How do you choose a chunking strategy?
Why C is the answer
From model and retrieval evaluation
The right chunk size depends on your documents and model, so measure retrieval and answers rather than guess.
In detail
- Evaluation shows which chunking strategy works best.
- Try options, then compare retrieval and answer quality.
- Keep a fixed test set, so the comparison stays fair.
Use case
An illustrative scenario, not a real incident.
A team tries three chunk sizes on its manuals and measures retrieval and answers. The middle size wins, so that is what ships.
See it in different ways
flowchart LR A["Source documents"] --> B["Extract text"] B --> C["Filter extraneous content"] C --> D["Chunk"] D --> E["Delta table in Unity Catalog"] E --> F["Vector Search index"]
sequenceDiagram participant E as Experiment participant M as Candidates participant D as Decision E->>M: Same data and tasks M-->>E: Metrics E->>D: Compare cost, latency and quality D-->>E: Choose, and record why
flowchart LR S0["Try strategies"] --> S1["Index each"] --> S2["Evaluate"] --> S3["Keep the best"]
Key point: Choose chunking from evaluation.
Why the other options are wrong
- A. From the price of the embedding model
- Embedding price does not show what retrieves well.
- B. From the size of the cluster that runs the indexing job
- Cluster size affects speed, not retrieval quality.
- D. From a fixed size used in a tutorial
- A tutorial's size may not suit your documents.