backend/routers/metadata_backfill.py

@@ -376,7 +376,11 @@ def locations_search(
         if q_norm in l_norm:
             scored.append((l, 1.0))
             continue
-        score = svc.similarity(q_norm, l_norm)
+        # Use the location-specific scorer (token_set_ratio + multi-digit
+        # penalty) instead of WRatio — same reason as the cluster-match
+        # path: location names share too much boilerplate vocabulary for
+        # WRatio to discriminate reliably.
+        score = svc.location_similarity(q_norm, l_norm)
         if score >= 0.50:
             scored.append((l, score))
 

backend/services/metadata_backfill.py

@@ -162,6 +162,11 @@ def similarity(a: str, b: str) -> float:
     too short to fuzzy-match safely (see _MIN_FUZZY_LEN comment) AND the
     strings don't exact-match.  This guardrails the 'one common word
     inside a longer phrase' false positive.
+
+    USE FOR: project names (where typos like '1-80' vs 'I-80' should
+    still match).  For location names use `location_similarity()` —
+    WRatio is too lenient on the shared boilerplate vocabulary in
+    location strings ('Area', 'Loc', 'Bridge', 'Dam', etc.).
     """
     if not a or not b:
         return 0.0
@@ -172,6 +177,50 @@ def similarity(a: str, b: str) -> float:
     return rapidfuzz.fuzz.WRatio(a, b) / 100.0
 
 
+# Multi-digit penalty applied when two location names have completely
+# disjoint multi-digit numeric tokens (e.g. "87 Jenks" vs "68 Jenks").
+# Single-digit numbers ("Loc 1", "Area 2") are often shared coincidentally,
+# but address-style multi-digit numbers are strong identifiers — if they
+# differ, the locations are usually different physical places.
+_LOCATION_DIGIT_MISMATCH_PENALTY = 0.30
+
+
+def location_similarity(a: str, b: str) -> float:
+    """Stricter similarity score for location-name matching.
+
+    Location names share so much boilerplate vocabulary ('Area', 'Loc',
+    'Bridge', 'Dam') that rapidfuzz.WRatio inflates obvious mismatches.
+    Example: 'Area 2 - Brookville Dam - Loc 2 East' vs 'Area 1 - Loc 1 -
+    87 Jenks' scores 85.5 via WRatio despite being unrelated locations.
+
+    This scorer uses `token_set_ratio` as the base (sensitive to actual
+    word overlap, not just substring containment).  It then applies a
+    multi-digit penalty: if both strings contain 2+-digit numbers and
+    none overlap, subtract 0.30.  Catches the "same project, different
+    address-style identifier" case ('87 Jenks' vs '68 Jenks') that pure
+    token-set scoring still rates above 0.90.
+
+    Single-digit numbers ('Loc 1', 'Area 2') are excluded from the
+    penalty because they're often shared boilerplate ("Loc 1" in every
+    project) rather than discriminating identifiers.
+    """
+    if not a or not b:
+        return 0.0
+    if a == b:
+        return 1.0
+    if min(len(a), len(b)) < _MIN_FUZZY_LEN:
+        return 0.0
+
+    base = rapidfuzz.fuzz.token_set_ratio(a, b) / 100.0
+
+    multidigits_a = set(re.findall(r"\d{2,}", a))
+    multidigits_b = set(re.findall(r"\d{2,}", b))
+    if multidigits_a and multidigits_b and not (multidigits_a & multidigits_b):
+        base = max(0.0, base - _LOCATION_DIGIT_MISMATCH_PENALTY)
+
+    return base
+
+
 # ── Cluster + Suggestion dataclasses ───────────────────────────────────────────
 
 
@@ -572,15 +621,24 @@ async def _scan_clusters(
 def _find_best_match(
     candidate_norm: str,
     candidates: list[tuple[str, str]],   # (id, normalised_name)
+    *,
+    kind: str = "project",               # "project" | "location"
 ) -> tuple[Optional[str], Optional[float], str]:
     """Return (best_id, best_score, classification).
 
     classification ∈ {"exact", "fuzzy", "ambiguous", "no_match"}
+
+    The `kind` parameter selects the scorer.  Project matching uses
+    rapidfuzz.WRatio (lenient — catches typos like '1-80' vs 'I-80').
+    Location matching uses `location_similarity` (stricter — catches
+    boilerplate-shared-but-actually-different strings like 'Loc 2 - 68
+    Jenks' vs 'Loc 1 - 87 Jenks').
     """
     if not candidate_norm or not candidates:
         return None, None, "no_match"
 
-    scored = [(cid, similarity(candidate_norm, cnorm)) for cid, cnorm in candidates]
+    scorer = location_similarity if kind == "location" else similarity
+    scored = [(cid, scorer(candidate_norm, cnorm)) for cid, cnorm in candidates]
     scored.sort(key=lambda x: x[1], reverse=True)
     best_id, best_score = scored[0]
 
@@ -725,7 +783,7 @@ def _build_suggestion(db: Session, cluster: Cluster) -> Suggestion:
         )
         location_candidates = [(l.id, _normalise(l.name)) for l in location_candidates_objs]
         if cluster.location_norm:
-            loc_id, loc_score, loc_match = _find_best_match(cluster.location_norm, location_candidates)
+            loc_id, loc_score, loc_match = _find_best_match(cluster.location_norm, location_candidates, kind="location")
         else:
             loc_id, loc_score, loc_match = None, None, "create_new"
     else: