Added semi-automatic mode to the metadata filter (#11886)

### What problem does this PR solve?

Retrieval metadata filtering adds semi-automatic mode, and users can
manually check the metadata key that participates in LLM to generate
filter conditions.
### Type of change

- [x] New Feature (non-breaking change which adds functionality)

agent/tools/retrieval.py

@@ -136,6 +136,16 @@ class Retrieval(ToolBase, ABC):
                 doc_ids.extend(meta_filter(metas, filters["conditions"], filters.get("logic", "and")))
                 if not doc_ids:
                     doc_ids = None
+            elif self._param.meta_data_filter.get("method") == "semi_auto":
+                selected_keys = self._param.meta_data_filter.get("semi_auto", [])
+                if selected_keys:
+                    filtered_metas = {key: metas[key] for key in selected_keys if key in metas}
+                    if filtered_metas:
+                        chat_mdl = LLMBundle(self._canvas.get_tenant_id(), LLMType.CHAT)
+                        filters: dict = gen_meta_filter(chat_mdl, filtered_metas, query)
+                        doc_ids.extend(meta_filter(metas, filters["conditions"], filters.get("logic", "and")))
+                        if not doc_ids:
+                            doc_ids = None
             elif self._param.meta_data_filter.get("method") == "manual":
                 filters = self._param.meta_data_filter["manual"]
                 for flt in filters:

api/apps/chunk_app.py

@@ -327,6 +327,40 @@ async def retrieval_test():
                 local_doc_ids.extend(meta_filter(metas, filters["conditions"], filters.get("logic", "and")))
                 if not local_doc_ids:
                     local_doc_ids = None
+            elif meta_data_filter.get("method") == "semi_auto":
+                selected_keys = meta_data_filter.get("semi_auto", [])
+                if selected_keys:
+                    filtered_metas = {key: metas[key] for key in selected_keys if key in metas}
+                    if filtered_metas:
+                        chat_mdl = LLMBundle(user_id, LLMType.CHAT, llm_name=search_config.get("chat_id", ""))
+                        filters: dict = gen_meta_filter(chat_mdl, filtered_metas, question)
+                        local_doc_ids.extend(meta_filter(metas, filters["conditions"], filters.get("logic", "and")))
+                        if not local_doc_ids:
+                            local_doc_ids = None
+            elif meta_data_filter.get("method") == "manual":
+                local_doc_ids.extend(meta_filter(metas, meta_data_filter["manual"], meta_data_filter.get("logic", "and")))
+                if meta_data_filter["manual"] and not local_doc_ids:
+                    local_doc_ids = ["-999"]
+        else:
+            meta_data_filter = req.get("meta_data_filter")
+            if meta_data_filter:
+                metas = DocumentService.get_meta_by_kbs(kb_ids)
+                if meta_data_filter.get("method") == "auto":
+                    chat_mdl = LLMBundle(user_id, LLMType.CHAT)
+                    filters: dict = gen_meta_filter(chat_mdl, metas, question)
+                    local_doc_ids.extend(meta_filter(metas, filters["conditions"], filters.get("logic", "and")))
+                    if not local_doc_ids:
+                        local_doc_ids = None
+                elif meta_data_filter.get("method") == "semi_auto":
+                    selected_keys = meta_data_filter.get("semi_auto", [])
+                    if selected_keys:
+                        filtered_metas = {key: metas[key] for key in selected_keys if key in metas}
+                        if filtered_metas:
+                            chat_mdl = LLMBundle(user_id, LLMType.CHAT)
+                            filters: dict = gen_meta_filter(chat_mdl, filtered_metas, question)
+                            local_doc_ids.extend(meta_filter(metas, filters["conditions"], filters.get("logic", "and")))
+                            if not local_doc_ids:
+                                local_doc_ids = None
                 elif meta_data_filter.get("method") == "manual":
                     local_doc_ids.extend(meta_filter(metas, meta_data_filter["manual"], meta_data_filter.get("logic", "and")))
                     if meta_data_filter["manual"] and not local_doc_ids:

api/apps/sdk/session.py

@@ -984,10 +984,45 @@ async def retrieval_test_embedded():
                 local_doc_ids.extend(meta_filter(metas, filters["conditions"], filters.get("logic", "and")))
                 if not local_doc_ids:
                     local_doc_ids = None
+            elif meta_data_filter.get("method") == "semi_auto":
+                selected_keys = meta_data_filter.get("semi_auto", [])
+                if selected_keys:
+                    filtered_metas = {key: metas[key] for key in selected_keys if key in metas}
+                    if filtered_metas:
+                        chat_mdl = LLMBundle(tenant_id, LLMType.CHAT, llm_name=search_config.get("chat_id", ""))
+                        filters: dict = gen_meta_filter(chat_mdl, filtered_metas, _question)
+                        local_doc_ids.extend(meta_filter(metas, filters["conditions"], filters.get("logic", "and")))
+                        if not local_doc_ids:
+                            local_doc_ids = None
             elif meta_data_filter.get("method") == "manual":
                 local_doc_ids.extend(meta_filter(metas, meta_data_filter["manual"], meta_data_filter.get("logic", "and")))
                 if meta_data_filter["manual"] and not local_doc_ids:
                     local_doc_ids = ["-999"]
+        else:
+            meta_data_filter = req.get("meta_data_filter")
+            if meta_data_filter:
+                metas = DocumentService.get_meta_by_kbs(kb_ids)
+                if meta_data_filter.get("method") == "auto":
+                    chat_mdl = LLMBundle(tenant_id, LLMType.CHAT)
+                    filters: dict = gen_meta_filter(chat_mdl, metas, question)
+                    local_doc_ids.extend(meta_filter(metas, filters["conditions"], filters.get("logic", "and")))
+                    if not local_doc_ids:
+                        local_doc_ids = None
+                elif meta_data_filter.get("method") == "semi_auto":
+                    selected_keys = meta_data_filter.get("semi_auto", [])
+                    if selected_keys:
+                        filtered_metas = {key: metas[key] for key in selected_keys if key in metas}
+                        if filtered_metas:
+                            chat_mdl = LLMBundle(tenant_id, LLMType.CHAT)
+                            filters: dict = gen_meta_filter(chat_mdl, filtered_metas, question)
+                            local_doc_ids.extend(meta_filter(metas, filters["conditions"], filters.get("logic", "and")))
+                            if not local_doc_ids:
+                                local_doc_ids = None
+                elif meta_data_filter.get("method") == "manual":
+                    local_doc_ids.extend(meta_filter(metas, meta_data_filter["manual"], meta_data_filter.get("logic", "and")))
+                    if meta_data_filter["manual"] and not local_doc_ids:
+                        local_doc_ids = ["-999"]
+
 
         tenants = UserTenantService.query(user_id=tenant_id)
         for kb_id in kb_ids:

api/db/services/dialog_service.py

@@ -425,6 +425,15 @@ async def async_chat(dialog, messages, stream=True, **kwargs):
             attachments.extend(meta_filter(metas, filters["conditions"], filters.get("logic", "and")))
             if not attachments:
                 attachments = None
+        elif dialog.meta_data_filter.get("method") == "semi_auto":
+            selected_keys = dialog.meta_data_filter.get("semi_auto", [])
+            if selected_keys:
+                filtered_metas = {key: metas[key] for key in selected_keys if key in metas}
+                if filtered_metas:
+                    filters: dict = gen_meta_filter(chat_mdl, filtered_metas, questions[-1])
+                    attachments.extend(meta_filter(metas, filters["conditions"], filters.get("logic", "and")))
+                    if not attachments:
+                        attachments = None
         elif dialog.meta_data_filter.get("method") == "manual":
             conds = dialog.meta_data_filter["manual"]
             attachments.extend(meta_filter(metas, conds, dialog.meta_data_filter.get("logic", "and")))
@@ -834,6 +843,15 @@ async def async_ask(question, kb_ids, tenant_id, chat_llm_name=None, search_conf
             doc_ids.extend(meta_filter(metas, filters["conditions"], filters.get("logic", "and")))
             if not doc_ids:
                 doc_ids = None
+        elif meta_data_filter.get("method") == "semi_auto":
+            selected_keys = meta_data_filter.get("semi_auto", [])
+            if selected_keys:
+                filtered_metas = {key: metas[key] for key in selected_keys if key in metas}
+                if filtered_metas:
+                    filters: dict = gen_meta_filter(chat_mdl, filtered_metas, question)
+                    doc_ids.extend(meta_filter(metas, filters["conditions"], filters.get("logic", "and")))
+                    if not doc_ids:
+                        doc_ids = None
         elif meta_data_filter.get("method") == "manual":
             doc_ids.extend(meta_filter(metas, meta_data_filter["manual"], meta_data_filter.get("logic", "and")))
             if meta_data_filter["manual"] and not doc_ids:
@@ -909,6 +927,15 @@ async def gen_mindmap(question, kb_ids, tenant_id, search_config={}):
             doc_ids.extend(meta_filter(metas, filters["conditions"], filters.get("logic", "and")))
             if not doc_ids:
                 doc_ids = None
+        elif meta_data_filter.get("method") == "semi_auto":
+            selected_keys = meta_data_filter.get("semi_auto", [])
+            if selected_keys:
+                filtered_metas = {key: metas[key] for key in selected_keys if key in metas}
+                if filtered_metas:
+                    filters: dict = gen_meta_filter(chat_mdl, filtered_metas, question)
+                    doc_ids.extend(meta_filter(metas, filters["conditions"], filters.get("logic", "and")))
+                    if not doc_ids:
+                        doc_ids = None
         elif meta_data_filter.get("method") == "manual":
             doc_ids.extend(meta_filter(metas, meta_data_filter["manual"], meta_data_filter.get("logic", "and")))
             if meta_data_filter["manual"] and not doc_ids:

web/src/components/metadata-filter/index.tsx

@@ -5,6 +5,7 @@ import { z } from 'zod';
 import { SelectWithSearch } from '../originui/select-with-search';
 import { RAGFlowFormItem } from '../ragflow-form';
 import { MetadataFilterConditions } from './metadata-filter-conditions';
+import { MetadataSemiAutoFields } from './metadata-semi-auto-fields';
 
 type MetadataFilterProps = {
   prefix?: string;
@@ -25,6 +26,9 @@ export const MetadataFilterSchema = {
           }),
         )
         .optional(),
+      semi_auto: z
+        .array(z.string()) // 修改为字符串数组
+        .optional(),
     })
     .optional(),
 };
@@ -76,6 +80,12 @@ export function MetadataFilter({
           canReference={canReference}
         ></MetadataFilterConditions>
       )}
+      {hasKnowledge && metadata === DatasetMetadata.SemiAutomatic && (
+        <MetadataSemiAutoFields
+          kbIds={kbIds}
+          prefix={prefix}
+        ></MetadataSemiAutoFields>
+      )}
     </>
   );
 }

web/src/components/metadata-filter/metadata-semi-auto-fields.tsx

@@ -0,0 +1,100 @@
+import { Button } from '@/components/ui/button';
+import {
+  DropdownMenu,
+  DropdownMenuContent,
+  DropdownMenuItem,
+  DropdownMenuTrigger,
+} from '@/components/ui/dropdown-menu';
+import {
+  FormControl,
+  FormField,
+  FormItem,
+  FormLabel,
+  FormMessage,
+} from '@/components/ui/form';
+import { Input } from '@/components/ui/input';
+import { useFetchKnowledgeMetadata } from '@/hooks/use-knowledge-request';
+import { Plus, X } from 'lucide-react';
+import { useCallback } from 'react';
+import { useFieldArray, useFormContext } from 'react-hook-form';
+import { useTranslation } from 'react-i18next';
+
+export function MetadataSemiAutoFields({
+  kbIds,
+  prefix = '',
+}: {
+  kbIds: string[];
+  prefix?: string;
+}) {
+  const { t } = useTranslation();
+  const form = useFormContext();
+  const name = prefix + 'meta_data_filter.semi_auto';
+  const metadata = useFetchKnowledgeMetadata(kbIds);
+
+  const { fields, remove, append } = useFieldArray({
+    name,
+    control: form.control,
+  });
+
+  const add = useCallback(
+    (key: string) => () => {
+      append(key); // 直接添加字符串而不是对象
+    },
+    [append],
+  );
+
+  return (
+    <section className="flex flex-col gap-2">
+      <div className="flex items-center justify-between">
+        <FormLabel>{t('chat.metadataKeys')}</FormLabel>
+        <DropdownMenu>
+          <DropdownMenuTrigger>
+            <Button variant={'ghost'} type="button">
+              <Plus />
+            </Button>
+          </DropdownMenuTrigger>
+          <DropdownMenuContent className="max-h-[300px] !overflow-y-auto scrollbar-auto">
+            {Object.keys(metadata.data).map((key, idx) => {
+              return (
+                <DropdownMenuItem key={idx} onClick={add(key)}>
+                  {key}
+                </DropdownMenuItem>
+              );
+            })}
+          </DropdownMenuContent>
+        </DropdownMenu>
+      </div>
+      <div className="space-y-5">
+        {fields.map((field, index) => {
+          // 修改字段名称以直接引用数组元素
+          const typeField = `${name}.${index}`;
+          return (
+            <section key={field.id} className="flex gap-2">
+              <div className="w-full space-y-2">
+                <FormField
+                  control={form.control}
+                  name={typeField}
+                  render={({ field }) => (
+                    <FormItem className="flex-1 overflow-hidden">
+                      <FormControl>
+                        <Input
+                          {...field}
+                          placeholder={t('common.pleaseInput')}
+                          readOnly
+                        ></Input>
+                      </FormControl>
+                      <FormMessage />
+                    </FormItem>
+                  )}
+                />
+              </div>
+              <Button variant={'ghost'} onClick={() => remove(index)}>
+                <X className="text-text-sub-title-invert " />
+              </Button>
+            </section>
+          );
+        })}
+      </div>
+    </section>
+  );
+}

web/src/constants/chat.ts

@@ -36,5 +36,6 @@ export const EmptyConversationId = 'empty';
 export enum DatasetMetadata {
   Disabled = 'disabled',
   Automatic = 'auto',
+  SemiAutomatic = 'semi_auto',
   Manual = 'manual',
 }

web/src/interfaces/request/knowledge.ts

@@ -7,6 +7,16 @@ export interface ITestRetrievalRequestBody {
   use_kg?: boolean;
   highlight?: boolean;
   kb_id?: string[];
+  meta_data_filter?: {
+    logic?: string;
+    method?: string;
+    manual?: Array<{
+      key: string;
+      op: string;
+      value: string;
+    }>;
+    semi_auto?: string[];
+  };
 }
 
 export interface IFetchKnowledgeListRequestBody {

web/src/locales/en.ts

@@ -737,11 +737,13 @@ This auto-tagging feature enhances retrieval by adding another layer of domain-s
       metadataTip:
         'Metadata filtering is the process of using metadata attributes (such as tags, categories, or access permissions) to refine and control the retrieval of relevant information within a system.',
       conditions: 'Conditions',
+      metadataKeys: 'Filterable items',
       addCondition: 'Add condition',
       meta: {
         disabled: 'Disabled',
         auto: 'Automatic',
         manual: 'Manual',
+        semi_auto: 'Semi-automatic',
       },
       cancel: 'Cancel',
       chatSetting: 'Chat setting',

web/src/locales/zh.ts

@@ -673,11 +673,13 @@ General：实体和关系提取提示来自 GitHub - microsoft/graphrag：基于
       metadataTip:
         '元数据过滤是使用元数据属性（例如标签、类别或访问权限）来优化和控制系统内相关信息检索的过程。',
       conditions: '条件',
+      metadataKeys: '可选过滤项',
       addCondition: '增加条件',
       meta: {
         disabled: '禁用',
         auto: '自动',
         manual: '手动',
+        semi_auto: '半自动',
       },
       cancel: '取消',
       chatSetting: '聊天设置',

web/src/pages/dataset/testing/testing-form.tsx

@@ -7,14 +7,18 @@ import { z } from 'zod';
 import { CrossLanguageFormField } from '@/components/cross-language-form-field';
 import { FormContainer } from '@/components/form-container';
 import {
-  initialTopKValue,
+  MetadataFilter,
+  MetadataFilterSchema,
+} from '@/components/metadata-filter';
+import {
   RerankFormFields,
+  initialTopKValue,
   topKSchema,
 } from '@/components/rerank';
 import {
+  SimilaritySliderFormField,
   initialSimilarityThresholdValue,
   initialVectorSimilarityWeightValue,
-  SimilaritySliderFormField,
   similarityThresholdSchema,
   vectorSimilarityWeightSchema,
 } from '@/components/similarity-slider';
@@ -33,6 +37,7 @@ import { trim } from 'lodash';
 import { Send } from 'lucide-react';
 import { useEffect } from 'react';
 import { useTranslation } from 'react-i18next';
+import { useParams } from 'umi';
 
 type TestingFormProps = Pick<
   ReturnType<typeof useTestRetrieval>,
@@ -45,6 +50,8 @@ export default function TestingForm({
   setValues,
 }: TestingFormProps) {
   const { t } = useTranslation();
+  const { id } = useParams(); // 正确解构出id参数
+  const knowledgeBaseId = id; // 现在knowledgeBaseId是字符串类型
 
   const formSchema = z.object({
     question: z.string().min(1, {
@@ -54,6 +61,8 @@ export default function TestingForm({
     ...vectorSimilarityWeightSchema,
     ...topKSchema,
     use_kg: z.boolean().optional(),
+    kb_ids: z.array(z.string()).optional(),
+    ...MetadataFilterSchema,
   });
 
   const form = useForm<z.infer<typeof formSchema>>({
@@ -63,6 +72,7 @@ export default function TestingForm({
       ...initialVectorSimilarityWeightValue,
       ...initialTopKValue,
       use_kg: false,
+      kb_ids: [knowledgeBaseId],
     },
   });
 
@@ -90,6 +100,8 @@ export default function TestingForm({
           <CrossLanguageFormField
             name={'cross_languages'}
           ></CrossLanguageFormField>
+          {/* 添加元数据过滤组件 */}
+          <MetadataFilter prefix=""></MetadataFilter>
         </FormContainer>
         <FormField
           control={form.control}

web/src/pages/next-search/search-view.tsx

@@ -202,7 +202,7 @@ export default function SearchingView({
                         <div className="w-full flex flex-col">
                           <div className="w-full highlightContent">
                             <ImageWithPopover
-                              id={chunk.img_id}
+                              id={chunk.image_id || chunk.img_id}
                             ></ImageWithPopover>
                             <Popover>
                               <PopoverTrigger asChild>