AI and Generative AI Foundations
Free CSA Trusted AI Safety Expert (TAISE) practice — 6 questions on AI and Generative AI Foundations, with explanations. No sign-up. Full 12-question mixed test →
A legal firm deployed a RAG chatbot over its contract repository. Users report the bot confidently cites clauses from contracts that don't exist, even though manual testing confirms the retrieval step consistently returns the correct top-k chunks containing the relevant clauses. Which issue is MOST likely causing this behavior?
This is a classic RAG 'groundedness' failure: even when retrieval returns the correct, relevant chunks, the generation step can still ignore or distort that context and fabricate plausible-sounding but false content. RAG improves access to relevant data but does not eliminate hallucination — it reduces the likelihood when the model is well-grounded, but doesn't guarantee faithfulness to the provided context.
An enterprise needs a system that answers employee questions using internal HR policies that are updated weekly, and every answer must cite the exact source policy document. The solution must avoid retraining a model each week. Which architecture BEST meets these requirements?
RAG decouples knowledge updates from the model itself — re-indexing the vector store when policies change is fast and cheap, and the retrieved chunks naturally support citing the exact source document, satisfying both the freshness and citation requirements without retraining.
A RAG pipeline for a medical knowledge base is configured with a chunk size of 2048 tokens and no overlap between chunks. Retrieval frequently returns chunks that are missing surrounding context needed to correctly answer multi-sentence clinical questions, because relevant information is split across chunk boundaries. What is the BEST configuration fix?
Adding overlap between chunks and using semantic/structure-aware chunking (splitting at sentence or paragraph boundaries rather than arbitrary token counts) directly addresses information being cut off at chunk boundaries, improving retrieval completeness without diluting embedding relevance.
What architectural innovation in the Transformer enables it to process all input tokens in parallel during training, rather than sequentially as in RNNs?
Self-attention computes pairwise relevance scores across all tokens in a sequence at once (via query, key, and value matrices), eliminating the step-by-step dependency chain that RNNs require and enabling full parallelization during training — this is the core innovation introduced in 'Attention Is All You Need.'
An AI agent has access to a calculator tool, a web search tool, and a code execution sandbox to answer multi-step quantitative questions. During testing, the agent sometimes invokes the code sandbox with syntactically invalid code, then — instead of reporting the execution error — fabricates a plausible-looking numeric answer and presents it with high confidence. From a TAISE AI safety perspective, what is the PRIMARY risk this behavior demonstrates?
This scenario illustrates a key agentic AI safety risk covered in TAISE: when a tool call fails or returns an error, an agent can mask that failure by generating a fabricated result and presenting it as fact, compounding the hallucination problem with false confidence — a significant risk when tool outputs are not explicitly verified before being surfaced to users.
A company wants to deploy a GenAI solution that autonomously plans multi-step tasks, invokes external APIs, maintains memory of intermediate results across steps, and adjusts its plan dynamically based on those results. Which category of GenAI application BEST describes this system?
Autonomous multi-step planning, tool/API invocation, persistent memory across steps, and dynamic replanning based on intermediate outcomes are the defining characteristics of agentic AI systems, which orchestrate an LLM as a reasoning/planning engine rather than treating it as a single-shot responder.
Ready for the real thing?
The full course has two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed answer explanations.
Start my full course on Udemy →