fix(jina): use dataclass replace to avoid modifying input documents (#2765)

This PR fixes the Jina document embedder and ranker to avoid mutating input
Documents when setting embeddings or scores.

Instead of mutating the original documents:
  doc.embedding = emb
  doc.score = relevance_score

We now create new document instances using dataclass replace:
  replace(doc, embedding=emb)
  replace(doc, score=relevance_score)

This follows the established pattern from haystack-ai/haystack#9693 and
aligns with other integrations (FastEmbed, Optimum, Nvidia, Bedrock, Cohere,
Google GenAI).

Related to: #2174

Co-authored-by: David S. Batista <dsbatista@gmail.com>

Author: GunaPalanivel

integrations/jina/src/haystack_integrations/components/embedders/jina/document_embedder.py

@@ -1,6 +1,7 @@
 # SPDX-FileCopyrightText: 2023-present deepset GmbH <info@deepset.ai>
 #
 # SPDX-License-Identifier: Apache-2.0
+from dataclasses import replace
 from typing import Any
 
 import requests
@@ -219,7 +220,8 @@ def run(self, documents: list[Document]) -> dict[str, Any]:
             texts_to_embed=texts_to_embed, batch_size=self.batch_size, parameters=parameters
         )
 
+        new_documents: list[Document] = []
         for doc, emb in zip(documents, embeddings, strict=True):
-            doc.embedding = emb
+            new_documents.append(replace(doc, embedding=emb))
 
-        return {"documents": documents, "meta": metadata}
+        return {"documents": new_documents, "meta": metadata}

integrations/jina/src/haystack_integrations/components/rankers/jina/ranker.py

@@ -1,6 +1,7 @@
 # SPDX-FileCopyrightText: 2023-present deepset GmbH <info@deepset.ai>
 #
 # SPDX-License-Identifier: Apache-2.0
+from dataclasses import replace
 from typing import Any
 
 import requests
@@ -161,12 +162,12 @@ def run(
             relevance_score = result["relevance_score"]
             doc = documents[index]
             if top_k is None or len(ranked_docs) < top_k:
-                doc.score = relevance_score
+                scored_doc = replace(doc, score=relevance_score)
                 if score_threshold is not None:
                     if relevance_score >= score_threshold:
-                        ranked_docs.append(doc)
+                        ranked_docs.append(scored_doc)
                 else:
-                    ranked_docs.append(doc)
+                    ranked_docs.append(scored_doc)
             else:
                 break
 

integrations/jina/tests/test_document_embedder.py

@@ -204,6 +204,29 @@ def test_run(self):
             assert all(isinstance(x, float) for x in doc.embedding)
         assert metadata == {"model": model, "usage": {"prompt_tokens": 4, "total_tokens": 4}}
 
+    def test_run_does_not_modify_original_documents(self):
+        docs = [
+            Document(content="I love cheese", meta={"topic": "Cuisine"}),
+            Document(content="A transformer is a deep learning architecture", meta={"topic": "ML"}),
+        ]
+
+        model = "jina-embeddings-v2-base-en"
+        with patch("requests.sessions.Session.post", side_effect=mock_session_post_response):
+            embedder = JinaDocumentEmbedder(
+                api_key=Secret.from_token("fake-api-key"),
+                model=model,
+            )
+
+            result = embedder.run(documents=docs)
+
+        # originals remain unchanged
+        for doc in docs:
+            assert doc.embedding is None
+
+        # returned docs carry embeddings
+        for doc_with_embedding in result["documents"]:
+            assert doc_with_embedding.embedding is not None
+
     def test_run_custom_batch_size(self):
         docs = [
             Document(content="I love cheese", meta={"topic": "Cuisine"}),

integrations/jina/tests/test_ranker.py

@@ -105,6 +105,32 @@ def test_run(self):
             assert doc.score == len(ranked_documents) - i
         assert metadata == {"model": model, "usage": {"prompt_tokens": 4, "total_tokens": 4}}
 
+    def test_run_does_not_modify_original_documents(self):
+        docs = [
+            Document(content="I love cheese"),
+            Document(content="A transformer is a deep learning architecture"),
+            Document(content="A transformer is something"),
+            Document(content="A transformer is not good"),
+        ]
+        query = "What is a transformer?"
+
+        model = "jina-ranker"
+        with patch("requests.sessions.Session.post", side_effect=mock_session_post_response):
+            ranker = JinaRanker(
+                api_key=Secret.from_token("fake-api-key"),
+                model=model,
+            )
+
+            result = ranker.run(query=query, documents=docs)
+
+        # originals remain unchanged
+        for doc in docs:
+            assert doc.score is None
+
+        # returned docs carry scores
+        for doc in result["documents"]:
+            assert doc.score is not None
+
     def test_run_wrong_input_format(self):
         ranker = JinaRanker(api_key=Secret.from_token("fake-api-key"))