Fable findings in data

pipeline/utils/fuzzy_join.py

@@ -11,7 +11,12 @@ from tqdm import tqdm
 
 from pipeline.local_temp import local_tmp_dir
 
-_NUMBER_RE = re.compile(r"\d+")
+# A house-number token includes any letter suffix: 8A, 8B and plain 8 are
+# three different properties on the same street, so digit-only extraction
+# (which collapsed all three to "8") is not enough. Addresses are passed
+# through normalize_address_key first, so tokens are uppercase and
+# space-separated and [A-Z] suffices for the suffix.
+_NUMBER_RE = re.compile(r"\d+[A-Z]?")
 _POSTCODE_RE = r"^[A-Z]{1,2}\d[A-Z\d]?\d[A-Z]{2}$"
 # A house number is a strong disambiguator, so a numbered, number-compatible
 # pair may match on a lower address-similarity score than a number-less one
@@ -61,8 +66,10 @@ def fuzzy_join_on_postcode(
     columns (index, address, postcode) via projection pushdown, and the
     final join reads the remaining columns lazily.
 
-    Returns a LazyFrame with all left and right columns.  Unmatched rows
-    have null right columns.
+    Returns a LazyFrame with all left and right columns, plus a
+    ``_match_score`` (UInt8) audit column holding the token_sort_ratio of
+    the accepted match (exact matches score 100).  Unmatched rows have null
+    right columns and a null score.
     """
 
     tmpdir = tempfile.mkdtemp(prefix="fuzzy_join_", dir=local_tmp_dir())
@@ -152,14 +159,17 @@ def fuzzy_join_on_postcode(
         # Sort descending by score so best matches are assigned first
         all_pairs.sort(key=lambda t: (t[0], -t[1]), reverse=True)
 
-        matches: list[tuple[int, int]] = []
+        # Keep the score alongside each accepted pair: it is emitted as the
+        # _match_score audit column so downstream consumers can distinguish
+        # exact (100) from looser fuzzy matches.
+        matches: list[tuple[int, int, int]] = []  # (left_idx, right_idx, score)
         matched_left: set[int] = set()
         matched_right: set[int] = set()
 
-        for _score, left_idx, right_idx in all_pairs:
+        for score, left_idx, right_idx in all_pairs:
             if left_idx in matched_left or right_idx in matched_right:
                 continue
-            matches.append((left_idx, right_idx))
+            matches.append((left_idx, right_idx, score))
             matched_left.add(left_idx)
             matched_right.add(right_idx)
 
@@ -171,6 +181,7 @@ def fuzzy_join_on_postcode(
                 {
                     "_left_idx": pl.Series([m[0] for m in matches], dtype=pl.UInt32),
                     "_right_idx": pl.Series([m[1] for m in matches], dtype=pl.UInt32),
+                    "_match_score": pl.Series([m[2] for m in matches], dtype=pl.UInt8),
                 }
             )
         else:
@@ -178,6 +189,7 @@ def fuzzy_join_on_postcode(
                 {
                     "_left_idx": pl.Series([], dtype=pl.UInt32),
                     "_right_idx": pl.Series([], dtype=pl.UInt32),
+                    "_match_score": pl.Series([], dtype=pl.UInt8),
                 }
             )
 
@@ -197,18 +209,26 @@ def fuzzy_join_on_postcode(
 
 
 def _numbers_compatible(a: str, b: str) -> bool:
-    """Check that numeric tokens (flat/house numbers) in the shorter set are a subset of the longer.
+    """Check that the number tokens (house/flat numbers, including any letter
+    suffix) of two addresses are IDENTICAL sets.
 
-    Returns False if one address has numbers and the other doesn't.
+    Equality, not subset: subset logic let "188 GREAT NORTH WAY" absorb
+    "FLAT 1 188 GREAT NORTH WAY" ({188} is a subset of {1, 188}), attaching a
+    single flat's EPC facts to the whole building — tens of thousands of
+    wrong-property matches. Likewise digit-only tokens made "8A" and "8B"
+    both look like {8} and match each other (and plain "8"). Precision over
+    recall: a pair whose two sources genuinely disagree on number tokens is
+    safer left unmatched.
+
+    One side numbered, the other not -> incompatible. Neither numbered ->
+    compatible; such pairs are scored against the stricter no-numbers
+    threshold instead.
     """
     nums_a = set(_NUMBER_RE.findall(a))
     nums_b = set(_NUMBER_RE.findall(b))
-    smaller, larger = (
-        (nums_a, nums_b) if len(nums_a) <= len(nums_b) else (nums_b, nums_a)
-    )
-    if not smaller and larger:
-        return False
-    return smaller.issubset(larger)
+    if not nums_a and not nums_b:
+        return True
+    return nums_a == nums_b
 
 
 def _score_bucket(