Answers lose context because long sections were cut at a fixed character count. What do you change?
Why B is the answer
The chunking strategy, so it respects document structure
Structure-aware chunking keeps related text together, which improves retrieval.
In detail
- A fixed character cut can split a section in the middle of an idea.
- Chunking that respects document structure keeps related text together.
- Check real chunks by eye before you change the embedding model.
Use case
An illustrative scenario, not a real incident.
A manual is cut every 1,000 characters, so a warning lands in one chunk and its condition in the next. Chunking by section keeps them together, and answers stop missing the condition.
See it in different ways
flowchart LR A["Larger chunks"] --> C["Fewer records"] B["Less overlap"] --> C D["Smaller chunks"] --> F["More records"] E["More overlap"] --> F
stateDiagram-v2 [*] --> Raw Raw --> Text: parse or OCR Text --> Clean: remove extraneous content Clean --> Chunked: split Chunked --> Governed: Delta table in Unity Catalog Governed --> Indexed: Vector Search Indexed --> [*]
flowchart TD
Q{"Do the documents have clear structure?"}
Q -->|"Yes"| A0["Split by headings and sections"]
Q -->|"No"| A1["Fixed size with overlap"]Key point: Chunk by structure, not by a fixed count.
Why the other options are wrong
- A. The LLM's temperature, so it fills gaps in the context
- A higher temperature can only make up what is missing.
- C. The Unity Catalog owner of the index
- The owner of an index has no effect on its content.
- D. The serving endpoint size and its scale-out limit
- Endpoint size affects capacity, not what a chunk contains.