fix: summary index bug (#31810)

Co-authored-by: autofix-ci[bot] <114827586+autofix-ci[bot]@users.noreply.github.com>
Co-authored-by: Jyong <76649700+JohnJyong@users.noreply.github.com>
Co-authored-by: zxhlyh <jasonapring2015@outlook.com>
Co-authored-by: Yansong Zhang <916125788@qq.com>
Co-authored-by: hj24 <mambahj24@gmail.com>
Co-authored-by: CodingOnStar <hanxujiang@dify.ai>
Co-authored-by: CodingOnStar <hanxujiang@dify.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

api/controllers/console/datasets/datasets_document.py

@@ -1339,6 +1339,18 @@ class DocumentGenerateSummaryApi(Resource):
             missing_ids = set(document_list) - found_ids
             raise NotFound(f"Some documents not found: {list(missing_ids)}")
 
+        # Update need_summary to True for documents that don't have it set
+        # This handles the case where documents were created when summary_index_setting was disabled
+        documents_to_update = [doc for doc in documents if not doc.need_summary and doc.doc_form != "qa_model"]
+
+        if documents_to_update:
+            document_ids_to_update = [str(doc.id) for doc in documents_to_update]
+            DocumentService.update_documents_need_summary(
+                dataset_id=dataset_id,
+                document_ids=document_ids_to_update,
+                need_summary=True,
+            )
+
         # Dispatch async tasks for each document
         for document in documents:
             # Skip qa_model documents as they don't generate summaries

api/core/indexing_runner.py

@@ -369,7 +369,9 @@ class IndexingRunner:
         # Generate summary preview
         summary_index_setting = tmp_processing_rule.get("summary_index_setting")
         if summary_index_setting and summary_index_setting.get("enable") and preview_texts:
-            preview_texts = index_processor.generate_summary_preview(tenant_id, preview_texts, summary_index_setting)
+            preview_texts = index_processor.generate_summary_preview(
+                tenant_id, preview_texts, summary_index_setting, doc_language
+            )
 
         return IndexingEstimate(total_segments=total_segments, preview=preview_texts)
 

api/core/llm_generator/prompts.py

@@ -441,11 +441,13 @@ DEFAULT_GENERATOR_SUMMARY_PROMPT = (
 
 Requirements:
 1. Write a concise summary in plain text
-2. Use the same language as the input content
+2. You must write in {language}. No language other than {language} should be used.
 3. Focus on important facts, concepts, and details
 4. If images are included, describe their key information
 5. Do not use words like "好的", "ok", "I understand", "This text discusses", "The content mentions"
 6. Write directly without extra words
+7. If there is not enough content to generate a meaningful summary, 
+   return an empty string without any explanation or prompt
 
 Output only the summary text. Start summarizing now:
 

api/core/rag/index_processor/index_processor_base.py

@@ -48,12 +48,22 @@ class BaseIndexProcessor(ABC):
 
     @abstractmethod
     def generate_summary_preview(
-        self, tenant_id: str, preview_texts: list[PreviewDetail], summary_index_setting: dict
+        self,
+        tenant_id: str,
+        preview_texts: list[PreviewDetail],
+        summary_index_setting: dict,
+        doc_language: str | None = None,
     ) -> list[PreviewDetail]:
         """
         For each segment in preview_texts, generate a summary using LLM and attach it to the segment.
         The summary can be stored in a new attribute, e.g., summary.
         This method should be implemented by subclasses.
+
+        Args:
+            tenant_id: Tenant ID
+            preview_texts: List of preview details to generate summaries for
+            summary_index_setting: Summary index configuration
+            doc_language: Optional document language to ensure summary is generated in the correct language
         """
         raise NotImplementedError
 

api/core/rag/index_processor/processor/paragraph_index_processor.py

@@ -275,7 +275,11 @@ class ParagraphIndexProcessor(BaseIndexProcessor):
             raise ValueError("Chunks is not a list")
 
     def generate_summary_preview(
-        self, tenant_id: str, preview_texts: list[PreviewDetail], summary_index_setting: dict
+        self,
+        tenant_id: str,
+        preview_texts: list[PreviewDetail],
+        summary_index_setting: dict,
+        doc_language: str | None = None,
     ) -> list[PreviewDetail]:
         """
         For each segment, concurrently call generate_summary to generate a summary
@@ -298,11 +302,15 @@ class ParagraphIndexProcessor(BaseIndexProcessor):
             if flask_app:
                 # Ensure Flask app context in worker thread
                 with flask_app.app_context():
-                    summary, _ = self.generate_summary(tenant_id, preview.content, summary_index_setting)
+                    summary, _ = self.generate_summary(
+                        tenant_id, preview.content, summary_index_setting, document_language=doc_language
+                    )
                     preview.summary = summary
             else:
                 # Fallback: try without app context (may fail)
-                summary, _ = self.generate_summary(tenant_id, preview.content, summary_index_setting)
+                summary, _ = self.generate_summary(
+                    tenant_id, preview.content, summary_index_setting, document_language=doc_language
+                )
                 preview.summary = summary
 
         # Generate summaries concurrently using ThreadPoolExecutor
@@ -356,6 +364,7 @@ class ParagraphIndexProcessor(BaseIndexProcessor):
         text: str,
         summary_index_setting: dict | None = None,
         segment_id: str | None = None,
+        document_language: str | None = None,
     ) -> tuple[str, LLMUsage]:
         """
         Generate summary for the given text using ModelInstance.invoke_llm and the default or custom summary prompt,
@@ -366,6 +375,8 @@ class ParagraphIndexProcessor(BaseIndexProcessor):
             text: Text content to summarize
             summary_index_setting: Summary index configuration
             segment_id: Optional segment ID to fetch attachments from SegmentAttachmentBinding table
+            document_language: Optional document language (e.g., "Chinese", "English")
+                to ensure summary is generated in the correct language
 
         Returns:
             Tuple of (summary_content, llm_usage) where llm_usage is LLMUsage object
@@ -381,8 +392,22 @@ class ParagraphIndexProcessor(BaseIndexProcessor):
             raise ValueError("model_name and model_provider_name are required in summary_index_setting")
 
         # Import default summary prompt
+        is_default_prompt = False
         if not summary_prompt:
             summary_prompt = DEFAULT_GENERATOR_SUMMARY_PROMPT
+            is_default_prompt = True
+
+        # Format prompt with document language only for default prompt
+        # Custom prompts are used as-is to avoid interfering with user-defined templates
+        # If document_language is provided, use it; otherwise, use "the same language as the input content"
+        # This is especially important for image-only chunks where text is empty or minimal
+        if is_default_prompt:
+            language_for_prompt = document_language or "the same language as the input content"
+            try:
+                summary_prompt = summary_prompt.format(language=language_for_prompt)
+            except KeyError:
+                # If default prompt doesn't have {language} placeholder, use it as-is
+                pass
 
         provider_manager = ProviderManager()
         provider_model_bundle = provider_manager.get_provider_model_bundle(

api/core/rag/index_processor/processor/parent_child_index_processor.py

@@ -358,7 +358,11 @@ class ParentChildIndexProcessor(BaseIndexProcessor):
         }
 
     def generate_summary_preview(
-        self, tenant_id: str, preview_texts: list[PreviewDetail], summary_index_setting: dict
+        self,
+        tenant_id: str,
+        preview_texts: list[PreviewDetail],
+        summary_index_setting: dict,
+        doc_language: str | None = None,
     ) -> list[PreviewDetail]:
         """
         For each parent chunk in preview_texts, concurrently call generate_summary to generate a summary
@@ -389,6 +393,7 @@ class ParentChildIndexProcessor(BaseIndexProcessor):
                         tenant_id=tenant_id,
                         text=preview.content,
                         summary_index_setting=summary_index_setting,
+                        document_language=doc_language,
                     )
                     preview.summary = summary
             else:
@@ -397,6 +402,7 @@ class ParentChildIndexProcessor(BaseIndexProcessor):
                     tenant_id=tenant_id,
                     text=preview.content,
                     summary_index_setting=summary_index_setting,
+                    document_language=doc_language,
                 )
                 preview.summary = summary
 

api/core/rag/index_processor/processor/qa_index_processor.py

@@ -241,7 +241,11 @@ class QAIndexProcessor(BaseIndexProcessor):
         }
 
     def generate_summary_preview(
-        self, tenant_id: str, preview_texts: list[PreviewDetail], summary_index_setting: dict
+        self,
+        tenant_id: str,
+        preview_texts: list[PreviewDetail],
+        summary_index_setting: dict,
+        doc_language: str | None = None,
     ) -> list[PreviewDetail]:
         """
         QA model doesn't generate summaries, so this method returns preview_texts unchanged.

api/core/workflow/nodes/knowledge_index/knowledge_index_node.py

@@ -78,12 +78,21 @@ class KnowledgeIndexNode(Node[KnowledgeIndexNodeData]):
                 indexing_technique = node_data.indexing_technique or dataset.indexing_technique
                 summary_index_setting = node_data.summary_index_setting or dataset.summary_index_setting
 
+                # Try to get document language if document_id is available
+                doc_language = None
+                document_id = variable_pool.get(["sys", SystemVariableKey.DOCUMENT_ID])
+                if document_id:
+                    document = db.session.query(Document).filter_by(id=document_id.value).first()
+                    if document and document.doc_language:
+                        doc_language = document.doc_language
+
                 outputs = self._get_preview_output_with_summaries(
                     node_data.chunk_structure,
                     chunks,
                     dataset=dataset,
                     indexing_technique=indexing_technique,
                     summary_index_setting=summary_index_setting,
+                    doc_language=doc_language,
                 )
                 return NodeRunResult(
                     status=WorkflowNodeExecutionStatus.SUCCEEDED,
@@ -315,6 +324,7 @@ class KnowledgeIndexNode(Node[KnowledgeIndexNodeData]):
         dataset: Dataset,
         indexing_technique: str | None = None,
         summary_index_setting: dict | None = None,
+        doc_language: str | None = None,
     ) -> Mapping[str, Any]:
         """
         Generate preview output with summaries for chunks in preview mode.
@@ -326,6 +336,7 @@ class KnowledgeIndexNode(Node[KnowledgeIndexNodeData]):
             dataset: Dataset object (for tenant_id)
             indexing_technique: Indexing technique from node config or dataset
             summary_index_setting: Summary index setting from node config or dataset
+            doc_language: Optional document language to ensure summary is generated in the correct language
         """
         index_processor = IndexProcessorFactory(chunk_structure).init_index_processor()
         preview_output = index_processor.format_preview(chunks)
@@ -365,6 +376,7 @@ class KnowledgeIndexNode(Node[KnowledgeIndexNodeData]):
                                 tenant_id=dataset.tenant_id,
                                 text=preview_item["content"],
                                 summary_index_setting=summary_index_setting,
+                                document_language=doc_language,
                             )
                             if summary:
                                 preview_item["summary"] = summary
@@ -374,6 +386,7 @@ class KnowledgeIndexNode(Node[KnowledgeIndexNodeData]):
                             tenant_id=dataset.tenant_id,
                             text=preview_item["content"],
                             summary_index_setting=summary_index_setting,
+                            document_language=doc_language,
                         )
                         if summary:
                             preview_item["summary"] = summary

api/services/dataset_service.py

@@ -16,6 +16,7 @@ from sqlalchemy.orm import Session
 from werkzeug.exceptions import Forbidden, NotFound
 
 from configs import dify_config
+from core.db.session_factory import session_factory
 from core.errors.error import LLMBadRequestError, ProviderTokenNotInitError
 from core.file import helpers as file_helpers
 from core.helper.name_generator import generate_incremental_name
@@ -1388,6 +1389,46 @@ class DocumentService:
         ).all()
         return documents
 
+    @staticmethod
+    def update_documents_need_summary(dataset_id: str, document_ids: Sequence[str], need_summary: bool = True) -> int:
+        """
+        Update need_summary field for multiple documents.
+
+        This method handles the case where documents were created when summary_index_setting was disabled,
+        and need to be updated when summary_index_setting is later enabled.
+
+        Args:
+            dataset_id: Dataset ID
+            document_ids: List of document IDs to update
+            need_summary: Value to set for need_summary field (default: True)
+
+        Returns:
+            Number of documents updated
+        """
+        if not document_ids:
+            return 0
+
+        document_id_list: list[str] = [str(document_id) for document_id in document_ids]
+
+        with session_factory.create_session() as session:
+            updated_count = (
+                session.query(Document)
+                .filter(
+                    Document.id.in_(document_id_list),
+                    Document.dataset_id == dataset_id,
+                    Document.doc_form != "qa_model",  # Skip qa_model documents
+                )
+                .update({Document.need_summary: need_summary}, synchronize_session=False)
+            )
+            session.commit()
+            logger.info(
+                "Updated need_summary to %s for %d documents in dataset %s",
+                need_summary,
+                updated_count,
+                dataset_id,
+            )
+            return updated_count
+
     @staticmethod
     def get_document_download_url(document: Document) -> str:
         """

api/services/rag_pipeline/rag_pipeline_transform_service.py

@@ -174,6 +174,10 @@ class RagPipelineTransformService:
         else:
             dataset.retrieval_model = knowledge_configuration.retrieval_model.model_dump()
 
+        # Copy summary_index_setting from dataset to knowledge_index node configuration
+        if dataset.summary_index_setting:
+            knowledge_configuration.summary_index_setting = dataset.summary_index_setting
+
         knowledge_configuration_dict.update(knowledge_configuration.model_dump())
         node["data"] = knowledge_configuration_dict
         return node

api/services/summary_index_service.py

@@ -49,11 +49,18 @@ class SummaryIndexService:
         # Use lazy import to avoid circular import
         from core.rag.index_processor.processor.paragraph_index_processor import ParagraphIndexProcessor
 
+        # Get document language to ensure summary is generated in the correct language
+        # This is especially important for image-only chunks where text is empty or minimal
+        document_language = None
+        if segment.document and segment.document.doc_language:
+            document_language = segment.document.doc_language
+
         summary_content, usage = ParagraphIndexProcessor.generate_summary(
             tenant_id=dataset.tenant_id,
             text=segment.content,
             summary_index_setting=summary_index_setting,
             segment_id=segment.id,
+            document_language=document_language,
         )
 
         if not summary_content:
@@ -558,6 +565,9 @@ class SummaryIndexService:
                     )
                     session.add(summary_record)
 
+            # Commit the batch created records
+            session.commit()
+
     @staticmethod
     def update_summary_record_error(
         segment: DocumentSegment,
@@ -762,7 +772,6 @@ class SummaryIndexService:
                 dataset=dataset,
                 status="not_started",
             )
-            session.commit()  # Commit initial records
 
             summary_records = []