Dataset Parsing Mode (BuiltIn vs Pipeline)#

Overview#

Every RAGFlow dataset has a parse_type that selects one of two mutually exclusive document-ingestion paths:

The enum is defined in web/src/constants/knowledge.ts.

How parse_type Controls the UI#

The dataset settings page (web/src/pages/dataset/dataset-setting/index.tsx) 1 watches parse_type via useWatch and conditionally renders two entirely different configuration panels 2:

• ParseType.BuiltIn: Shows the ChunkMethodItem dropdown (to pick a chunk method) and the ChunkMethodForm (the per-method configuration panel). A "Learn More" sidebar also appears.
• ParseType.Pipeline: Hides chunk-method fields entirely and shows DataFlowSelect — a searchable dropdown that lets the user attach any existing DataFlow Canvas pipeline via pipeline_id.

Switching modes triggers a side-effect 3:

• Switching to BuiltIn clears pipeline_id.
• Switching to Pipeline resets chunk_method back to DocumentParserType.Naive (default).

The toggle itself is rendered by ParseTypeItem — a radio group with two options labeled "Built-In" and "Manual Setup" (i18n keys builtIn / manualSetup).

BuiltIn Mode: Chunk Methods#

When parse_type === BuiltIn, the active chunk_method is read from DocumentParserType, which maps to 14 parser slugs:

naive, qa, resume, manual, table, paper, book, laws, presentation, picture, one, audio, email, tag, knowledge_graph

ChunkMethodForm dispatches to a per-method configuration component via ConfigurationComponentMap (in chunk-method-form.tsx). See the Dataset Configuration UI article for the full dispatch table and per-method field breakdown.

The default is DocumentParserType.Naive 4.

Pipeline Mode: DataFlow Canvas#

When parse_type === Pipeline, the form field pipeline_id stores the ID of a DataFlow Canvas agent. DataFlowSelect (in web/src/components/data-pipeline-select/) fetches available pipelines via the /agents endpoint filtered by canvas_category: AgentCategory.DataflowCanvas.

A pipeline executes a fixed linear chain: File → Parser → Chunker → Indexer. See the Pipeline Canvas Architecture article for runtime details.

Backend Behavior#

parse_type is a request-only field — it is used at API validation time but is not persisted to the database. The service layer explicitly deletes it before saving (in api/apps/dataset_api_service.py). What gets stored instead are:

• parser_id — the chosen chunk method slug (e.g. "naive")
• pipeline_id — the DataFlow Canvas ID (nullable)

The API validation model (CreateDatasetReq) enforces mutual exclusivity: if chunk_method (i.e. parser_id) is provided, pipeline_id must be absent, and vice versa. If neither is present, the system defaults to chunk_method="naive".

The Knowledgebase and Document DB models both carry parser_id and pipeline_id fields, allowing per-document overrides at the document level as well.

Zod Form Validation#

The frontend enforces the same mutual-exclusivity constraint in formSchema (Zod): a superRefine validator requires a non-empty pipeline_id when parse_type === Pipeline, and requires a valid chunk_method otherwise.

Key Source Files#

Related articles:

• Dataset Configuration UI — chunk method dispatch and per-method form components
• Pipeline Canvas Architecture — DataFlow Pipeline runtime, components, and execution model