Choose the right setup
Decide whether a project needs retrieval, fine-tuning, or both.
Start from the failure mode
The best setup depends on what would make the assistant fail.
If the failure would be missing or outdated facts, start with retrieval.
If the failure would be poor tone, weak structure, or inconsistent judgment, add fine-tuning.
Use retrieval when answers must come from sources
Retrieval is the right starting point when the assistant should answer from:
- product documentation
- policies and procedures
- support articles
- research notes
- internal playbooks
Retrieval improves factual grounding because the assistant can pull relevant source material into the answer flow.
Use fine-tuning when behavior matters more than lookup
Fine-tuning matters when you want the assistant to:
- sound like your team
- follow a specific answer format
- mirror strong example responses
- make more consistent choices across repeated situations
Fine-tuning improves how the assistant answers, not just what it can look up.
A practical default
If you are unsure, start with retrieval and a small source set.
Add fine-tuning when:
- the facts are mostly right but the delivery is weak
- answers miss your preferred structure
- similar prompts get noticeably inconsistent treatment