Grounding measures how closely a prompt, agent profile, or response fits the relevant rubric criteria, objectives, context, and instructions. A well-grounded result stays on topic and meets the specified standards; a poorly grounded result misses them. section 9.5
A practical grounding workflow includes:
- Define the purpose, objectives, and rubric criteria for the prompt or profile. section 9.5
- Score each criterion, using summed or weighted totals when appropriate. section 9.5
- Evaluate recommendations manually or with a second LLM prompt, then review results for consistency and iterate on the rubric. section 9.5 section 9.6
Grounding a recommender
A recommender produces three items for specified subjects, formats, and genres. An evaluation prompt scores each recommendation against the rubric; an example output rates The Butterfly Effect, Primer, and Time Bandits on subject, format, and genre. Prompt flow can run the evaluation automatically after generation. section 9.5 section 9.6
Table 9.2 Rubric ratings
|
Rating
|
Description
|
|---|---|
|
1
|
Poor alignment: this is the opposite of what is expected given the criteria.
|
|
2
|
Bad alignment: this isn’t a good fit for the given criteria.
|
|
3
|
Mediocre alignment: it may or may not fit well with the given criteria.
|
|
4
|
Good alignment: it may not align 100% with the criteria but is a good fit otherwise.
|
|
5
|
Excellent alignment: this is a good recommendation for the given criteria.
|
Use the rating table to distinguish poor, bad, mediocre, good, and excellent alignment. section 9.5
