Change document question answering pipeline to always return an array (#19071)

Co-authored-by: Ankur Goyal <ankur@impira.com>

src/transformers/pipelines/document_question_answering.py

@@ -383,8 +383,6 @@ class DocumentQuestionAnsweringPipeline(Pipeline):
             answers = self.postprocess_extractive_qa(model_outputs, top_k=top_k, **kwargs)
 
         answers = sorted(answers, key=lambda x: x.get("score", 0), reverse=True)[:top_k]
-        if len(answers) == 1:
-            return answers[0]
         return answers
 
     def postprocess_donut(self, model_outputs, **kwargs):

tests/pipelines/test_pipelines_document_question_answering.py

@@ -267,7 +267,7 @@ class DocumentQuestionAnsweringPipelineTests(unittest.TestCase, metaclass=Pipeli
         image = INVOICE_URL
         question = "What is the invoice number?"
         outputs = dqa_pipeline(image=image, question=question, top_k=2)
-        self.assertEqual(nested_simplify(outputs, decimals=4), {"answer": "us-001"})
+        self.assertEqual(nested_simplify(outputs, decimals=4), [{"answer": "us-001"}])
 
     @require_tf
     @unittest.skip("Document question answering not implemented in TF")