Retrieved chunks are relevant but often cut a clause in half, so answers miss conditions. What do you change first?
Why A is the answer
Use a chunking strategy that follows document structure
If the chunk boundary loses meaning, no model can restore it. Fix chunking before changing the model.
In detail
- Fix chunks first.
- Follow document structure, so clauses stay whole.
- Check chunks by reading them.
Use case
An illustrative scenario, not a real incident.
A liability clause splits across two chunks, so the exception is lost. Chunking by document structure keeps it together.
See it in different ways
flowchart TD
A{"Document type"} -->|"Uniform short text"| B["Fixed-size chunking"]
A -->|"Clear sections and hierarchy"| C["Hierarchical chunking"]
A -->|"Meaning shifts within the text"| D["Semantic chunking"]mindmap
root((RAG))
Ingest
Sync data source
Detect personal data
Chunk
By structure
Hierarchical
Search
Vector
Keyword
Rank
Rerankerflowchart TD
Q{"Do chunks cut clauses?"}
Q -->|"Yes"| A0["Chunk by structure"]
Q -->|"No"| A1["Look elsewhere"]Key point: Fix chunk boundaries first.
Why the other options are wrong
- B. Increase the temperature so the model fills in the missing words
- Higher temperature invents words.
- C. Add more embedding dimensions to capture finer meaning
- More dimensions do not change where text was cut.
- D. Move to a bigger foundation model that reasons over partial text
- A bigger model cannot read missing text.