Skip to content

RAGFlow is an agent-native database: a Retrieval-Augmented Generation (RAG) engine that grew into context infrastructure.

Fix: Promote the visual enhancement option to a global setting. (#20397)

balibabu9bb5ad99f25c
diff --git a/web/src/locales/en.ts b/web/src/locales/en.ts
index c82ab72d8c..4c8bfd6480 100644
--- a/web/src/locales/en.ts
+++ b/web/src/locales/en.ts
@@ -2764,9 +2764,9 @@ Best for: Documents with flowing, contextually connected content — such as boo
oneChunkTitle: 'Note',
oneChunkDescription:
'All parsed sections will be merged in order into a single chunk.',
- flattenMediaToText: 'Disable vision model',
- flattenMediaToTextTip:
- 'Treat image and table sections as plain text and skip vision enhancement.',
+ enableVisionEnhancement: 'Enable vision enhancement',
+ enableVisionEnhancementTip:
+ 'Use a vision model to parse image and table blocks; when off, they are treated as plain text.',
enableChildrenDelimiters: 'Child chunk are used for retrieval',
merge: 'Merge',
split: 'Split',
diff --git a/web/src/locales/zh.ts b/web/src/locales/zh.ts
index e893e24ed4..732f9708c3 100644
--- a/web/src/locales/zh.ts
+++ b/web/src/locales/zh.ts
@@ -2344,8 +2344,9 @@ NER:使用 spaCy NER 和基于规则的关键词提取来抽取 Entities 和 R
oneChunkTitle: 'Note',
oneChunkDescription:
'所有解析后的 sections 会按原始顺序合并为 1 个 chunk。',
- flattenMediaToText: '禁用视觉模型',
- flattenMediaToTextTip: '将图片和表格区块按普通文本处理,并跳过视觉增强。',
+ enableVisionEnhancement: '启用视觉增强',
+ enableVisionEnhancementTip:
+ '使用视觉模型解析图片和表格区块,关闭时按普通文本处理。',
enableChildrenDelimiters: '子块用于检索',
merge: '合并',
split: '拆分',
diff --git a/web/src/pages/agent/constant/pipeline.tsx b/web/src/pages/agent/constant/pipeline.tsx
index 2c3f7f4afa..4852cb1590 100644
--- a/web/src/pages/agent/constant/pipeline.tsx
+++ b/web/src/pages/agent/constant/pipeline.tsx
@@ -64,11 +64,10 @@ export const OutputFormatMap = {
[FileType.Audio]: AudioOutputFormat,
};
-// The video parser defaults to the tenant's VLM model and the audio parser to
-// the ASR model, keyed by the useFetchDefaultModelDictionary fields. A file
-// type without a configured tenant default keeps its empty model id.
+// The audio parser defaults to the tenant's ASR model, keyed by the
+// useFetchDefaultModelDictionary fields. (The vision model is a single global
+// parser option, not per file type, so video no longer appears here.)
export const FileTypeDefaultModelFieldMap: Partial<Record<FileType, string>> = {
- [FileType.Video]: ModelTypeToField.vision,
[FileType.Audio]: ModelTypeToField.asr,
};
@@ -168,13 +167,16 @@ export const initialParserValues = {
html: { type: 'string', value: '' },
json: { type: 'Array<object>', value: [] },
},
+ // Global vision enhancement, off by default; the model id is prefilled from
+ // the tenant's image2text default when a node is created.
+ vlm: { llm_id: '' },
+ enable_vision_enhancement: false,
setups: [
{
fileFormat: FileType.PDF,
output_format: PdfOutputFormat.Json,
parse_method: ParseDocumentType.DeepDOC,
preprocess: PreprocessValue.main_content,
- flatten_media_to_text: false,
remove_header_footer: false,
pages: [{ from: 1, to: 100000 }],
},
@@ -183,7 +185,6 @@ export const initialParserValues = {
output_format: SpreadsheetOutputFormat.Json,
parse_method: ParseDocumentType.DeepDOC,
preprocess: PreprocessValue.main_content,
- flatten_media_to_text: false,
},
{
fileFormat: FileType.Image,
@@ -202,7 +203,6 @@ export const initialParserValues = {
fileFormat: FileType.TextMarkdown,
output_format: TextMarkdownOutputFormat.Text,
preprocess: PreprocessValue.main_content,
- flatten_media_to_text: false,
},
{
fileFormat: FileType.Code,
@@ -219,14 +219,12 @@ export const initialParserValues = {
fileFormat: FileType.Doc,
output_format: DocxOutputFormat.Json,
preprocess: PreprocessValue.main_content,
- flatten_media_to_text: false,
remove_header_footer: false,
},
{
fileFormat: FileType.Docx,
output_format: DocxOutputFormat.Json,
preprocess: PreprocessValue.main_content,
- flatten_media_to_text: false,
remove_header_footer: false,
},
{
@@ -238,7 +236,6 @@ export const initialParserValues = {
{
fileFormat: FileType.Video,
output_format: VideoOutputFormat.Text,
- vlm: { llm_id: '' },
},
{
fileFormat: FileType.Audio,
diff --git a/web/src/pages/agent/form/parser-form/video-form-fields.tsx b/web/src/pages/agent/form/parser-form/audio-form-fields.tsx
similarity index 62%
rename from web/src/pages/agent/form/parser-form/video-form-fields.tsx
rename to web/src/pages/agent/form/parser-form/audio-form-fields.tsx
index ee536c45be..1da002522f 100644
--- a/web/src/pages/agent/form/parser-form/video-form-fields.tsx
+++ b/web/src/pages/agent/form/parser-form/audio-form-fields.tsx
@@ -24,21 +24,3 @@ export function AudioFormFields({ prefix }: OutputFormatFormFieldProps) {
</>
);
}
-
-export function VideoFormFields({ prefix }: OutputFormatFormFieldProps) {
- const { t } = useTranslation();
- const ownerTenantId = useOwnerTenantId();
-
- return (
- <>
- {/* Multimodal Model */}
- <ModelTreeSelectFormField
- name={buildFieldNameWithPrefix('vlm.llm_id', prefix)}
- label={t('chat.model')}
- modelTypes={ModelTypeMap.img2txt_id}
- allowClear
- ownerTenantId={ownerTenantId}
- />
- </>
- );
-}
diff --git a/web/src/pages/agent/form/parser-form/common-form-fields.tsx b/web/src/pages/agent/form/parser-form/common-form-fields.tsx
index 23f343a65d..b18bec32c4 100644
--- a/web/src/pages/agent/form/parser-form/common-form-fields.tsx
+++ b/web/src/pages/agent/form/parser-form/common-form-fields.tsx
@@ -1,5 +1,8 @@
import { useCrossLanguageOptions } from '@/components/cross-language-form-field';
-import { LayoutRecognizeFormField } from '@/components/layout-recognize-form-field';
+import {
+ LayoutRecognizeFormField,
+ ParseDocumentType,
+} from '@/components/layout-recognize-form-field';
import {
SelectWithSearch,
SelectWithSearchFlagOptionType,
@@ -8,6 +11,8 @@ import { RAGFlowFormItem } from '@/components/ragflow-form';
import { Switch } from '@/components/ui/switch';
import { FileType } from '@/constants/file';
import { upperCase, upperFirst } from 'lodash';
+import { useEffect } from 'react';
+import { useFormContext, useWatch } from 'react-hook-form';
import { useTranslation } from 'react-i18next';
import { useOwnerTenantId } from '../../context';
import {
@@ -74,29 +79,94 @@ export function ParserMethodFormField({
);
}
-export function FlattenMediaToTextFormField({ prefix }: CommonProps) {
- const { t } = useTranslation();
+const TableResultTypeOptions: SelectWithSearchFlagOptionType[] = [
+ { label: 'Markdown', value: '0' },
+ { label: 'HTML', value: '1' },
+];
+
+const MarkdownImageResponseTypeOptions: SelectWithSearchFlagOptionType[] = [
+ { label: 'URL', value: '0' },
+ { label: 'Text', value: '1' },
+];
+
+type TcadpSelectFieldProps = CommonProps & {
+ name: string;
+ label: string;
+ options: SelectWithSearchFlagOptionType[];
+};
+
+function TcadpSelectField({
+ prefix,
+ name,
+ label,
+ options,
+}: TcadpSelectFieldProps) {
return (
<RAGFlowFormItem
- name={buildFieldNameWithPrefix(`flatten_media_to_text`, prefix)}
- label={t('flow.flattenMediaToText')}
- tooltip={t('flow.flattenMediaToTextTip')}
- horizontal={true}
- labelClassName="w-full"
- valueClassName="w-8"
+ name={buildFieldNameWithPrefix(name, prefix)}
+ label={label}
>
{(field) => (
- <Switch
- checked={field.value}
- onCheckedChange={(checked) => {
- field.onChange?.(checked);
- }}
- />
+ <SelectWithSearch
+ value={field.value}
+ onChange={field.onChange}
+ options={options}
+ ></SelectWithSearch>
)}
</RAGFlowFormItem>
);
}
+// TCADP parser options shared by the PDF and spreadsheet forms. Visible only
+// when TCADP is the parse method, and seeds the default values on selection.
+export function TcadpFormFields({ prefix }: CommonProps) {
+ const { t } = useTranslation();
+ const form = useFormContext();
+
+ const parseMethod = useWatch({
+ name: buildFieldNameWithPrefix('parse_method', prefix),
+ });
+ const shown = !!parseMethod && parseMethod === ParseDocumentType.TCADPParser;
+
+ // Set default values for TCADP options when TCADP is selected
+ useEffect(() => {
+ if (!shown) {
+ return;
+ }
+ const names = ['table_result_type', 'markdown_image_response_type'];
+ names.forEach((name) => {
+ const fieldName = buildFieldNameWithPrefix(name, prefix);
+ if (!form.getValues(fieldName)) {
+ form.setValue(fieldName, '1', {
+ shouldValidate: true,
+ shouldDirty: true,
+ });
+ }
+ });
+ }, [shown, form, prefix]);
+
+ if (!shown) {
+ return null;
+ }
+
+ return (
+ <>
+ <TcadpSelectField
+ prefix={prefix}
+ name="table_result_type"
+ label={t('flow.tableResultType') || '表格返回形式'}
+ options={TableResultTypeOptions}
+ />
+ <TcadpSelectField
+ prefix={prefix}
+ name="markdown_image_response_type"
+ label={t('flow.markdownImageResponseType') || '图片返回形式'}
+ options={MarkdownImageResponseTypeOptions}
+ />
+ </>
+ );
+}
+
export function TwoColumnCheckFormField({ prefix }: CommonProps) {
const { t } = useTranslation();
return (
diff --git a/web/src/pages/agent/form/parser-form/index.tsx b/web/src/pages/agent/form/parser-form/index.tsx
index 835a3b3f30..22ec8c25fd 100644
--- a/web/src/pages/agent/form/parser-form/index.tsx
+++ b/web/src/pages/agent/form/parser-form/index.tsx
@@ -24,6 +24,7 @@ import { useWatchFormChange } from '../../hooks/use-watch-form-change';
import { INextOperatorForm } from '../../interface';
import { buildOutputList } from '../../utils/build-output-list';
import { Output } from '../components/output';
+import { AudioFormFields } from './audio-form-fields';
import { OutputFormatFormField } from './common-form-fields';
import { EmailFormFields } from './email-form-fields';
import { ImageFormFields } from './image-form-fields';
@@ -35,8 +36,8 @@ import {
HtmlFormFields,
TextMarkdownFormFields,
} from './text-html-form-fields';
-import { buildInitialParserSetup } from './utils';
-import { AudioFormFields, VideoFormFields } from './video-form-fields';
+import { buildInitialParserSetup, normalizeParserFormValues } from './utils';
+import { VisionEnhancementFormFields } from './vision-form-fields';
import { WordFormFields } from './word-form-fields';
export { FormSchema } from './schema';
@@ -50,7 +51,6 @@ const FileFormatWidgetMap = {
[FileType.PowerPoint]: PptFormFields,
[FileType.Doc]: WordFormFields,
[FileType.Docx]: WordFormFields,
- [FileType.Video]: VideoFormFields,
[FileType.Audio]: AudioFormFields,
[FileType.Email]: EmailFormFields,
[FileType.Image]: ImageFormFields,
@@ -152,7 +152,13 @@ const ParserForm = ({
// Show the saved values as-is: an empty llm_id means the user cleared the
// model, and the backend falls back to the tenant default at parse time.
// Prefilling it here would make a cleared model reappear and be written back.
- const defaultValues = useFormValues(initialParserValues, node);
+ const formValues = useFormValues(initialParserValues, node);
+ // Lifts any legacy per-setup vision options onto the shared top-level
+ // fields; already-normalized values pass through unchanged.
+ const defaultValues = useMemo(
+ () => normalizeParserFormValues(formValues),
+ [formValues],
+ );
const form = useForm<z.infer<typeof FormSchema>>({
defaultValues,
@@ -196,6 +202,7 @@ const ParserForm = ({
return (
<Form {...form}>
<form className="space-y-5 px-5">
+ <VisionEnhancementFormFields />
{fields.map((field, index) => {
return (
<ParserItem
diff --git a/web/src/pages/agent/form/parser-form/pdf-form-fields.tsx b/web/src/pages/agent/form/parser-form/pdf-form-fields.tsx
index 3d197a5ec8..08ffc33687 100644
--- a/web/src/pages/agent/form/parser-form/pdf-form-fields.tsx
+++ b/web/src/pages/agent/form/parser-form/pdf-form-fields.tsx
@@ -1,24 +1,13 @@
import { ParseDocumentType } from '@/components/layout-recognize-form-field';
-import {
- ModelTreeSelectFormField,
- ModelTypeMap,
-} from '@/components/model-tree-select';
-import {
- SelectWithSearch,
- SelectWithSearchFlagOptionType,
-} from '@/components/originui/select-with-search';
-import { RAGFlowFormItem } from '@/components/ragflow-form';
import { isEmpty } from 'lodash';
-import { useEffect, useMemo } from 'react';
-import { useFormContext, useWatch } from 'react-hook-form';
-import { useTranslation } from 'react-i18next';
-import { useOwnerTenantId } from '../../context';
+import { useMemo } from 'react';
+import { useWatch } from 'react-hook-form';
import {
- FlattenMediaToTextFormField,
LanguageFormField,
ParserMethodFormField,
RemoveHeaderFooterFormField,
RmdirFormField,
+ TcadpFormFields,
TwoColumnCheckFormField,
} from './common-form-fields';
import { CommonProps } from './interface';
@@ -26,28 +15,11 @@ import { DynamicPageRange } from './dynamic-page-range';
import { useSetInitialLanguage } from './use-set-initial-language';
import { buildFieldNameWithPrefix } from './utils';
-const tableResultTypeOptions: SelectWithSearchFlagOptionType[] = [
- { label: 'Markdown', value: '0' },
- { label: 'HTML', value: '1' },
-];
-
-const markdownImageResponseTypeOptions: SelectWithSearchFlagOptionType[] = [
- { label: 'URL', value: '0' },
- { label: 'Text', value: '1' },
-];
-
export function PdfFormFields({ prefix }: CommonProps) {
- const { t } = useTranslation();
- const form = useFormContext();
- const ownerTenantId = useOwnerTenantId();
-
const parseMethodName = buildFieldNameWithPrefix('parse_method', prefix);
const parseMethod = useWatch({
name: parseMethodName,
});
- const flattenMediaToText = useWatch({
- name: buildFieldNameWithPrefix('flatten_media_to_text', prefix),
- });
const languageShown = useMemo(() => {
return (
@@ -58,41 +30,8 @@ export function PdfFormFields({ prefix }: CommonProps) {
);
}, [parseMethod]);
- const tcadpOptionsShown = useMemo(() => {
- return (
- !isEmpty(parseMethod) && parseMethod === ParseDocumentType.TCADPParser
- );
- }, [parseMethod]);
-
useSetInitialLanguage({ prefix, languageShown });
- // Set default values for TCADP options when TCADP is selected
- useEffect(() => {
- if (tcadpOptionsShown) {
- const tableResultTypeName = buildFieldNameWithPrefix(
- 'table_result_type',
- prefix,
- );
- const markdownImageResponseTypeName = buildFieldNameWithPrefix(
- 'markdown_image_response_type',
- prefix,
- );
-
- if (isEmpty(form.getValues(tableResultTypeName))) {
- form.setValue(tableResultTypeName, '1', {
- shouldValidate: true,
- shouldDirty: true,
- });
- }
- if (isEmpty(form.getValues(markdownImageResponseTypeName))) {
- form.setValue(markdownImageResponseTypeName, '1', {
- shouldValidate: true,
- shouldDirty: true,
- });
- }
- }
- }, [tcadpOptionsShown, form, prefix]);
-
return (
<>
<TwoColumnCheckFormField prefix={prefix} />
@@ -100,48 +39,9 @@ export function PdfFormFields({ prefix }: CommonProps) {
<RemoveHeaderFooterFormField prefix={prefix} />
<ParserMethodFormField prefix={prefix}></ParserMethodFormField>
<DynamicPageRange prefix={prefix} />
- <FlattenMediaToTextFormField prefix={prefix} />
- {!flattenMediaToText && (
- <ModelTreeSelectFormField
- name={buildFieldNameWithPrefix('vlm.llm_id', prefix)}
- label={t('chat.model')}
- modelTypes={ModelTypeMap.img2txt_id}
- allowClear
- ownerTenantId={ownerTenantId}
- />
- )}
+
{languageShown && <LanguageFormField prefix={prefix}></LanguageFormField>}
- {tcadpOptionsShown && (
- <>
- <RAGFlowFormItem
- name={buildFieldNameWithPrefix('table_result_type', prefix)}
- label={t('flow.tableResultType') || '表格返回形式'}
- >
- {(field) => (
- <SelectWithSearch
- value={field.value}
- onChange={field.onChange}
- options={tableResultTypeOptions}
- ></SelectWithSearch>
- )}
- </RAGFlowFormItem>
- <RAGFlowFormItem
- name={buildFieldNameWithPrefix(
- 'markdown_image_response_type',
- prefix,
- )}
- label={t('flow.markdownImageResponseType') || '图片返回形式'}
- >
- {(field) => (
- <SelectWithSearch
- value={field.value}
- onChange={field.onChange}
- options={markdownImageResponseTypeOptions}
- ></SelectWithSearch>
- )}
- </RAGFlowFormItem>
- </>
- )}
+ <TcadpFormFields prefix={prefix} />
</>
);
}
diff --git a/web/src/pages/agent/form/parser-form/schema.ts b/web/src/pages/agent/form/parser-form/schema.ts
index 0af3f8dae0..0ef1b39f89 100644
--- a/web/src/pages/agent/form/parser-form/schema.ts
+++ b/web/src/pages/agent/form/parser-form/schema.ts
@@ -10,8 +10,9 @@ export const SetupSchema = z
parse_method: z.string().optional(),
lang: z.string().optional(),
fields: z.array(z.string()).optional(),
+ // Per-setup vlm is only used by Audio (its ASR model); the vision model
+ // lives at the top level, shared by all vision-capable file types.
vlm: z.object({ llm_id: z.string().optional() }).optional(),
- flatten_media_to_text: z.boolean().optional(),
system_prompt: z.string().optional(),
table_result_type: z.string().optional(),
markdown_image_response_type: z.string().optional(),
@@ -54,6 +55,11 @@ export const SetupSchema = z
export const FormSchema = z.object({
setups: z.array(SetupSchema).min(1, i18n.t('flow.atLeastOneFileType')),
+ // Global vision enhancement: one switch + one img2txt model shared by every
+ // vision-capable file type, sitting at the params top level alongside the
+ // per-family setups (see the backend Parser component contract).
+ vlm: z.object({ llm_id: z.string().optional() }).optional(),
+ enable_vision_enhancement: z.boolean().optional(),
});
export type ParserFormSchemaType = z.infer<typeof FormSchema>;
diff --git a/web/src/pages/agent/form/parser-form/spreadsheet-form-fields.tsx b/web/src/pages/agent/form/parser-form/spreadsheet-form-fields.tsx
index 934d5526e5..7dc99bbbd6 100644
--- a/web/src/pages/agent/form/parser-form/spreadsheet-form-fields.tsx
+++ b/web/src/pages/agent/form/parser-form/spreadsheet-form-fields.tsx
@@ -1,49 +1,8 @@
import { ParseDocumentType } from '@/components/layout-recognize-form-field';
-import {
- ModelTreeSelectFormField,
- ModelTypeMap,
-} from '@/components/model-tree-select';
-import {
- SelectWithSearch,
- SelectWithSearchFlagOptionType,
-} from '@/components/originui/select-with-search';
-import { RAGFlowFormItem } from '@/components/ragflow-form';
-import { isEmpty } from 'lodash';
-import { useEffect, useMemo } from 'react';
-import { useFormContext, useWatch } from 'react-hook-form';
-import { useTranslation } from 'react-i18next';
-import { useOwnerTenantId } from '../../context';
-import {
- FlattenMediaToTextFormField,
- ParserMethodFormField,
-} from './common-form-fields';
+import { ParserMethodFormField, TcadpFormFields } from './common-form-fields';
import { CommonProps } from './interface';
-import { buildFieldNameWithPrefix } from './utils';
-
-const tableResultTypeOptions: SelectWithSearchFlagOptionType[] = [
- { label: 'Markdown', value: '0' },
- { label: 'HTML', value: '1' },
-];
-
-const markdownImageResponseTypeOptions: SelectWithSearchFlagOptionType[] = [
- { label: 'URL', value: '0' },
- { label: 'Text', value: '1' },
-];
export function SpreadsheetFormFields({ prefix }: CommonProps) {
- const { t } = useTranslation();
- const form = useFormContext();
- const ownerTenantId = useOwnerTenantId();
-
- const parseMethodName = buildFieldNameWithPrefix('parse_method', prefix);
-
- const parseMethod = useWatch({
- name: parseMethodName,
- });
- const flattenMediaToText = useWatch({
- name: buildFieldNameWithPrefix('flatten_media_to_text', prefix),
- });
-
// Spreadsheet only supports DeepDOC and TCADPParser
const optionsWithoutLLM = [
{ label: ParseDocumentType.DeepDOC, value: ParseDocumentType.DeepDOC },
@@ -53,86 +12,13 @@ export function SpreadsheetFormFields({ prefix }: CommonProps) {
},
];
- const tcadpOptionsShown = useMemo(() => {
- return (
- !isEmpty(parseMethod) && parseMethod === ParseDocumentType.TCADPParser
- );
- }, [parseMethod]);
-
- // Set default values for TCADP options when TCADP is selected
- useEffect(() => {
- if (tcadpOptionsShown) {
- const tableResultTypeName = buildFieldNameWithPrefix(
- 'table_result_type',
- prefix,
- );
- const markdownImageResponseTypeName = buildFieldNameWithPrefix(
- 'markdown_image_response_type',
- prefix,
- );
-
- if (isEmpty(form.getValues(tableResultTypeName))) {
- form.setValue(tableResultTypeName, '1', {
- shouldValidate: true,
- shouldDirty: true,
- });
- }
- if (isEmpty(form.getValues(markdownImageResponseTypeName))) {
- form.setValue(markdownImageResponseTypeName, '1', {
- shouldValidate: true,
- shouldDirty: true,
- });
- }
- }
- }, [tcadpOptionsShown, form, prefix]);
-
return (
<>
<ParserMethodFormField
prefix={prefix}
optionsWithoutLLM={optionsWithoutLLM}
></ParserMethodFormField>
- <FlattenMediaToTextFormField prefix={prefix} />
- {!flattenMediaToText && (
- <ModelTreeSelectFormField
- name={buildFieldNameWithPrefix('vlm.llm_id', prefix)}
- label={t('chat.model')}
- modelTypes={ModelTypeMap.img2txt_id}
- allowClear
- ownerTenantId={ownerTenantId}
- />
- )}
- {tcadpOptionsShown && (
- <>
- <RAGFlowFormItem
- name={buildFieldNameWithPrefix('table_result_type', prefix)}
- label={t('flow.tableResultType') || '表格返回形式'}
- >
- {(field) => (
- <SelectWithSearch
- value={field.value}
- onChange={field.onChange}
- options={tableResultTypeOptions}
- ></SelectWithSearch>
- )}
- </RAGFlowFormItem>
- <RAGFlowFormItem
- name={buildFieldNameWithPrefix(
- 'markdown_image_response_type',
- prefix,
- )}
- label={t('flow.markdownImageResponseType') || '图片返回形式'}
- >
- {(field) => (
- <SelectWithSearch
- value={field.value}
- onChange={field.onChange}
- options={markdownImageResponseTypeOptions}
- ></SelectWithSearch>
- )}
- </RAGFlowFormItem>
- </>
- )}
+ <TcadpFormFields prefix={prefix} />
</>
);
}
diff --git a/web/src/pages/agent/form/parser-form/text-html-form-fields.tsx b/web/src/pages/agent/form/parser-form/text-html-form-fields.tsx
index 33b480da4b..1df1a71e58 100644
--- a/web/src/pages/agent/form/parser-form/text-html-form-fields.tsx
+++ b/web/src/pages/agent/form/parser-form/text-html-form-fields.tsx
@@ -1,38 +1,13 @@
import {
- ModelTreeSelectFormField,
- ModelTypeMap,
-} from '@/components/model-tree-select';
-import { useWatch } from 'react-hook-form';
-import { useTranslation } from 'react-i18next';
-import { useOwnerTenantId } from '../../context';
-import {
- FlattenMediaToTextFormField,
RemoveHeaderFooterFormField,
RmdirFormField,
} from './common-form-fields';
import { CommonProps } from './interface';
-import { buildFieldNameWithPrefix } from './utils';
export function TextMarkdownFormFields({ prefix }: CommonProps) {
- const { t } = useTranslation();
- const ownerTenantId = useOwnerTenantId();
- const flattenMediaToText = useWatch({
- name: buildFieldNameWithPrefix('flatten_media_to_text', prefix),
- });
-
return (
<>
<RmdirFormField prefix={prefix} />
- <FlattenMediaToTextFormField prefix={prefix} />
- {!flattenMediaToText && (
- <ModelTreeSelectFormField
- name={buildFieldNameWithPrefix('vlm.llm_id', prefix)}
- label={t('chat.model')}
- modelTypes={ModelTypeMap.img2txt_id}
- allowClear
- ownerTenantId={ownerTenantId}
- />
- )}
</>
);
}
diff --git a/web/src/pages/agent/form/parser-form/utils.test.ts b/web/src/pages/agent/form/parser-form/utils.test.ts
index 764ec77573..93774e3a41 100644
--- a/web/src/pages/agent/form/parser-form/utils.test.ts
+++ b/web/src/pages/agent/form/parser-form/utils.test.ts
@@ -1,7 +1,11 @@
import { FileType } from '@/constants/file';
import { ModelTypeToField } from '@/constants/llm';
import { initialParserValues } from '../../constant/pipeline';
-import { buildInitialParserSetup, buildInitialParserValues } from './utils';
+import {
+ buildInitialParserSetup,
+ buildInitialParserValues,
+ normalizeParserFormValues,
+} from './utils';
describe('parser-form utils', () => {
describe('buildInitialParserSetup', () => {
@@ -15,14 +19,16 @@ describe('parser-form utils', () => {
);
});
- it('prefills the tenant default model for video and audio', () => {
+ it('prefills the tenant default model for audio only', () => {
const dictionary = {
[ModelTypeToField.vision]: 'gpt-4o@OpenAI',
[ModelTypeToField.asr]: 'whisper@OpenAI',
};
- expect(buildInitialParserSetup(FileType.Video, dictionary)?.vlm).toEqual({
- llm_id: 'gpt-4o@OpenAI',
- });
+ // The vision model is a global parser option, so video no longer gets
+ // a per-setup prefill.
+ expect(buildInitialParserSetup(FileType.Video, dictionary)?.vlm).toBe(
+ undefined,
+ );
expect(buildInitialParserSetup(FileType.Audio, dictionary)?.vlm).toEqual({
llm_id: 'whisper@OpenAI',
});
@@ -50,18 +56,19 @@ describe('parser-form utils', () => {
});
describe('buildInitialParserValues', () => {
- it('prefills tenant default models across every default file type', () => {
+ it('prefills the global vision model and keeps the enhancement off', () => {
const values = buildInitialParserValues({
[ModelTypeToField.vision]: 'gpt-4o@OpenAI',
[ModelTypeToField.asr]: 'whisper@OpenAI',
});
+ expect(values.vlm).toEqual({ llm_id: 'gpt-4o@OpenAI' });
+ expect(values.enable_vision_enhancement).toBe(false);
+
const byFileType = new Map(
values.setups.map((setup: any) => [setup.fileFormat, setup]),
);
- expect(byFileType.get(FileType.Video)?.vlm).toEqual({
- llm_id: 'gpt-4o@OpenAI',
- });
+ expect(byFileType.get(FileType.Video)?.vlm).toBe(undefined);
expect(byFileType.get(FileType.Audio)?.vlm).toEqual({
llm_id: 'whisper@OpenAI',
});
@@ -76,4 +83,88 @@ describe('parser-form utils', () => {
);
});
});
+
+ describe('normalizeParserFormValues', () => {
+ it('lifts legacy per-setup vision values onto the top level', () => {
+ const values = normalizeParserFormValues({
+ setups: [
+ {
+ fileFormat: FileType.PDF,
+ flatten_media_to_text: false,
+ vlm: { llm_id: 'gpt-4o@OpenAI' },
+ },
+ { fileFormat: FileType.Audio, vlm: { llm_id: 'whisper@OpenAI' } },
+ ],
+ });
+
+ expect(values.enable_vision_enhancement).toBe(true);
+ expect(values.vlm).toEqual({ llm_id: 'gpt-4o@OpenAI' });
+ expect(values.setups[0]).toEqual({ fileFormat: FileType.PDF });
+ // Audio keeps its per-setup ASR model.
+ expect(values.setups[1]).toEqual({
+ fileFormat: FileType.Audio,
+ vlm: { llm_id: 'whisper@OpenAI' },
+ });
+ });
+
+ it('turns the enhancement off when no setup used the vision model', () => {
+ const values = normalizeParserFormValues({
+ setups: [
+ { fileFormat: FileType.PDF, flatten_media_to_text: true },
+ { fileFormat: FileType.Docx, flatten_media_to_text: true },
+ ],
+ });
+
+ expect(values.enable_vision_enhancement).toBe(false);
+ expect(values.vlm).toEqual({ llm_id: '' });
+ });
+
+ it('treats a non-empty per-setup model as enabled enhancement', () => {
+ const values = normalizeParserFormValues({
+ setups: [
+ { fileFormat: FileType.Video, vlm: { llm_id: 'gpt-4o@OpenAI' } },
+ ],
+ });
+
+ expect(values.enable_vision_enhancement).toBe(true);
+ expect(values.vlm).toEqual({ llm_id: 'gpt-4o@OpenAI' });
+ expect(values.setups[0]).toEqual({ fileFormat: FileType.Video });
+ });
+
+ it('is idempotent on already-normalized values', () => {
+ const values = {
+ vlm: { llm_id: 'gpt-4o@OpenAI' },
+ enable_vision_enhancement: true,
+ setups: [
+ {
+ fileFormat: FileType.PDF,
+ flatten_media_to_text: true,
+ vlm: { llm_id: 'stale@Model' },
+ },
+ { fileFormat: FileType.Audio, vlm: { llm_id: 'whisper@OpenAI' } },
+ ],
+ };
+
+ const normalized = normalizeParserFormValues(values);
+ expect(normalized).toEqual({
+ vlm: { llm_id: 'gpt-4o@OpenAI' },
+ enable_vision_enhancement: true,
+ setups: [
+ { fileFormat: FileType.PDF },
+ { fileFormat: FileType.Audio, vlm: { llm_id: 'whisper@OpenAI' } },
+ ],
+ });
+ expect(normalizeParserFormValues(normalized)).toEqual(normalized);
+ });
+
+ it('keeps an explicitly cleared top-level model', () => {
+ const values = normalizeParserFormValues({
+ vlm: { llm_id: '' },
+ enable_vision_enhancement: true,
+ setups: [{ fileFormat: FileType.PDF, vlm: { llm_id: 'stale@Model' } }],
+ });
+
+ expect(values.vlm).toEqual({ llm_id: '' });
+ });
+ });
});
diff --git a/web/src/pages/agent/form/parser-form/utils.ts b/web/src/pages/agent/form/parser-form/utils.ts
index 05d35248d3..a06f50b6c9 100644
--- a/web/src/pages/agent/form/parser-form/utils.ts
+++ b/web/src/pages/agent/form/parser-form/utils.ts
@@ -1,5 +1,8 @@
import { FileType } from '@/constants/file';
-import { cloneDeep } from 'lodash';
+import { ModelTypeToField } from '@/constants/llm';
+import cloneDeep from 'lodash/cloneDeep';
+import isEmpty from 'lodash/isEmpty';
+import omit from 'lodash/omit';
import {
FileTypeDefaultModelFieldMap,
initialParserValues,
@@ -9,8 +12,59 @@ export function buildFieldNameWithPrefix(name: string, prefix: string) {
return `${prefix}.${name}`;
}
+// File types whose image/table blocks go through the vision model when the
+// global enhancement is on. Audio is excluded: its per-setup model is an ASR
+// model, not a vision one.
+export const VisionEnhancementFileTypes: FileType[] = [
+ FileType.PDF,
+ FileType.Spreadsheet,
+ FileType.Doc,
+ FileType.Docx,
+ FileType.TextMarkdown,
+ FileType.Video,
+];
+
+// Lifts the legacy per-setup vision options (vlm.llm_id / flatten_media_to_text)
+// onto the top level. Idempotent, so every boundary (form defaults, canvas
+// save, dataset load) can apply it: values already carrying a top-level
+// enable_vision_enhancement keep it, and per-setup leftovers are always
+// stripped (except Audio's vlm, which holds the ASR model).
+export function normalizeParserFormValues<T extends Record<string, any>>(
+ values: T,
+): T & { vlm: { llm_id: string }; enable_vision_enhancement: boolean } {
+ const setups = Array.isArray(values?.setups) ? values.setups : [];
+ const visionSetups = setups.filter((x) =>
+ VisionEnhancementFileTypes.includes(x?.fileFormat),
+ );
+
+ const enableVisionEnhancement =
+ typeof values?.enable_vision_enhancement === 'boolean'
+ ? values.enable_vision_enhancement
+ : visionSetups.some(
+ (x) => x?.flatten_media_to_text === false || !isEmpty(x?.vlm?.llm_id),
+ );
+
+ const llmId =
+ values?.vlm?.llm_id ??
+ visionSetups.find((x) => !isEmpty(x?.vlm?.llm_id))?.vlm?.llm_id ??
+ '';
+
+ const nextSetups = setups.map((x) =>
+ x?.fileFormat === FileType.Audio
+ ? x
+ : omit(x, ['vlm', 'flatten_media_to_text']),
+ );
+
+ return {
+ ...values,
+ vlm: { ...values?.vlm, llm_id: llmId },
+ enable_vision_enhancement: enableVisionEnhancement,
+ setups: nextSetups,
+ };
+}
+
// Builds the default setup for a file type being added to the parser form,
-// prefilling the tenant default model for types that have one (video/audio).
+// prefilling the tenant default model for types that have one (audio).
export function buildInitialParserSetup(
fileType: FileType,
defaultModelDictionary: Record<string, string>,
@@ -32,13 +86,16 @@ export function buildInitialParserSetup(
}
// Form values for a freshly added Parser node: every default file type with its
-// tenant default model prefilled. Only applies to node creation — an existing
-// form is shown as saved, so a cleared model stays cleared.
+// tenant default model prefilled, and the global vision model prefilled from
+// the tenant's image2text default (enhancement itself stays off). Only applies
+// to node creation — an existing form is shown as saved, so a cleared model
+// stays cleared.
export function buildInitialParserValues(
defaultModelDictionary: Record<string, string>,
) {
return {
...initialParserValues,
+ vlm: { llm_id: defaultModelDictionary[ModelTypeToField.vision] ?? '' },
setups: initialParserValues.setups.map(
(setup) =>
buildInitialParserSetup(
diff --git a/web/src/pages/agent/form/parser-form/vision-form-fields.tsx b/web/src/pages/agent/form/parser-form/vision-form-fields.tsx
new file mode 100644
index 0000000000..7dbf164962
--- /dev/null
+++ b/web/src/pages/agent/form/parser-form/vision-form-fields.tsx
@@ -0,0 +1,82 @@
+import {
+ ModelTreeSelectFormField,
+ ModelTypeMap,
+} from '@/components/model-tree-select';
+import { RAGFlowFormItem } from '@/components/ragflow-form';
+import { Switch } from '@/components/ui/switch';
+import { ModelTypeToField } from '@/constants/llm';
+import { useFetchDefaultModelDictionary } from '@/hooks/use-llm-request';
+import { useCallback } from 'react';
+import { useFormContext, useWatch } from 'react-hook-form';
+import { useTranslation } from 'react-i18next';
+import { useOwnerTenantId } from '../../context';
+import { ParserFormSchemaType } from './schema';
+
+type VisionEnhancementSwitchProps = {
+ enabled: boolean | undefined;
+ onEnabledChange: (value: boolean) => void;
+};
+
+// Turning the enhancement on prefills an empty model with the tenant's
+// image2text default.
+function VisionEnhancementSwitch({
+ enabled,
+ onEnabledChange,
+}: VisionEnhancementSwitchProps) {
+ const form = useFormContext<ParserFormSchemaType>();
+ const defaultModelDictionary = useFetchDefaultModelDictionary();
+
+ const handleCheckedChange = useCallback(
+ (checked: boolean) => {
+ onEnabledChange(checked);
+ if (checked && !form.getValues('vlm.llm_id')) {
+ const modelId = defaultModelDictionary[ModelTypeToField.vision];
+ if (modelId) {
+ form.setValue('vlm.llm_id', modelId, { shouldDirty: true });
+ }
+ }
+ },
+ [onEnabledChange, form, defaultModelDictionary],
+ );
+
+ return <Switch checked={!!enabled} onCheckedChange={handleCheckedChange} />;
+}
+
+// Global vision enhancement: one switch + img2txt model shared by every
+// vision-capable file type. The fields live at the form's top level
+// (`enable_vision_enhancement` and `vlm.llm_id`) alongside `setups`, matching
+// the backend Parser component's params contract.
+export function VisionEnhancementFormFields() {
+ const { t } = useTranslation();
+ const ownerTenantId = useOwnerTenantId();
+ const enabled = useWatch({ name: 'enable_vision_enhancement' });
+
+ return (
+ <>
+ <RAGFlowFormItem
+ name="enable_vision_enhancement"
+ label={t('flow.enableVisionEnhancement')}
+ tooltip={t('flow.enableVisionEnhancementTip')}
+ horizontal={true}
+ labelClassName="w-full"
+ valueClassName="w-8"
+ >
+ {(field) => (
+ <VisionEnhancementSwitch
+ enabled={field.value}
+ onEnabledChange={field.onChange}
+ />
+ )}
+ </RAGFlowFormItem>
+ {enabled && (
+ <ModelTreeSelectFormField
+ name="vlm.llm_id"
+ label={t('chat.model')}
+ modelTypes={ModelTypeMap.img2txt_id}
+ allowClear
+ ownerTenantId={ownerTenantId}
+ />
+ )}
+ </>
+ );
+}
diff --git a/web/src/pages/agent/form/parser-form/word-form-fields.tsx b/web/src/pages/agent/form/parser-form/word-form-fields.tsx
index 77ce5314fa..9be304d48d 100644
--- a/web/src/pages/agent/form/parser-form/word-form-fields.tsx
+++ b/web/src/pages/agent/form/parser-form/word-form-fields.tsx
@@ -1,39 +1,14 @@
import {
- ModelTreeSelectFormField,
- ModelTypeMap,
-} from '@/components/model-tree-select';
-import { useWatch } from 'react-hook-form';
-import { useTranslation } from 'react-i18next';
-import { useOwnerTenantId } from '../../context';
-import {
- FlattenMediaToTextFormField,
OutputFormatFormFieldProps,
RemoveHeaderFooterFormField,
RmdirFormField,
} from './common-form-fields';
-import { buildFieldNameWithPrefix } from './utils';
export function WordFormFields({ prefix }: OutputFormatFormFieldProps) {
- const { t } = useTranslation();
- const ownerTenantId = useOwnerTenantId();
- const flattenMediaToText = useWatch({
- name: buildFieldNameWithPrefix('flatten_media_to_text', prefix),
- });
-
return (
<>
<RmdirFormField prefix={prefix} />
<RemoveHeaderFooterFormField prefix={prefix} />
- <FlattenMediaToTextFormField prefix={prefix} />
- {!flattenMediaToText && (
- <ModelTreeSelectFormField
- name={buildFieldNameWithPrefix('vlm.llm_id', prefix)}
- label={t('chat.model')}
- modelTypes={ModelTypeMap.img2txt_id}
- allowClear
- ownerTenantId={ownerTenantId}
- />
- )}
</>
);
}
diff --git a/web/src/pages/agent/utils.ts b/web/src/pages/agent/utils.ts
index 042a2a1e37..03e88f8fc3 100644
--- a/web/src/pages/agent/utils.ts
+++ b/web/src/pages/agent/utils.ts
@@ -41,6 +41,7 @@ import { BeginFormSchemaType } from './form/begin-form/schema';
import { DataOperationsFormSchemaType } from './form/data-operations-form';
import { ExtractorFormSchemaType } from './form/extractor-form';
import { ParserFormSchemaType } from './form/parser-form';
+import { normalizeParserFormValues } from './form/parser-form/utils';
import { TitleChunkerFormSchemaType } from './form/title-chunker-form';
import { TokenChunkerFormSchemaType } from './form/token-chunker-form';
import { BeginQuery, IPosition } from './interface';
@@ -204,7 +205,12 @@ function transformObjectArrayToPureArray(
}
export function transformParserParams(params: ParserFormSchemaType) {
- const setups = params.setups.reduce<
+ // The vision options (enable_vision_enhancement / vlm.llm_id) live at the
+ // params top level alongside the per-family setups — the backend reads them
+ // there, so they pass through as-is. normalizeParserFormValues also lifts
+ // legacy per-setup values of nodes saved before the move.
+ const normalizedParams = normalizeParserFormValues(params);
+ const setups = normalizedParams.setups.reduce<
Record<string, ParserFormSchemaType['setups'][0]>
>((pre, cur, index) => {
if (cur.fileFormat) {
@@ -227,8 +233,6 @@ export function transformParserParams(params: ParserFormSchemaType) {
...filteredSetup,
parse_method: cur.parse_method,
lang: cur.lang,
- vlm: { llm_id: cur.vlm?.llm_id },
- flatten_media_to_text: cur.flatten_media_to_text,
enable_multi_column: cur.enable_multi_column,
remove_toc: cur.remove_toc,
remove_header_footer: cur.remove_header_footer || false,
@@ -245,8 +249,6 @@ export function transformParserParams(params: ParserFormSchemaType) {
filteredSetup = {
...filteredSetup,
parse_method: cur.parse_method,
- vlm: { llm_id: cur.vlm?.llm_id },
- flatten_media_to_text: cur.flatten_media_to_text,
};
// Only include TCADP parameters if TCADP Parser is selected
if (cur.parse_method?.toLowerCase() === 'tcadp parser') {
@@ -284,8 +286,6 @@ export function transformParserParams(params: ParserFormSchemaType) {
case FileType.Doc:
filteredSetup = {
...filteredSetup,
- vlm: { llm_id: cur.vlm?.llm_id },
- flatten_media_to_text: cur.flatten_media_to_text,
remove_toc: cur.remove_toc,
remove_header_footer: cur.remove_header_footer || false,
};
@@ -293,8 +293,6 @@ export function transformParserParams(params: ParserFormSchemaType) {
case FileType.Docx:
filteredSetup = {
...filteredSetup,
- vlm: { llm_id: cur.vlm?.llm_id },
- flatten_media_to_text: cur.flatten_media_to_text,
remove_toc: cur.remove_toc,
remove_header_footer: cur.remove_header_footer || false,
};
@@ -309,13 +307,12 @@ export function transformParserParams(params: ParserFormSchemaType) {
case FileType.TextMarkdown:
filteredSetup = {
...filteredSetup,
- vlm: { llm_id: cur.vlm?.llm_id },
- flatten_media_to_text: cur.flatten_media_to_text,
remove_toc: cur.remove_toc,
};
break;
- case FileType.Video:
case FileType.Audio:
+ // Audio keeps its own per-setup model: it is an ASR model, not the
+ // shared vision one.
filteredSetup = {
...filteredSetup,
vlm: { llm_id: cur.vlm?.llm_id },
diff --git a/web/src/utils/pipeline-operator.ts b/web/src/utils/pipeline-operator.ts
index 1be071b63e..ae22069e9b 100644
--- a/web/src/utils/pipeline-operator.ts
+++ b/web/src/utils/pipeline-operator.ts
@@ -67,14 +67,24 @@ export function transformParserConfigSetups(
// a Python canvas) nest the per-family setups under a "setups" key; the Go
// shape lists file families at the top level. Lift the nested group —
// top-level families win — and drop the non-family protocol keys so they
- // never render as a bogus file format.
+ // never render as a bogus file format. The vision options
+ // (enable_vision_enhancement / vlm) are global params-level keys, not
+ // families, so they are dropped here too (the caller lifts them back onto
+ // the form's top level).
let source = setups;
const nested = source.setups;
if (nested && typeof nested === 'object' && !Array.isArray(nested)) {
source = { ...nested, ...omit(source, ['setups']) };
}
- return Object.entries(omit(source, ['outputs', 'allowed_output_format']))
+ return Object.entries(
+ omit(source, [
+ 'outputs',
+ 'allowed_output_format',
+ 'enable_vision_enhancement',
+ 'vlm',
+ ]),
+ )
.map(([fileFormat, config]) => {
const { pages, ...rest } = (config ?? {}) as Record<string, any>;
const normalizedPages = Array.isArray(pages)
@@ -333,7 +343,16 @@ export function transformApiConfigToForm(
): Record<string, any> {
switch (operatorType) {
case Operator.Parser:
- return { setups: transformParserConfigSetups(config) };
+ // Lift the global vision options back onto the form's top level; omit
+ // them when absent so an explicit undefined cannot clobber template or
+ // default baselines downstream (Object.assign/spread copies undefined).
+ return {
+ setups: transformParserConfigSetups(config),
+ ...(config?.enable_vision_enhancement !== undefined
+ ? { enable_vision_enhancement: config.enable_vision_enhancement }
+ : {}),
+ ...(config?.vlm !== undefined ? { vlm: config.vlm } : {}),
+ };
case Operator.Extractor:
return transformExtractorConfigToForm(config);
case Operator.Tokenizer: