Contracts have sections, clauses and subclauses. Which chunking suits them?
Why B is the answer
Hierarchical chunking
Hierarchical chunking respects parent and child structure.
In detail
- Hierarchical chunking follows sections and clauses.
- Search small, then read the larger context.
Use case
An illustrative scenario, not a real incident.
Contract clauses depend on their sections. Hierarchical chunking retrieves a small passage with its parent's context.
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
Rerankermindmap
root(("Contract"))
n1["Section"]
n2["Clause"]
n3["Subclause"]Key point: Hierarchical chunking respects structure.
Why the other options are wrong
- A. Fixed-size chunks of a few words, cut with no regard for structure
- Fixed chunks cut through clauses.
- C. One chunk holding the entire contract, whatever its overall length
- One huge chunk retrieves poorly.
- D. No chunking at all, with the raw file stored directly in the prompt
- A raw file in a prompt wastes tokens.