Configure Search Center retrieval
Search Center retrieval settings control how much of the selected dataset can feed an answer. Change them with a repeatable question set, not by judging one response.
Choose the context mode
Use the full dataset as context for every answer is on by default. When it is turned off, Documents retrieved per search limits the records selected for each query.
| Dataset situation | Useful starting choice | What to watch |
|---|---|---|
| Small, tightly scoped collection | Full dataset | Conflicting or redundant records |
| Large documentation set | Limited retrieval | Missing a necessary supporting document |
| Answers combine several policies | Compare both modes | Conditions dropped from a multi-source answer |
| Dataset contains unrelated sections | Fix the scope first | Retrieval settings masking poor content selection |
The setting does not determine which source is authoritative. Remove obsolete or conflicting records from the dataset instead of expecting the model to resolve policy ownership.
Decide whether to remember searches
Remember previous searches between visits is off by default. This gives each visit a clean starting state.
Turn it on only when returning visitors benefit from continuity. Test whether an earlier query changes a later answer in a confusing way, especially when people share a browser.
Run a controlled comparison
Prepare questions in four groups:
- one-source facts;
- answers requiring two or more documents;
- similar terms from different content sections; and
- unsupported questions.
For each setting, record the answer, documents shown, missing conditions, and unsupported behavior. A fluent response is not enough; the expected sources must appear and support the claims.
Keep the simplest setting that passes the representative questions. Re-run the comparison when the dataset grows or its content boundaries change.