Applied Generative AI and Retrieval Augmented Generation

About this course

Ship a grounded question answering service over a private document set, with measured retrieval quality, guardrails and a cost budget you can defend.

This is an 8-week specialist track - our most in-depth program, taking you to a professional, hireable standard. Learn online at your own pace. Practice the concepts with the listed assessments.

What's included

• Self-paced lessons and practice (40 estimated hours).
• Hands-on coding quests you solve inside the EchoLens browser compiler - nothing to install.
• Gems, stages and a leaderboard that keep you moving instead of grade anxiety.
• A verified certificate with a scannable QR code, ready to share on LinkedIn, when you satisfy the course requirements.
• Completely free - no fee, just create an account and start.

What you will learn

Course outline - level by level

12 leveles, each with hands-on quests you clear in the portal.

• Level 1. Embeddings and Semantic Similarity - Analogy: Giving every book in a library a map coordinate so that books about the same subject end up on neighbouring streets. Covered: Dense vector representations, embedding model choice and dimensionality, cosine similarity against Euclidean distance, normalisation, batching and token cost, and why similarity is not relevance.
• Level 2. Chunking Strategies - Analogy: Cutting a legal manual into index cards along clause boundaries rather than slicing straight through the middle of a sentence. Covered: Fixed character splitting, recursive splitting on structure, token aware splitting, overlap sizing, markdown and code aware splitters, and attaching parent document context to each chunk.
• Level 3. Vector Databases and Metadata Filtering - Analogy: A filing system that pulls the right folder by subject, then lets you narrow to a single department and date range before it hands anything over. Covered: Approximate nearest neighbour indexes, HNSW parameters, collections and namespaces, metadata filters, upserts and deletions, and choosing between Chroma, Qdrant and pgvector.
• Level 4. Prompting, Grounding and Injection Defence - Analogy: Briefing an expert witness with the case file and telling them plainly to say 'not in the record' rather than guess. Covered: System instruction design, few shot examples, context grounding, citation formats, explicit refusal behaviour, and treating retrieved text as untrusted input rather than instructions.
• Level 5. LCEL Chains and Streaming - Analogy: An assembly line where each station hands its output to the next without anyone stacking parts on the floor in between. Covered: Runnable composition with the pipe operator, RunnableParallel and RunnablePassthrough, output parsers, streaming tokens, batching, and adding LangSmith tracing.
• Level 6. Hybrid Retrieval and Re-ranking - Analogy: A reference librarian who rewrites a vague question into three sharper ones, gathers everything, then puts the three best results on top. Covered: BM25 keyword search, reciprocal rank fusion, multi query generation, HyDE, and cross encoder re ranking with Cohere or a local model.
• Level 7. Function Calling and Tool Schemas - Analogy: Handing an assistant a calculator and a booking terminal, with written rules about which one to reach for. Covered: JSON tool schemas, argument validation, forcing or restricting tool choice, returning results to the model, parallel tool calls, and error handling when a tool fails.
• Level 8. ReAct Loops and Agent Control - Analogy: A detective working the board in cycles of observe, think, act, then observe again, with a rule that they stop after ten leads. Covered: Thought, action and observation loops, scratchpad management, step limits and timeouts, cost ceilings, and recognising when an agent is the wrong tool for a deterministic job.
• Level 9. LangGraph State Machines and Multi-agent Graphs - Analogy: A newsroom where the researcher gathers, the writer drafts and the editor sends work back until it is fit to print. Covered: Typed state schemas, nodes and conditional edges, cycles, checkpointing and resumption, human in the loop interrupts, and supervisor patterns.
• Level 10. Evaluation with Ragas and a Golden Set - Analogy: A bar exam that marks answers against the statute book, not against how confident the candidate sounded. Covered: Building a golden question set, faithfulness, answer relevancy, context precision and recall, the limits of model graded evaluation, and tracking scores across changes.
• Level 11. Guardrails, PII Redaction and Structured Output - Analogy: A scanner at the exit rather than the entrance, checking what is leaving the building as well as what came in. Covered: Input and output filtering, PII detection and redaction, topic restriction, schema constrained generation, refusal handling, and logging blocked events for review.
• Level 12. Serving, Caching, Cost and Observability - Analogy: A drive through that keeps the popular orders hot on the counter and prices every item before it leaves the window. Covered: Serving with FastAPI, semantic and exact match caching, rate limiting, streaming responses, per request token and cost accounting, and structured logging.

Who it's for

Applied Generative AI and Retrieval Augmented Generation suits learners at a intermediate to advanced level who want a practical, project-based route into Applied Generative AI and Retrieval Augmented Generation. You need only a browser and an internet connection - all coding runs inside the EchoLens compiler, so there is nothing to set up.

Certificate

Pass every required assessment at its stated threshold to earn your verified certificate. Optional practice and watching videos do not determine eligibility. Anyone can scan its QR code to verify it on our site. You can add it to your CV or share it to LinkedIn in one click.