diff --git a/src/promptflow-devkit/promptflow/_sdk/operations/_local_storage_operations.py b/src/promptflow-devkit/promptflow/_sdk/operations/_local_storage_operations.py
index a3437de5348..4a7b31dc678 100644
--- a/src/promptflow-devkit/promptflow/_sdk/operations/_local_storage_operations.py
+++ b/src/promptflow-devkit/promptflow/_sdk/operations/_local_storage_operations.py
@@ -515,6 +515,14 @@ def load_inputs_and_outputs(self) -> Tuple["DataFrame", "DataFrame"]:
                 outputs = self._outputs_padding(outputs, inputs[LINE_NUMBER].tolist())
                 outputs.fillna(value="(Failed)", inplace=True)  # replace nan with explicit prompt
                 outputs = outputs.set_index(LINE_NUMBER)
+        # Sort inputs by line_number so their row order matches the sorted
+        # outputs index.  Without this, get_details() merges inputs and
+        # outputs positionally and rows become misaligned whenever the
+        # async executor wrote results in a different order than the input
+        # data was originally listed.
+        # See https://github.com/microsoft/promptflow/issues/2646
+        if LINE_NUMBER in inputs.columns:
+            inputs = inputs.sort_values(LINE_NUMBER, ascending=True).reset_index(drop=True)
         return inputs, outputs
 
     def _collect_io_from_debug_info(self) -> Tuple["DataFrame", "DataFrame"]:
diff --git a/src/promptflow-devkit/tests/sdk_cli_test/unittests/test_local_storage_operations.py b/src/promptflow-devkit/tests/sdk_cli_test/unittests/test_local_storage_operations.py
index 78e07b4516e..bb06bec8bba 100644
--- a/src/promptflow-devkit/tests/sdk_cli_test/unittests/test_local_storage_operations.py
+++ b/src/promptflow-devkit/tests/sdk_cli_test/unittests/test_local_storage_operations.py
@@ -34,3 +34,62 @@ def test_outputs_padding(self) -> None:
         assert df_with_padding.iloc[0].to_dict() == {LINE_NUMBER: 1, "col": "a"}
         assert df_with_padding.iloc[1].to_dict() == {LINE_NUMBER: 2, "col": "b"}
         assert df_with_padding.iloc[2].to_dict() == {LINE_NUMBER: 4, "col": ""}
+
+    def test_load_inputs_and_outputs_sorts_inputs_by_line_number(
+        self, tmp_path
+    ) -> None:
+        """Inputs returned by load_inputs_and_outputs must be sorted by line_number.
+
+        Regression test for https://github.com/microsoft/promptflow/issues/2646
+
+        When the async executor writes results in a different order than the
+        input dataset, the raw inputs JSON file can have rows in a non-sequential
+        order (e.g. [2, 0, 1]).  _outputs_padding() already sorts outputs by
+        line_number, but inputs were left unsorted.  get_details() merges the
+        two DataFrames positionally, so mismatched orderings caused input rows
+        to be paired with the wrong output rows.
+        """
+        import json
+
+        # Write inputs file with rows in out-of-order line_number sequence
+        inputs_data = [
+            {LINE_NUMBER: 2, "query": "third"},
+            {LINE_NUMBER: 0, "query": "first"},
+            {LINE_NUMBER: 1, "query": "second"},
+        ]
+        inputs_path = tmp_path / "inputs.jsonl"
+        with open(inputs_path, "w") as f:
+            for row in inputs_data:
+                f.write(json.dumps(row) + "\n")
+
+        # Write outputs file with rows in sorted line_number sequence
+        outputs_data = [
+            {LINE_NUMBER: 0, "answer": "ans0"},
+            {LINE_NUMBER: 1, "answer": "ans1"},
+            {LINE_NUMBER: 2, "answer": "ans2"},
+        ]
+        outputs_path = tmp_path / "outputs.jsonl"
+        with open(outputs_path, "w") as f:
+            for row in outputs_data:
+                f.write(json.dumps(row) + "\n")
+
+        # Patch the LocalStorageOperations instance to point at our temp files
+        # without needing a full Run / directory tree.
+        ops = object.__new__(LocalStorageOperations)
+        ops._sdk_inputs_path = inputs_path
+        ops._sdk_output_path = outputs_path
+
+        inputs, outputs = ops.load_inputs_and_outputs()
+
+        # inputs must be in ascending line_number order after the fix
+        assert list(inputs[LINE_NUMBER]) == [0, 1, 2], (
+            "load_inputs_and_outputs() returned inputs in wrong order; "
+            f"got line_numbers={list(inputs[LINE_NUMBER])}"
+        )
+        assert list(inputs["query"]) == ["first", "second", "third"], (
+            "Input rows are not aligned with their line_numbers after sort"
+        )
+
+        # outputs must also be in ascending line_number order (already guaranteed
+        # by _outputs_padding / set_index, just verify the contract is intact)
+        assert list(outputs.index) == [0, 1, 2]