Knowledge Graph#

RAGFlow's Knowledge Graph (KG) feature lets you enrich a knowledge base with structured entity/relation/community data extracted from documents, then inject that graph context into retrieval. Queries routed through the KG path receive a synthesized result block (labeled "Related content in Knowledge Graph" with similarity: 1.0) prepended to the normal chunk results 1.

Status as of v0.27.0: GraphRAG is superseded by the Knowledge Compilation Engine, but the existing GraphRAG code remains in the repository and is still used by datasets created before the migration. 2

Architecture: Document-Store–Backed Graph#

Rather than a dedicated graph database, all graph data is stored as special chunks in the same document store (Elasticsearch / Infinity) used for ordinary chunks. Each chunk carries a knowledge_graph_kwd field that identifies its type:

The KGSearch class queries all three types at retrieval time, combining entity rows, relation rows, and community reports into a single returned block. Community reports are ordered by weight_flt and their entities are used as the community-matching key 3.

Graph indexing constants (batch sizes, timeouts, retry logic) live at the top of rag/graphrag/general/index.py:

• Default batch chunk token size: 4 096 tokens
• Entity resolution timeout: 1 800 s
• Community detection timeout: 1 800 s

Three built-in extraction methods are registered in the same file 4:

Four-Step User Workflow#

The "Knowledge Graph" chunk method was removed from the dataset-creation dialog in September 2025 5 to fix race conditions, inefficient reparsing, and premature community detection 6. KG generation is now a separate post-parsing task. The required order is strict — doing steps out of order means the KG won't be used during retrieval 7:

1. Configure — Set entity types in the dataset's Configuration page before uploading files. Settings are locked at import time.
2. Parse — Upload and parse documents until visible chunks appear.
3. Generate — Click Generate → Knowledge graph on the Files page.
4. Enable — Toggle "Use Knowledge Graph in search" when creating or editing a retrieval query / chat assistant.

KG task progress is tracked via the graphrag_task_id and graphrag_task_finish_at fields on the Knowledgebase model 8.

The global kg_retriever singleton (a KGSearch instance backed by the shared docStoreConn) is initialized in common/settings.py init_settings() 9.

Knowledge Compilation Engine (v0.27.0)#

As of v0.27.0, the GraphRAG system is replaced by the Knowledge Compilation Engine, introduced in PR #16515 10. This new engine lives in rag/advanced_rag/knowlege_compile/ and provides a template-driven, concurrent compilation pipeline that covers multiple artifact types — not just knowledge graphs:

Key modules:

• structure.py — Compiles document structure (list/set/hypergraph chain kinds) using LLMCallPool for concurrent LLM calls.
• wiki.py — KB-wide MAP → REDUCE → PLAN → REFINE pipeline for artifact/wiki compilation.
• runner.py — Handler-free core for document-scoped compilation.

Task orchestration runs through rag/svr/task_executor_refactor/ — the dataset_wiki_generator.py handles KB-wide artifact tasks (MAP batch size: 64 chunks) and the chunk_post_processor.py handles structure compilation 11 12.

Compilation templates are YAML files stored under api/db/init_data/compilation_templates/ and managed via REST endpoints at api/apps/restful_apis/compilation_template_api.py. The Knowledgebase model tracks compilation progress via artifact_task_id / artifact_task_finish_at 13.

Key Source References#