diff --git a/backend/src/main/java/org/raddatz/familienarchiv/document/DocumentService.java b/backend/src/main/java/org/raddatz/familienarchiv/document/DocumentService.java
index d25cf03c..73c60e43 100644
--- a/backend/src/main/java/org/raddatz/familienarchiv/document/DocumentService.java
+++ b/backend/src/main/java/org/raddatz/familienarchiv/document/DocumentService.java
@@ -669,6 +669,18 @@ public class DocumentService {
     public DocumentSearchResult searchDocuments(String text, LocalDate from, LocalDate to, UUID sender, UUID receiver, List<String> tags, String tagQ, DocumentStatus status, DocumentSort sort, String dir, TagOperator tagOperator, boolean undated, Pageable pageable) {
         boolean hasText = StringUtils.hasText(text);
 
+        // Pure-text RELEVANCE: push pagination + ts_rank ordering into SQL — skip
+        // findAllMatchingIdsByFts entirely (ADR-008). This must run BEFORE any
+        // findAllMatchingIdsByFts call so the fast path is preserved. An active undated
+        // filter must NOT take this path: it bypasses buildSearchSpec, so the
+        // undatedOnly predicate would be silently dropped. By definition this path has
+        // no date/sender/receiver/tag/status filters, and undated documents are valid
+        // FTS hits already folded into the ranked page, so there is no separate undated
+        // count to report here.
+        if (!undated && isPureTextRelevance(hasText, sort, from, to, sender, receiver, tags, tagQ, status)) {
+            return relevanceSortedPageFromSql(text, pageable);
+        }
+
         List<UUID> rankedIds = null;
         if (hasText) {
             rankedIds = documentRepository.findAllMatchingIdsByFts(text);
@@ -706,13 +718,8 @@ public class DocumentService {
                                            List<String> tags, String tagQ, DocumentStatus status,
                                            DocumentSort sort, String dir, TagOperator tagOperator,
                                            boolean undated, Pageable pageable) {
-        // Pure-text RELEVANCE: push pagination into SQL — skip findAllMatchingIdsByFts entirely (ADR-008).
-        // An active undated filter must NOT take this path: it bypasses buildSearchSpec, so the
-        // undatedOnly predicate would be silently dropped.
-        if (!undated && isPureTextRelevance(hasText, sort, from, to, sender, receiver, tags, tagQ, status)) {
-            return relevanceSortedPageFromSql(text, pageable);
-        }
-
+        // The pure-text RELEVANCE fast path is handled by the caller (searchDocuments)
+        // before findAllMatchingIdsByFts runs, so it never reaches here (ADR-008).
         Specification<Document> spec = buildSearchSpec(
                 hasText, rankedIds, from, to, sender, receiver, tags, tagQ, status, tagOperator, undated);