← Curriculum

G02 Stage C · Applied GenAI 56 h (18 T / 38 P)

Retrieval-Augmented Generation

Builds and measures a grounded retrieval system end to end.

Prerequisite
G01, F05
Modules
6
NOS
Statutory mapping

Tech stack

LangChainLlamaIndexpgvectorQdrantUnstructuredsentence-transformers

Modules

6 modules · 56 hours

M1 · RAG architecture

8 h
TOPICS
  • Why retrieval
  • Indexing versus query pipeline
  • Grounding and citation
  • Failure modes
  • When RAG is the wrong tool

Lab: Baseline RAG implementation end to end

Course material for this module is in production.

M2 · Ingestion and parsing

10 h
TOPICS
  • Document loaders
  • PDF/HTML/DOCX and table extraction
  • OCR
  • Layout preservation
  • Metadata extraction
  • Incremental sync
  • Deduplication
  • Refresh scheduling

Lab: Ingest a heterogeneous corpus including scanned and tabular documents

Course material for this module is in production.

M3 · Chunking strategy

10 h
TOPICS
  • Fixed, recursive, semantic, parent-document and hierarchical chunking
  • Overlap
  • Chunk size versus recall
  • Metadata design for filtering

Lab: Controlled chunking experiment measured on retrieval metrics

Course material for this module is in production.

M4 · Embeddings and vector stores

12 h
TOPICS
  • Embedding model selection and benchmarking
  • Dimensionality
  • Normalisation
  • Index types: flat, HNSW, IVF
  • Distance metrics
  • Filtering and hybrid metadata queries
  • Upserts
  • Sharding and scale

Lab: Build and tune a vector index for recall and latency

Course material for this module is in production.

M5 · Retrieval quality

10 h
TOPICS
  • Dense versus sparse (BM25)
  • Hybrid fusion
  • Reranking with cross-encoders
  • Query rewriting and expansion
  • Multi-hop retrieval
  • Top-k selection
  • Context assembly and ordering

Lab: Raise recall@k and precision with hybrid search plus reranking

Course material for this module is in production.

M6 · Generation layer

6 h
TOPICS
  • Grounded prompting
  • Citation enforcement
  • Context-window packing
  • Conflicting-source handling
  • Answerability and abstention

Lab: Citation-enforced answering with abstention on insufficient evidence

Course material for this module is in production.

Real-world work scenario

An internal policy assistant answers confidently from outdated documents. Add document versioning and recency metadata, implement hybrid retrieval with reranking, enforce citations, and prove the improvement with retrieval and answer metrics on a fixed evaluation set.

Assessment

Retrieval-tuning study with measured metrics · grounded answering service