Knowledge Bases, Retrieval-Augmented Generation and Grounding
What this domain tests
This domain tests how knowledge bases supply grounded answers: ingestion, chunking, retrieval, answer generation, citation and fallback. The exam expects you to separate retrieval quality from generation quality.
Core rules
- A knowledge base is built from sources such as uploaded documents or connected repositories. Content must be current, owned and permissioned.
- Ingestion splits documents into chunks; chunk size and overlap affect retrieval. Tables, headers and FAQ pairs often need special handling.
- Retrieval finds candidate passages, optionally reranks them, and passes the best context to an answer step or LLM. Generation should be instructed to answer from that context.
- Grounding is not a guarantee. Use confidence thresholds, citation display and an explicit “I don’t know” path.
- Respect source permissions. A user must not receive a passage they could not open in the source system.
- Evaluate retrieval separately: did the right passage appear? Then evaluate answer faithfulness: does the answer match the passage?
Worked example
An HR assistant answers “How many parental leave weeks do I get?” Retrieval returns the parental-leave policy section and an older FAQ with a different number. The better-configured flow prefers the current policy, passes only that passage, and cites section 4.2. If no passage clears the threshold, it says it cannot confirm and links to HR. Reasoning: conflicting sources must be resolved by source authority and freshness, and a low-confidence fallback is safer than a confident wrong number.
Traps
Believing RAG eliminates hallucination; indexing confidential documents without access controls; using one giant chunk; ignoring stale duplicates; and testing only questions whose answers are easy to find. Also, do not confuse a knowledge answer with a transactional action that must call an API.