Emits candidate pairs in a flat, review-ready form so a human or an LLM can
apply a label (TRUE match / FALSE non-match). The result feeds two goals:
calibrating tier thresholds, and training the supervised combiner used by
np_score() method = "model". It is the same schema the refinement hand-off
will use.
Usage
np_label_frame(pairs, k = 3)
Arguments
- pairs
Scored + vetoed np_pairs.
- k
Number of top-scoring candidates to keep per query. Default 3.
Value
A data frame (class np_labelframe) with per-field similarities,
score, veto info, a candidate_type (why the pair was kept), and an empty
label column to be filled in.
Details
To make labelling efficient it keeps, per query, the top-k scoring pairs
plus the best name-only and best address-only pairs (so the eventual training
set contains hard negatives near the decision boundary, not just easy ones).