Knowledge Integration & Data Handling: NCP-AAI Domain 6 (10%)

What this domain covers

Knowledge Integration and Data Handling is about connecting an agent to information it was not trained on — documents, databases, APIs — and using that information reliably. The dominant pattern is retrieval-augmented generation (RAG). At 10%, it ties with Cognition, Planning, and Memory as the heaviest domain outside the top four.

Retrieval-augmented generation (RAG)

RAG grounds the model in external data: you retrieve passages relevant to the query, put them in the prompt, and have the model answer from them. It is how agents stay current and cite sources without retraining. The core RAG quality question is whether the right context was retrieved — if retrieval misses, the model has nothing accurate to work from.

query -> embed -> search vector store -> top-k passages
      -> put passages in prompt -> model answers from them
# Garbage retrieval in -> wrong / hallucinated answer out

Embeddings, chunking, and vector search

Documents are split into chunks, each converted to an embedding (a vector capturing meaning) and stored in a vector database. At query time the question is embedded and the nearest chunks are retrieved by similarity. Chunk size is a real trade-off: chunks too large dilute relevance and waste context; too small lose surrounding meaning. Hybrid search — combining keyword and vector — often beats either alone.

Chunk too large:  retrieves noise, wastes context tokens
Chunk too small:  loses surrounding meaning, fragments facts
Hybrid search:    keyword + vector — catches exact terms AND meaning

Retrieval quality and grounding

Grounding means the answer is supported by retrieved evidence, not the model's parametric guesswork. Techniques that improve it: re-rank retrieved passages by relevance, instruct the model to answer only from the provided context and say it does not know otherwise, and return citations so claims are traceable. These are the main defenses against hallucination in agentic systems.

Data handling: freshness, structure, and governance

Beyond retrieval, agents must handle data responsibly: keep the knowledge base fresh (stale data produces confidently wrong answers), query structured sources such as SQL and APIs as well as unstructured text, and respect access control so an agent never retrieves data the user is not allowed to see. Data quality and permissions are part of this domain, not an afterthought.

Exam tip

When a RAG scenario gives wrong answers even though the information exists in the knowledge base, suspect retrieval — chunking, embeddings, or ranking — before blaming the model. And remember the grounding rule: answer from retrieved context and admit uncertainty, rather than letting the model fill gaps from its own memory.