feat: Weights and score normalization for JoinDocuments node with reciprocal rank fusion

haystack/nodes/other/join_docs.py

@@ -155,8 +155,16 @@ def _calculate_rrf(self, results):
         K = 61
 
         scores_map = defaultdict(int)
-        for result in results:
+        weights = self.weights if self.weights else [1 / len(results)] * len(results)
+
+        # Calculate weighted reciprocal rank fusion score
+        for result, weight in zip(results, weights):
             for rank, doc in enumerate(result):
-                scores_map[doc.id] += 1 / (K + rank)
+                scores_map[doc.id] += (weight * len(results)) / (K + rank)
+
+        # Normalize scores. Note: len(results) / K is the maximum possible score,
+        # achieved by being ranked first in all results with non-zero weight.
+        for id in scores_map:
+            scores_map[id] = scores_map[id] / (len(results) / K)
 
         return scores_map

releasenotes/notes/join-docs-weighting-rrf-c52ba00a25004fd4.yaml

@@ -0,0 +1,6 @@
+---
+enhancements:
+  - |
+    Make `JoinDocuments` sensitive to `weights` parameter when
+    `join_mode` is reciprocal rank fusion. Add score normalization
+    for `JoinDocuments` when `join_mode` is reciprocal rank fusion.

test/nodes/test_join_documents.py

@@ -3,6 +3,7 @@
 
 from haystack import Document
 from haystack.nodes.other.join_docs import JoinDocuments
+from copy import deepcopy
 
 
 @pytest.mark.unit
@@ -113,3 +114,38 @@ def test_joindocuments_concatenate_duplicate_docs_null_score():
     result, _ = join_docs.run(inputs)
     assert len(result["documents"]) == 3
     assert result["documents"] == expected_outputs["documents"]
+
+
+@pytest.mark.unit
+def test_joindocuments_rrf_weights():
+    """
+    Test that the reciprocal rank fusion method correctly handles weights.
+    """
+    inputs_none = [
+        {
+            "documents": [
+                Document(content="text document 1", content_type="text", score=0.2),
+                Document(content="text document 2", content_type="text", score=0.3),
+            ]
+        },
+        {
+            "documents": [
+                Document(content="text document 3", content_type="text", score=0.7),
+                Document(content="text document 4", content_type="text", score=None),
+            ]
+        },
+    ]
+
+    inputs_even = deepcopy(inputs_none)
+    inputs_uneven = deepcopy(inputs_none)
+
+    join_docs_none = JoinDocuments(join_mode="reciprocal_rank_fusion")
+    result_none, _ = join_docs_none.run(inputs_none)
+    join_docs_even = JoinDocuments(join_mode="reciprocal_rank_fusion", weights=[0.5, 0.5])
+    result_even, _ = join_docs_even.run(inputs_even)
+    join_docs_uneven = JoinDocuments(join_mode="reciprocal_rank_fusion", weights=[0.7, 0.3])
+    result_uneven, _ = join_docs_uneven.run(inputs_uneven)
+
+    assert result_none["documents"] == result_even["documents"]
+    assert result_uneven["documents"] != result_none["documents"]
+    assert result_uneven["documents"][0].score > result_none["documents"][0].score