RAGFlow is an agent-native database: a Retrieval-Augmented Generation (RAG) engine that grew into context infrastructure.
Go agent-harnessagent-native-databaseagentic-aiagentic-retrievalagentic-searchaiai-agentscontext-enginecontext-engineeringcontext-managementharness-engineeringknowledge-compilationragretrieval-augmented-generationretrieval-harnesssearch-harness
import { FileType } from '@/constants/file';
import i18n from '@/locales/config';
import { z } from 'zod';
export const SetupSchema = z
.object({
fileFormat: z.string().nullish(),
// preprocess: z.array(z.string()).optional(),
parse_method: z.string().optional(),
mineru_parse_method: z.string().optional(),
mineru_lang: z.string().optional(),
mineru_formula_enable: z.boolean().optional(),
mineru_table_enable: z.boolean().optional(),
// Image OCR runs independently of the global vision-enhancement switch.
// The dropdown used to encode this choice in parse_method; ocr_enabled is
// the new boolean contract. Legacy setups may still carry a parse_method
// only; normalizeParserFormValues migrates them on load.
ocr_enabled: z.boolean().optional(),
ocr_model_id: 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(),
system_prompt: z.string().optional(),
table_result_type: z.string().optional(),
markdown_image_response_type: z.string().optional(),
enable_multi_column: z.boolean().optional(),
remove_toc: z.boolean().optional(),
remove_header_footer: z.boolean().optional(),
pages: z
.array(
z
.object({
// Keep these checks on the fields themselves: an object-level
// `superRefine` is skipped whenever the base shape fails to
// parse, so one missing sibling would silently swallow the
// other field's error.
from: z.coerce
.number()
.int(i18n.t('knowledgeDetails.pageRangeFromInvalid'))
.min(1, i18n.t('knowledgeDetails.pageRangeFromInvalid')),
to: z.coerce
.number()
.int(i18n.t('knowledgeDetails.pageRangeToInvalid'))
.min(1, i18n.t('knowledgeDetails.pageRangeToInvalid')),
})
.refine(({ from, to }) => to >= from, {
path: ['to'],
message: i18n.t('knowledgeDetails.pageRangeToInvalid'),
}),
)
.optional(),
})
.superRefine((values, ctx) => {
if (values.fileFormat === FileType.Email && !values.fields?.length) {
ctx.addIssue({
code: z.ZodIssueCode.custom,
path: ['fields'],
message: 'Fields is required',
});
}
});
export const FormSchema = z
.object({
setups: z.array(SetupSchema).min(1, i18n.t('flow.atLeastOneFileType')),
// Global vision enhancement: one switch + img2txt model + description prompt,
// sitting at the params top level alongside the per-family setups (see the
// backend Parser component contract). Response language is intentionally not
// here: captions are indexed text, so they follow the knowledge base language
// that also drives tokenization, and can never disagree with it.
vlm: z
.object({
llm_id: z.string().optional(),
system_prompt: z.string().optional(),
})
.optional(),
enable_vision_enhancement: z.boolean().optional(),
})
.superRefine((values, ctx) => {
// An image needs at least one text source: local OCR or the global vision
// enhancement. Reject the dead configuration at save time so a canvas that
// indexes nothing cannot be written (mirrors the backend Check()).
if (values.enable_vision_enhancement) {
return;
}
const imageIndex = values.setups.findIndex(
(setup) =>
setup.fileFormat === FileType.Image && setup.ocr_enabled === false,
);
if (imageIndex >= 0) {
ctx.addIssue({
code: z.ZodIssueCode.custom,
path: ['setups', imageIndex, 'ocr_enabled'],
message: i18n.t('flow.imageTextSourceRequired'),
});
}
});
export type ParserFormSchemaType = z.infer<typeof FormSchema>;