Permission-aware from the start
- Modules
- 8
- Format
- Online
- Mentorship
- Weekly 1:1
- Project
- Production
Why this track
What you get that a general AI course does not cover.
Retrieval evaluated separately
Most RAG failures are retrieval failures, but almost nobody measures retrieval on its own. You will build a retrieval eval set first, so you know whether a bad answer came from the retriever or the generator.
Hybrid search by default
Dense vectors find paraphrase. BM25 finds exact identifiers, error codes and proper nouns. Production systems use both, and you will build and tune the hybrid from module 3.
Reranking is not optional
A cross-encoder reranker over the top candidates is the single biggest quality gain available for modest latency cost. This is a core module rather than an optimisation exercise.
Written for messy real data
PDFs with broken tables, duplicated FAQs, conflicting versions and stale documents. Clean-document RAG tutorials are not the problem you will have at work.
Curriculum
Eight modules, each ending in something you can show.
- M01
Why RAG fails and where to look first
A taxonomy of failure: bad chunking, bad queries, bad ranking, bad generation. Learning to attribute an error before fixing it.
- Failure taxonomy
- Attribution workflow
- Building a query set
- M02
Chunking and document preparation
Fixed, semantic, hierarchical and parent-child chunking. Parsing real documents without losing structure.
- Chunking strategies
- Overlapping windows
- Table and layout parsing
- Metadata design
Tools
- Unstructured
- langchain
- marker
- M03
Embeddings and vector search
Choosing an embedding model on your own data, indexing at scale, and understanding where vector search stops being enough.
- Embedding model choice
- ANN indexes
- Filtering and metadata
Tools
- pgvector
- FAISS
- sentence-transformers
- M04
Hybrid search and query rewriting
BM25 alongside dense retrieval, reciprocal rank fusion, and rewriting vague user questions into something retrievable.
- BM25
- Reciprocal rank fusion
- Query rewriting
- Multi-query expansion
Tools
- rank_bm25
- Elasticsearch
- pgvector
- M05
Reranking and context assembly
Cross-encoder rerankers, deduplication, and assembling a context window that stays coherent instead of just top-k.
- Cross-encoder rerankers
- Deduplication
- Context assembly
- Lost in the middle
Tools
- Cohere rerank
- bge-reranker
- M06
Grounded generation and citations
Making the model answer only from retrieved context, refuse cleanly when the answer is absent, and cite the source span.
- Grounded prompting
- Refusal behaviour
- Span-level citations
- Answer verification
- M07
Retrieval evaluation
Recall@k, MRR and nDCG on a labelled set. Separating retriever quality from generation quality in CI.
- Recall@k
- MRR and nDCG
- Context precision
- Regression testing
Tools
- ragas
- deepeval
- M08
Production concerns
Freshness and re-indexing, access control per user, latency budgets, and observability for retrieval pipelines.
- Incremental indexing
- Permission-aware retrieval
- Caching layers
- Pipeline observability
Stack
- Langfuse
- Helm
- pg_cron
Roles this maps to
Where this track lines up with real job descriptions.
RAG Engineer
The direct match. You own the retrieval stack end to end and are judged on answer quality over a query set.
Search Engineer
A good fit if you enjoy ranking, relevance work and query understanding more than generation.
AI Backend Engineer
RAG systems are mostly backend systems with a model in the middle, and this track is weighted accordingly.
What you can do afterwards
Concrete capabilities, not a certificate.
- Build a hybrid retrieval pipeline with reciprocal rank fusion and a reranking stage
- Attribute a bad answer to the retriever or the generator using a labelled eval set
- Handle PDFs, tables and conflicting document versions without losing structure
- Add permission-aware retrieval so users never see documents they should not
Questions
Do I need a vector database to start?
No, and starting without one is instructive. Modules 2 and 3 use a simple in-process index so you can see how embeddings behave before adding infrastructure. pgvector and FAISS come in once there is a reason.
How do you decide between chunking strategies?
You measure. Module 7 has you run the same query set across at least three chunking strategies and pick on retrieval metrics, not on which one looks tidiest in a diagram.
Is this still relevant with long context models?
Yes, for a specific reason: retrieval still handles more documents than fit in a context window, and access control has to happen before the model sees anything. Module 7 covers where long context genuinely replaces retrieval, so you can tell when to use which.
What if my documents are messy?
Good, because that is the realistic case. Table extraction, duplicate detection and version conflicts each get explicit treatment in modules 2 and 8.
Related tracks
Ready when you are
Applications for the next cohort are open. You pick your track during the first month, so you can change if the material does not suit you.