fix: KB dashboard search was applying the strict fiscal name-match rule (used to authorize automatic writes) to human browsing, so a natural query (e.g. missing the unit word) found nothing; now a separate, lenient, global, paginated server-side search — filter over the whole KB, then count/order/paginate. 1.3.6 (Hub only)

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

central/app/enrichment/observability.py

@@ -100,23 +100,50 @@ def _candidate_cards(session, products) -> list[dict]:
     return out
 
 
-def kb_search(session, query: str, limit: int = 6) -> list[dict]:
-    from ..models import KbProduct
-    from . import kb as kbm
+def kb_search_ids(session, query: str) -> list[int]:
+    """GLOBAL, server-side match over the whole KB (never limited to a rendered page): GTIN exact, or every normalized query word present somewhere among the
+    product's known names (any alias row OR its own normalized_description - words can come from different aliases of the same product). Deliberately lenient
+    (a human exploring the KB, e.g. "COCA COLA LATA 350" without the unit, must still find "COCA-COLA ORIGINAL LATA 350 ML"); this is DIFFERENT from
+    `kb.alias_match`, the strict rule that authorizes an automatic fiscal write, which stays untouched."""
+    import re
+
+    from ..models import KbKnowledge, KbProduct
+    from . import kb as kbm, textnorm
     q = (query or "").strip()
     if not q:
         return []
     digits = "".join(ch for ch in q if ch.isdigit())
-    products = []
-    if len(digits) >= 8 and len(digits) == len(q.replace(" ", "")):
+    if len(digits) >= 8 and re.fullmatch(r"[\d\s.\-]+", q):
         p = kbm.find_by_gtin(session, digits.zfill(14))
-        products = [p] if p else []
-    else:
-        products = kbm.name_candidates(session, q)[:limit]
-    return _candidate_cards(session, products)
+        return [p.id] if p else []
+    tokens = [t for t in textnorm.tokens(q) if t not in textnorm.MARKETING]
+    if not tokens:
+        return []
+    ids: set[int] | None = None
+    for token in tokens:                                        # SQL ILIKE narrows the candidates (portable: Postgres in production, SQLite in tests); the exact whole-word
+        like = f"%{token}%"                                      # check happens in Python so "ML" never matches inside "SIMPLES", etc.
+        matched = {pid for pid, val in session.execute(select(KbKnowledge.kb_product_id, KbKnowledge.value).where(KbKnowledge.knowledge_key == "alias", KbKnowledge.value.ilike(like))) if token in (val or "").split()}
+        matched |= {pid for pid, val in session.execute(select(KbProduct.id, KbProduct.normalized_description).where(KbProduct.normalized_description.ilike(like))) if token in (val or "").split()}
+        ids = matched if ids is None else (ids & matched)
+        if not ids:
+            return []
+    return sorted(ids)
 
 
-def kb_summary(session, query: str | None = None) -> dict:
+def kb_search(session, query: str, *, page: int = 1, page_size: int = 10) -> dict:
+    """query/filter GLOBAL over the whole KB -> count -> order -> pagination (never the other way round)."""
+    from ..models import KbProduct
+    ids = kb_search_ids(session, query)
+    total = len(ids)
+    page = max(1, page)
+    pages = max(1, -(-total // page_size))
+    page = min(page, pages)
+    page_ids = ids[(page - 1) * page_size: page * page_size]
+    products = [session.get(KbProduct, i) for i in page_ids]
+    return {"results": _candidate_cards(session, products), "total": total, "page": page, "pages": pages, "page_size": page_size}
+
+
+def kb_summary(session, query: str | None = None, page: int = 1) -> dict:
     """Compact Knowledge Base panel: known products, statuses of the NCM knowledge, hits today, and the latest conflicts with their evidence (no dashboard of its own)."""
     from ..models import KbKnowledge, KbProduct
     today = datetime.now(timezone.utc).strftime("%Y-%m-%d")
@@ -134,4 +161,5 @@ def kb_summary(session, query: str | None = None) -> dict:
                           "candidates": [{"ncm": i.value, "clients": sorted({c for c in (i.evidence or {}).get("clients", [])}), "documents": sum(int(l.get("docs", 0)) for l in ((i.evidence or {}).get("lineages") or {}).values()),
                                           "providers": sorted(f for f in ((i.evidence or {}).get("families") or {}) if f.startswith("PROVIDER:"))} for i in items]})
     return {"products": products, "verified": by.get("VERIFIED_KB", 0), "corroborated": by.get("CORROBORATED", 0) + by.get("EXTERNAL_CORROBORATED", 0), "observed": by.get("OBSERVED", 0),
-            "conflicts": by.get("CONFLICT", 0), "stale": by.get("STALE", 0), "hits_today": hits, "recent_conflicts": conflicts, "search": kb_search(session, query) if query else []}
+            "conflicts": by.get("CONFLICT", 0), "stale": by.get("STALE", 0), "hits_today": hits, "recent_conflicts": conflicts,
+            "search": kb_search(session, query, page=page) if query else {"results": [], "total": 0, "page": 1, "pages": 1, "page_size": 10}}