Architect a retrieval-augmented generation pipeline that grounds answers in verified sources.
Digital Transformation & Artificial Intelligence
Grounding, Citation and Hallucination Control in LLM Applications
Ground large language model applications in verified sources, generate reliable citations and detect hallucination before it reaches production users.
🗓️Duration5 training days
📚Content4 modules · 8 sessions
🎓On completionAccredited attendance certificate
About the programme
Course Overview
An AI application loses a user's trust the moment it states something false with total confidence, and retrieval alone does not guarantee that an answer stays faithful to its sources. This course is built around the practical craft of grounding: designing retrieval pipelines where chunking, embedding and re-ranking choices genuinely improve precision, generating citations that map each claim to a verifiable passage rather than a plausible-looking reference, and catching hallucination before it ever reaches a user. Chain-of-verification and self-consistency techniques are applied to a working retrieval-augmented generation pipeline, alongside automated faithfulness scoring and the human review it cannot fully replace. A dedicated module builds evaluation sets, measures retrieval quality and citation accuracy, and applies LLM-as-judge methods without over-trusting them. By the end, delegates can defend not just what an LLM application answered, but why the answer can be trusted and what should happen when it cannot be.
Expected Learning Outcomes
Design chunking, embedding and re-ranking strategies that improve retrieval precision.
Generate inline citations that map each claim to a specific, verifiable source passage.
Detect ungrounded claims using entailment checks, self-consistency and guardrail models.
Build evaluation sets that score faithfulness, completeness and citation accuracy.
Set confidence thresholds that route uncertain answers to human review instead of guessing.
Monitor hallucination and citation-error rates in production and trace them to root cause.
Who Should Attend
AI engineers and machine learning teams building retrieval-augmented generation applications.
Product managers responsible for the accuracy and trustworthiness of AI-generated answers.
Data scientists evaluating and monitoring large language model applications in production.
Knowledge management teams curating the source content an AI application retrieves from.
Risk and compliance staff assessing generative AI applications before customer-facing release.
Technical leads choosing between prompting, fine-tuning and guardrail approaches to faithfulness.
Course Modules
Select any module to see its sessions and points.
01Architecting Retrieval for Grounded Answers
2 sessions · 8 points
Session 1Retrieval Design: Chunking, Embeddings and Indexing
- Choosing chunk size and overlap strategies that preserve semantic units instead of splitting them apart.
- Selecting embedding models suited to domain vocabulary and multilingual content.
- Combining hybrid keyword and vector search to catch exact terms that embeddings alone miss.
- Refreshing and re-indexing a knowledge base on a schedule that matches how often source documents change.
Session 2Query Understanding and Multi-Step Retrieval
- Rewriting and expanding user queries to improve recall of relevant passages before generation.
- Decomposing multi-hop questions into sequential retrieval steps that each answer part of the query.
- Re-ranking retrieved passages with cross-encoder models before they reach the generation step.
- Setting context window budgets across the system prompt, retrieved passages and conversation history.
02Generating Reliable Citations and Source Attribution
2 sessions · 8 points
Session 1Designing Citation Formats and Attribution Logic
- Producing inline citations that map each claim to a specific retrieved span rather than a whole document.
- Distinguishing direct quotation from paraphrase in how a citation is presented to the user.
- Attaching confidence indicators to answers whose source support is weak or partial.
- Handling citation formatting when multiple sources partially agree or directly conflict.
Session 2Verifying Citations Match the Retrieved Evidence
- Running automated entailment checks between a generated claim and its cited passage.
- Flagging citations that point to the correct document but the wrong section or paragraph.
- Sampling generated answers regularly for manual citation audits by a subject expert.
- Tracing systematic citation errors back to a specific weakness in prompt or retrieval design.
03Detecting and Reducing Hallucination
2 sessions · 8 points
Session 1Techniques for Faithfulness and Self-Verification
- Applying chain-of-verification prompting so a model checks its own claims before the final answer.
- Using self-consistency sampling to flag answers that vary meaningfully across repeated generations.
- Constraining generation to retrieved content through structured prompts and enforced output formats.
- Routing low-confidence or weakly grounded answers to human review instead of letting the model guess.
Session 2Automated Detection and Guardrail Models
- Deploying a separate faithfulness-checking model that scores groundedness before an answer is shown.
- Applying natural language inference models to compare generated claims against source text directly.
- Setting thresholds that trigger a safe fallback response when grounding confidence falls too low.
- Logging flagged hallucinations for root-cause analysis by the retrieval or prompt engineering team.
04Evaluating, Monitoring and Governing Grounded Applications
2 sessions · 8 points
Session 1Building Evaluation Sets and Metrics
- Constructing a labelled evaluation set of questions with known correct sources and answers.
- Measuring retrieval quality with recall at k and mean reciprocal rank across test queries.
- Scoring generated answers for faithfulness, completeness and citation accuracy against the evaluation set.
- Using LLM-as-judge evaluation carefully, cross-checked against human spot checks on a sample.
Session 2Production Monitoring and Governance
- Tracking hallucination and citation-error rates on live traffic rather than only at launch.
- Alerting when retrieval coverage drops for a topic and answers start to go unsupported.
- Disclosing to users when an answer could not be fully grounded in a verified source.
- Documenting known failure modes and escalation paths for regulated or safety-critical use cases.
What the participant receives
4 course modules
A structured syllabus
8 training sessions
across 5 days
32 detailed points
Applied, detailed content
Accredited attendance certificate
On completing the programme
Complete your registration
We will contact you within one business day to confirm.
Ready to start?
Reserve your seat and start building the skill.