Knowledge Base Summarization Pipeline#

Dify's knowledge base summarization pipeline automatically generates LLM-powered summaries for each DocumentSegment when a document is uploaded and indexed. Each summary is independently embedded and stored in the vector database, creating a secondary retrieval layer alongside the original chunk vectors.

Prerequisites#

Summarization only runs when all three conditions hold 1:

1. document.need_summary is True
2. dataset.indexing_technique == "high_quality" (economy mode is excluded)
3. dataset.summary_index_setting.enable == True

The feature is also skipped for qa_model documents 2.

Configuration — SummaryIndexSettingDict#

Configuration is stored as a JSON column on the Dataset model 3 and typed as SummaryIndexSettingDict:

If summary_prompt is omitted, the default prompt is used with a {language} placeholder filled from the document's detected language 4.

Pipeline Orchestration#

Trigger: Post-Indexing#

After document indexing completes in document_indexing_task.py, the indexing task checks each completed document and enqueues a separate Celery task 5:

generate_summary_index_task.delay(dataset.id, document.id, None)

Celery Task: generate_summary_index_task#

generate_summary_index_task runs on the dataset_summary Celery queue (separate from the main dataset queue). It:

1. Validates dataset, document, and configuration preconditions 6
2. Determines whether to process only parent chunks (for parent_child_index datasets) 7
3. Delegates to SummaryIndexService.generate_summaries_for_document() 8

SummaryIndexService.generate_summaries_for_document()#

This method iterates over all enabled, completed segments for the document:

1. Batch-creates stub DocumentSegmentSummary records with NOT_STARTED status upfront 9 — this enables status tracking before generation completes.
2. Loops over segments, calling generate_and_vectorize_summary() for each. Errors on individual segments are caught and logged; processing continues for remaining segments 10.

Summary Generation — ParagraphIndexProcessor.generate_summary()#

The core LLM call in ParagraphIndexProcessor:

1. Resolves model/provider from summary_index_setting 11
2. Checks if the model supports vision; if so, extracts images from SegmentAttachmentBinding or from markdown ![]() links in the segment text 12
3. Invokes the LLM with stream=False and returns (summary_content, LLMUsage) 13
4. Deducts token quota via deduct_llm_quota 14

Vectorization — SummaryIndexService.vectorize_summary()#

After generation, the summary text is embedded and stored in the vector DB:

• Creates a Document with metadata is_summary=True and original_chunk_id linking it back to the source DocumentSegment 15
• Retries up to 3 times with exponential backoff for transient connection errors 16
• Updates DocumentSegmentSummary.status to COMPLETED (or ERROR) and saves the summary_index_node_id and token count 17

Summary Lifecycle#

DocumentSegmentSummary.enabled stays in sync with the parent DocumentSegment.enabled — disabling a segment removes its summary vector but preserves the record 18. Re-enabling a segment re-vectorizes the existing summary content 19.

Manual Triggers#

Beyond the automatic post-indexing flow, summaries can be triggered or updated via:

• REST API: POST /datasets/<dataset_id>/documents/generate-summary — queues generate_summary_index_task per document 20
• Segment update: When a user edits a segment with a manual summary, update_summary_for_segment() re-vectorizes immediately (synchronously) so it is searchable at once

Key Files#