Quality score

After Auto-Pilot finishes configuring your index it runs 10 validation queries and prints a score out of 100. The score is the average per-query score (0–10) × 10.

# How each query is scored

A test query is a real keyword (e.g. "alternator", "scaun tractor", "BK99651") sampled either from your catalogue or generated from your domain profile. Skryx runs it, then scores the result on:

1. Does it return any hits at all? Zero results = 0.
2. Is the first hit relevant? A keyword in the title gets full credit; a keyword only in the description gets partial credit.
3. How many of the top 5 hits are relevant? More = higher.
4. For exact-token queries (SKUs, model numbers): does the top hit match exactly? Yes = bonus.

The result is a number between 0 (broken) and 10 (perfect for that query). The score in the validation panel is the average across all 10 queries × 10, giving the familiar 0-100 scale.

# What good looks like

# When Auto-Pilot self-corrects

If the first validation pass scores under 70, Auto-Pilot:

1. Looks at which queries failed.
2. Bumps relevance bonuses (e.g. title_prefix, first_word) for the field types those queries should have matched.
3. Re-validates once.

If the second pass also scores under 70, Auto-Pilot stops and leaves the adjusted weights in place — the score you see is the post-adjustment number.

# Improving the score yourself

Three lever-by-lever ways to push a low score up:

1. Synonyms. If queries like "alternator" return wrong items, add alternator ↔ dinam ↔ generator auto in Synonyms.
2. Ranking rules. If discontinued products dominate, add a rule that penalises them.
3. Re-run Auto-Pilot. After adding a few hundred more products in one category, Auto-Pilot can detect a much sharper domain pattern.

You can find sample weak queries in the Auto-Pilot completion page's Validation results panel.