Search results contain several different ranking values, such as vector distance, a result's rank within one search channel, and the score used after combining channels. Callers cannot always tell which value they received or whether a larger or smaller number is better.
These values must not look like confidence percentages when they are only useful for sorting results.
What should change:
- Give each returned value a clear name and meaning.
- Include the distance metric and state whether lower or higher values rank better.
- If VidXP converts a raw distance into another score, identify the conversion used.
- Keep the rank from each search channel separate from the final combined rank and score.
- Show which search channels were run and which ones contributed to the returned moment.
- Mark uncalibrated scores as
ordering_only; only call a value confidence or probability after it has been measured and calibrated.
- Keep the same meaning across CLI, API, MCP, stored job results, and evidence artifacts.
Use #76 for calibration and end-to-end ranking evaluation.
Search results contain several different ranking values, such as vector distance, a result's rank within one search channel, and the score used after combining channels. Callers cannot always tell which value they received or whether a larger or smaller number is better.
These values must not look like confidence percentages when they are only useful for sorting results.
What should change:
ordering_only; only call a value confidence or probability after it has been measured and calibrated.Use #76 for calibration and end-to-end ranking evaluation.