Matching dimensions and weights

Once a lead is qualified, the platform scores your own records against them and shows your team the best few. A matching dimension is one comparison in that score: one field on your records, one answer from the lead, and a way of comparing them.

Building a dimension

The match types

When either side is missing, the dimension scores 0.5 — the middle. Not zero. A record that has not filled in a field is unknown, not unsuitable, and scoring it zero would bury every incomplete record beneath every complete one regardless of fit.

How weights work

Weights are relative, not percentages. They are summed and each dimension takes its share, so weights of 3, 1 and 1 give the first dimension three-fifths of the score. You never have to make them add up to anything.

The final score is out of 100. What moves it is the ratio between your weights — doubling every weight changes nothing at all.

A useful starting point: give the one thing you would never compromise on a weight several times larger than everything else, leave the rest equal, and adjust after you have watched a dozen real matches.

Hard filters

A hard filter is different in kind from a weight. A dimension marked as a hard filter removes a failing record from consideration entirely — it does not score badly, it is not there.

Use one where the answer is genuinely disqualifying: outside the service area, unavailable on the date, does not hold the licence. Use a weight for everything else. Two hard filters that each rule out half your records leave you with a quarter of them, and the usual symptom of over-filtering is an empty matches list with no explanation.

Three details:

• A hard filter only fires when both sides have a value. An unknown never eliminates a record.
• RANGE and CATEGORY_OVERLAP hard filters can be tuned — how far outside the range, how much overlap is enough. EXACT and CONTAINS are all-or-nothing.
• CUSTOM_LLM cannot be a hard filter, and the product refuses to make it one. A model's judgement is the wrong thing to delete a record on: it is not reproducible, and a filter that silently drops a good record on a bad day is worse than a low score you can see. This is a deliberate rule, not a missing feature.

What your team actually sees

Records are scored, the ones that fail a hard filter are dropped, the rest are ranked, and the top few go to the model for a short written explanation of why each one fits this lead. Your team sees the score, the breakdown by dimension, and that explanation.

How many are shown is a setting in Configuration → Settings.

Records already rejected or finalised for that lead are skipped when matching runs again, so re-running never resurrects a decision your team already made.

Only records that are active and available are considered. A record marked allocated or on hold is out of the pool until you change it back — which is the intended way to take someone off the board for a fortnight.

Reading a bad result

• No matches at all — almost always a hard filter. Turn them off one at a time.
• Everything scores about the same — your weights are too even, or the dimensions all measure much the same thing.
• The wrong record is on top — look at the breakdown before changing weights. Usually one dimension is scoring 0.5 because a field is empty on your records rather than because the comparison is wrong.