Dataset Configuration UI#

The dataset settings page in RAGFlow uses a dynamic, chunk-method-driven form architecture: selecting a chunk method swaps in a method-specific configuration component, which in turn renders only the form fields relevant to that method.

Entry Point#

The top-level page is DatasetSettings (web/src/pages/dataset/dataset-setting/index.tsx). It owns the react-hook-form instance validated against formSchema (Zod), renders the general dataset fields, and conditionally mounts ChunkMethodForm when parse_type === ParseType.BuiltIn.

The form defaults chunk_method to DocumentParserType.Naive and initialises parser_config.compilation_template_group_id as an empty array 1. When the user switches to a Pipeline parse type, chunk_method is reset to Naive and the chunk-method section is hidden 2.

Dynamic Configuration Dispatch#

ChunkMethodForm (chunk-method-form.tsx) is the dispatch layer. It watches chunk_method via useWatch and looks up the active component from ConfigurationComponentMap 3:

If no method is selected, an EmptyComponent renders a blank <div> 4.

All configuration components live under web/src/pages/dataset/dataset-setting/configuration/.

Per-Method Configuration Components#

Each component uses ConfigurationFormContainer and MainContainer as layout wrappers, then composes reusable atomic form field components.

NaiveConfiguration 5 is the most complete example — it includes:

• CompilationTemplateFormField — compilation template group selector (see below)
• LayoutRecognizeFormField, MaxTokenNumberFormField, DelimiterFormField, ChildrenDelimiterForm, EnableTocToggle, ImageContextWindow, AutoMetadata, OverlappedPercent
• AutoKeywordsFormField, AutoQuestionsFormField, ExcelToHtmlFormField

Other methods are narrower: KnowledgeGraphConfiguration shows EntityTypesFormField + MaxTokenNumberFormField + DelimiterFormField; AudioConfiguration, EmailConfiguration, and PictureConfiguration show only the auto-tagging/metadata fields. QAConfiguration, ResumeConfiguration, and TagConfiguration are currently placeholder implementations returning empty fragments 6 7 8.

Shared per-section form items (ChunkMethodItem, ParseTypeItem, EnableTocToggle, ImageContextWindow, OverlappedPercent, AutoMetadata, LLMModelItem, etc.) are centralised in common-item.tsx.

CompilationTemplateFormField#

CompilationTemplateFormField (web/src/components/compilation-template-form-field.tsx) binds to parser_config.compilation_template_group_id by default. It wraps CompilationTemplateMultiSelect, which:

1. Fetches all user-created template groups via useFetchAllCompilationTemplateGroups.
2. Derives selectedScopes from the current value to enforce the at-most-one-group-per-scope rule 9: once a "file"- or "dataset"-scope group is selected, all other groups sharing that scope are disabled in the picker.
3. Appends the scope label ((file) / (dataset)) as a suffix on each option 10.
4. Normalises legacy single-string values into arrays during migration 11.

The field is used both in NaiveConfiguration (dataset-settings page) and directly in ChunkMethodDialog (the per-document chunk method dialog), where it appears when parseType === ParseType.BuiltIn.

For details on template group scopes, CRUD API, and the downstream compilation workflow, see the Compilation Template Management article.

Form Schema#

formSchema (Zod) covers the full settings payload including parser_config.compilation_template_group_id: z.array(z.string()).optional(). Custom superRefine validators enforce that a pipeline_id must be present when parse_type === Pipeline, that RAPTOR requires a prompt, and that GraphRAG requires at least one entity_type.

The same compilation_template_group_id array field is also declared in the ChunkMethodDialog's inline Zod schema 12.

Key Source Files#