Creating AI Applications Using Retrieval-Augmented Generation (RAG) Course

Lesson 1 • Anatomy of a RAG Pipeline

Breaks the pipeline into retriever, reader, and generator stages. Shows how data flows from raw documents to a final grounded response.

Lesson 2 • What RAG Is and Why It Matters

Defines RAG by contrasting it with standard LLM inference and parametric memory. Establishes the core motivation for grounding AI outputs in external knowledge.

Lesson 3 • RAG Variants and Taxonomy

Surveys naive, advanced, and modular RAG architectures. Helps students select the right variant for a given application context.

Lesson 1 • Chunking Strategies for Retrieval

Teaches fixed-size, semantic, and hierarchical chunking methods. Chunk design is the single biggest lever on retrieval accuracy.

Lesson 2 • Building a Reproducible Ingestion Pipeline

Combines all preprocessing steps into a versioned, testable pipeline. Reproducibility ensures consistent retrieval behavior across updates.

Lesson 3 • Loading Documents from Multiple Sources

Covers connectors for files, databases, APIs, and web sources. Establishes source-agnostic ingestion as the pipeline entry point.

Lesson 4 • Metadata Extraction and Enrichment

Shows how to attach structured metadata to chunks for filtered retrieval. Metadata enables hybrid search and fine-grained access control.

Lesson 1 • Fine-Tuning Embeddings for Your Domain

Covers contrastive learning, triplet loss, and domain-specific fine-tuning workflows. Fine-tuned embeddings significantly boost retrieval on specialized corpora.

Lesson 2 • Selecting the Right Embedding Model

Compares general-purpose, domain-specific, and multilingual embedding models. Selection criteria include domain fit, latency, and cost.

Lesson 3 • How Embedding Models Work

Explains encoder architectures, pooling strategies, and embedding dimensionality. Provides the technical basis for informed model selection.

Lesson 1 • Metadata Filtering and Faceted Search

Applies structured filters alongside vector similarity to narrow result sets. Filtering reduces noise and enforces access boundaries.

Lesson 2 • Hybrid Search: Dense and Sparse Retrieval

Combines dense vector search with sparse keyword search using reciprocal rank fusion. Hybrid search consistently outperforms either method alone.

Lesson 3 • Scaling and Managing Vector Stores

Addresses sharding, replication, and index refresh strategies for production scale. Operational practices ensure consistent latency under growing data volumes.

Lesson 4 • Indexing Algorithms Deep Dive

Covers HNSW, IVF, and product quantization indexing methods. Algorithm choice governs the speed-recall frontier for a given workload.

Lesson 1 • Diversity and Coverage in Retrieval

Applies maximal marginal relevance and clustering to reduce redundancy in retrieved chunks. Diverse context sets improve answer completeness.

Lesson 2 • Multi-Stage Retrieval Pipelines

Designs retrieval as a cascade: broad recall followed by precision-focused re-ranking. Staged pipelines balance latency and relevance at scale.

Lesson 3 • Contextual and Conversational Retrieval

Adapts retrieval to multi-turn conversations by tracking context and resolving coreferences. Enables coherent, session-aware RAG applications.

Lesson 4 • Query Understanding and Transformation

Covers query expansion, rewriting, and decomposition to improve retrieval recall. Better query representation is the cheapest retrieval improvement.

Lesson 1 • Faithfulness and Citation Prompting

Teaches prompts that instruct the model to cite sources and refuse unsupported claims. Faithfulness prompting is the primary defense against hallucination.

Lesson 2 • Context Injection Techniques

Covers stuffing, map-reduce, and refine patterns for injecting multiple chunks. Each pattern suits different context lengths and latency budgets.

Lesson 3 • Output Format Control

Uses structured output prompts to produce JSON, tables, and markdown reliably. Consistent formats enable downstream parsing and UI rendering.

Lesson 4 • Prompt Versioning and Testing

Applies prompt registries and A/B evaluation to manage prompt changes safely. Versioned prompts prevent regressions when models or data are updated.

Lesson 1 • Faithfulness and Answer Relevance Metrics

Measures whether answers are grounded in retrieved context and relevant to the query. These two metrics are the primary quality signals for RAG outputs.

Lesson 2 • Context Precision and Recall Metrics

Evaluates whether retrieved chunks are necessary and sufficient for correct answers. Context quality metrics diagnose retrieval rather than generation failures.

Lesson 3 • Iterative Improvement Workflows

Connects evaluation results to targeted fixes in chunking, retrieval, or prompting. Structured iteration prevents random tinkering and accelerates quality gains.

Lesson 4 • RAG Evaluation Frameworks

Introduces RAGAS, TruLens, and custom evaluation pipelines for holistic RAG assessment. Frameworks decompose quality into measurable, actionable sub-metrics.

Lesson 1 • Observability and Monitoring

Instruments RAG pipelines with traces, metrics, and logs for full operational visibility. Observability enables rapid incident response and continuous quality monitoring.

Lesson 2 • API Design and Integration

Covers REST and streaming API design for RAG endpoints consumed by front-end and backend clients. Well-designed APIs decouple the RAG core from consuming applications.

Lesson 3 • Safety, Guardrails, and Access Control

Implements input/output guardrails, PII redaction, and role-based document access. Safety controls protect users and organizations from harmful or unauthorized outputs.

Lesson 4 • Caching and Latency Optimization

Applies semantic caching, embedding caching, and response caching to reduce latency and cost. Caching is the highest-leverage production optimization for RAG.