diff --git a/core/src/main/java/org/opensearch/sql/executor/analyze/ExpensiveSortRule.java b/core/src/main/java/org/opensearch/sql/executor/analyze/ExpensiveSortRule.java
index 0db47504df3..fb93e3e8ff8 100644
--- a/core/src/main/java/org/opensearch/sql/executor/analyze/ExpensiveSortRule.java
+++ b/core/src/main/java/org/opensearch/sql/executor/analyze/ExpensiveSortRule.java
@@ -29,7 +29,11 @@ public List<Recommendation> apply(ProfileView view) {
       return recommendations;
     }
     for (PlanNode node : view.planNodes()) {
-      if (!node.getNode().toLowerCase(Locale.ROOT).contains("sort")) {
+      // Match standalone sorts and bounded sorts fused with a limit. A top-level PPL sort is
+      // typically merged with the query-size-limit into CalciteEnumerableTopK, whose name contains
+      // no "sort" -- but a TopK is always a sort (it extends EnumerableLimitSort), so it qualifies.
+      String name = node.getNode().toLowerCase(Locale.ROOT);
+      if (!name.contains("sort") && !name.contains("topk")) {
         continue;
       }
       long rowsIn = ProfileView.rowsIn(node);
diff --git a/core/src/test/java/org/opensearch/sql/executor/analyze/AnalyzeRecommendationBuilderTest.java b/core/src/test/java/org/opensearch/sql/executor/analyze/AnalyzeRecommendationBuilderTest.java
index 95624ee161f..ded1b20f3e0 100644
--- a/core/src/test/java/org/opensearch/sql/executor/analyze/AnalyzeRecommendationBuilderTest.java
+++ b/core/src/test/java/org/opensearch/sql/executor/analyze/AnalyzeRecommendationBuilderTest.java
@@ -99,6 +99,21 @@ void expensiveSortFiresOnLargeSlowSort() {
     assertTrue(r.getMessage().contains("30% of execution"));
   }
 
+  @Test
+  void expensiveSortFiresForFusedTopKNode() {
+    // A top-level sort fuses with the query-size-limit into CalciteEnumerableTopK (no "sort" in
+    // the name). It is still a sort, so it must trip the rule: 30ms self-time of 100 (30% > 20%),
+    // 60k input rows (> 50k).
+    PlanNode scan = new PlanNode("EnumerableMergeJoin", 10.0, 60_000, null);
+    PlanNode topK = new PlanNode("CalciteEnumerableTopK", 40.0, 100, List.of(scan));
+    List<Recommendation> recs = new AnalyzeRecommendationBuilder(profile(1, 100, topK)).build();
+
+    Recommendation r = ruleOf(recs, "Expensive Sort").orElseThrow();
+    assertEquals(RecommendationSeverityLevel.WARNING, r.getSeverity());
+    assertEquals("CalciteEnumerableTopK", r.getAffected_node());
+    assertTrue(r.getMessage().contains("Sorting 60000 rows"));
+  }
+
   @Test
   void expensiveSortSilentWhenSelfTimeIsSmall() {
     // sort cumulative 90ms but its child took 89ms -> self-time only 1ms, not expensive
diff --git a/docs/user/ppl/interfaces/endpoint.md b/docs/user/ppl/interfaces/endpoint.md
index fd61c212a97..300f99331d0 100644
--- a/docs/user/ppl/interfaces/endpoint.md
+++ b/docs/user/ppl/interfaces/endpoint.md
@@ -247,11 +247,22 @@ Expected output (trimmed):
 | `size` | Integer | Number of result rows returned. |
 
 
+### Rules For Recommendations
+| Rule | Severity | Trigger | Default threshold | Suggestion |
+|------|----------|---------------|-------------------|------------|
+| Ineffective Filter | WARNING | An in-memory filter operator passes through nearly all of its input rows (`rows_out / rows_in`). | ratio > 0.95 | Consider removing the filter or making it more selective. |
+| Join Row Explosion | WARNING / CRITICAL | A join produces far more rows than its combined inputs (`rows_out / rows_in`). | ratio > 5 (WARNING); ≥ 20 (CRITICAL) | Add filters to the subqueries before the join to reduce rows. |
+| Expensive Sort | WARNING | A sort (including a limit-fused top-N) runs in-memory over a large input and consumes a large share of execution time. | self-time > 20% of execute and input rows > 50,000 | Filter or limit rows before sorting (e.g. add `head` or a `where`). |
+| Bottleneck Stage | INFO | A single operator's self-time dominates total execution time. | self-time > 75% of execute | — |
+| Optimize Phase Dominates | INFO | Query planning takes longer than execution. | optimize > execute **and** optimize > 75 ms | — |
+
+
 ### Notes
 - Analyze output is only returned when the query finishes successfully.
 - Analyze requires the Calcite engine to be enabled (`plugins.calcite.enabled=true`).
 - The `profile` section uses the same format as the `profile` endpoint.
-
+- In rule calculation: `rows_in` is the sum of an operator's child row counts; self-time is an operator's own duration (`time_ms − max(child time_ms)`), since profile `time_ms` is cumulative
+  wall-time.
 
 
 ## Profile (Experimental) (Deprecated)