Users report that your Copilot Studio agent frequently returns answers that are technically correct but cite only a small fragment of a larger document, missing important context from surrounding paragraphs. Which knowledge source configuration adjustment should you evaluate?

When Copilot Studio indexes knowledge source documents, it divides them into chunks — segments of text that are individually embedded and stored for semantic retrieval. The chunk size determines how much text is in each segment. If chunks are too small, a retrieved chunk may contain the direct answer but lack the surrounding context that gives it meaning.

Adjusting the chunk size or context window settings for the knowledge source to include more surrounding content per retrieved segment is the right configuration to evaluate. Configuring larger chunk sizes, or adjusting context window settings to retrieve adjacent chunks along with the primary match, ensures the LLM has more surrounding context when synthesizing its answer. This addresses the 'technically correct but lacking context' symptom.

Increasing the number of knowledge sources to provide more diverse content adds breadth, not depth of context. The problem is not lack of sources but insufficient context per retrieved chunk.

Switching from generative answers to classic topic-based answers for all responses abandons the generative capability entirely. Classic topics require manually authored responses — not a scaling solution.

Reducing the confidence threshold so the agent returns more results causes the agent to return lower-confidence, potentially less relevant content. It does not address the context window problem and may introduce noise.

Chunking is an advanced knowledge source tuning consideration. On the exam, recognize: partial/context-light answers → investigate chunk size and context window. Large document collections with dependent concepts may need larger chunks or overlap between chunks.