feat(knowledge): add configurable ASR model in knowledge system settings

Expose optional speech-to-text model on the knowledge-base model tab and
resolve media transcription from knowledge LLM config instead of workbench ASR.

Co-authored-by: Cursor <cursoragent@cursor.com>

src/backend/bisheng/knowledge/domain/services/media_transcription_service.py

@@ -91,11 +91,11 @@ class KnowledgeMediaTranscriptionService:
 
     @classmethod
     def _resolve_asr_model(cls, tenant_id: int | None) -> tuple[LLMModel, LLMServer]:
-        workbench_llm = LLMService.get_workbench_llm_sync(tenant_id=tenant_id)
-        if not workbench_llm.asr_model or not workbench_llm.asr_model.id:
+        knowledge_llm = LLMService.get_knowledge_llm(tenant_id=tenant_id)
+        if not knowledge_llm.asr_model_id:
             raise NoAsrModelConfigError()
 
-        model_info = LLMDao.get_model_by_id(int(workbench_llm.asr_model.id))
+        model_info = LLMDao.get_model_by_id(int(knowledge_llm.asr_model_id))
         if not model_info:
             raise AsrModelConfigDeletedError()
         if model_info.model_type != LLMModelType.ASR.value:

src/backend/bisheng/llm/domain/schemas.py

@@ -131,6 +131,7 @@ class KnowledgeLLMConfig(BaseModel):
         None, description="Documentation Knowledge Base Extraction Header Model'sID"
     )
     qa_similar_model_id: int | None = Field(None, description="QAThe Knowledge Base Similarity Question Model'sID")
+    asr_model_id: int | None = Field(None, description="Knowledge base media transcription ASR model ID")
     abstract_enabled: bool = Field(default=True, description="Whether to generate file summaries after parsing")
     auto_tag_enabled: bool = Field(default=True, description="Whether to generate file tags after upload parsing")
     abstract_prompt: str | None = Field(None, description="Summary Prompt")

src/backend/bisheng/llm/domain/services/llm.py

@@ -910,7 +910,13 @@ class LLMService:
         """Update default model configuration for knowledge base"""
         target = _resolve_tenant_id(tenant_id)
         await avalidate_system_model_refs(
-            [data.embedding_model_id, data.source_model_id, data.extract_title_model_id, data.qa_similar_model_id],
+            [
+                data.embedding_model_id,
+                data.source_model_id,
+                data.extract_title_model_id,
+                data.qa_similar_model_id,
+                data.asr_model_id,
+            ],
             target,
         )
         await cls._base_update_llm_config(

src/backend/test/knowledge/test_media_transcription_service.py

@@ -1,12 +1,17 @@
 import subprocess
+from types import SimpleNamespace
+from unittest.mock import patch
 
 import pytest
 
 from bisheng.common.errcode.knowledge import KnowledgeMediaNoRecognizableAudioError
+from bisheng.common.errcode.server import NoAsrModelConfigError
 from bisheng.knowledge.domain.services.media_transcription_service import (
     KnowledgeMediaTranscriptionService,
     TranscriptSegment,
 )
+from bisheng.llm.domain.const import LLMModelType, LLMServerType
+from bisheng.llm.domain.schemas import KnowledgeLLMConfig
 
 
 def test_normalize_segments_keeps_aliyun_millisecond_timestamps_consistent() -> None:
@@ -113,6 +118,45 @@ def test_convert_to_wav_reports_missing_audio_stream(monkeypatch, tmp_path) -> N
         KnowledgeMediaTranscriptionService._convert_to_wav(str(media_path))
 
 
+def test_resolve_asr_model_reads_knowledge_config(monkeypatch) -> None:
+    model_info = SimpleNamespace(
+        id=42,
+        model_name="paraformer-realtime-v2",
+        model_type=LLMModelType.ASR.value,
+        server_id=7,
+        online=True,
+        config={},
+    )
+    server_info = SimpleNamespace(name="Aliyun", type=LLMServerType.QWEN.value, config={"api_key": "sk-test"})
+
+    monkeypatch.setattr(
+        "bisheng.knowledge.domain.services.media_transcription_service.LLMService.get_knowledge_llm",
+        lambda tenant_id=None: KnowledgeLLMConfig(asr_model_id=42),
+    )
+    monkeypatch.setattr(
+        "bisheng.knowledge.domain.services.media_transcription_service.LLMDao.get_model_by_id",
+        lambda model_id: model_info if model_id == 42 else None,
+    )
+    monkeypatch.setattr(
+        "bisheng.knowledge.domain.services.media_transcription_service.LLMDao.get_server_by_id",
+        lambda server_id: server_info if server_id == 7 else None,
+    )
+
+    resolved_model, resolved_server = KnowledgeMediaTranscriptionService._resolve_asr_model(tenant_id=1)
+
+    assert resolved_model is model_info
+    assert resolved_server is server_info
+
+
+def test_resolve_asr_model_requires_knowledge_config() -> None:
+    with patch(
+        "bisheng.knowledge.domain.services.media_transcription_service.LLMService.get_knowledge_llm",
+        return_value=KnowledgeLLMConfig(asr_model_id=None),
+    ):
+        with pytest.raises(NoAsrModelConfigError):
+            KnowledgeMediaTranscriptionService._resolve_asr_model(tenant_id=1)
+
+
 def test_empty_asr_text_reports_missing_recognizable_audio(monkeypatch, tmp_path) -> None:
     media_path = tmp_path / "silent.mp4"
     wav_path = tmp_path / "silent.wav"

src/backend/test/llm/test_llm_share_fallback.py

@@ -367,7 +367,7 @@ async def test_update_knowledge_llm_calls_validator_with_all_model_ids():
 
     payload = KnowledgeLLMConfig(
         embedding_model_id=4, source_model_id=7,
-        extract_title_model_id=7, qa_similar_model_id=7,
+        extract_title_model_id=7, qa_similar_model_id=7, asr_model_id=9,
     )
     with patch(
         'bisheng.llm.domain.services.llm.avalidate_system_model_refs',
@@ -380,7 +380,7 @@ async def test_update_knowledge_llm_calls_validator_with_all_model_ids():
 
     args, kwargs = mock_validate.call_args
     passed_ids = list(args[0])
-    assert set(passed_ids) == {4, 7}
+    assert set(passed_ids) == {4, 7, 9}
     assert (kwargs.get('target_tenant_id') if 'target_tenant_id' in kwargs else args[1]) == 1
 
 

src/frontend/platform/public/locales/en-US/model.json

@@ -17,6 +17,7 @@
     "sessionTitleGenerationModel": "Session Title Generation Model",
     "asrModel": "Speech-to-Text (ASR) Model",
     "asrModelTooltip": "Used for workbench speech-to-text scenarios",
+    "knowledgeAsrModelTooltip": "Used for speech-to-text in audio/video parsing",
     "ttsModel": "Text-to-Speech (TTS) Model",
     "ttsModelTooltip": "Used for workbench text-to-speech scenarios",
     "saveFailed": "Save failed, please try again!",

src/frontend/platform/public/locales/ja/model.json

@@ -15,6 +15,7 @@
     "sessionTitleGenerationModel": "セッションタイトル生成モデル",
     "asrModel": "音声認識（ASR）モデル",
     "asrModelTooltip": "ワークベンチ音声認識シナリオに使用",
+    "knowledgeAsrModelTooltip": "音声・動画解析の音声認識シナリオに使用",
     "ttsModel": "テキスト読み上げ（TTS）モデル",
     "ttsModelTooltip": "ワークベンチテキスト読み上げシナリオに使用",
     "saveFailed": "保存に失敗しました。再試行してください！",

src/frontend/platform/public/locales/zh-Hans/model.json

@@ -15,6 +15,7 @@
     "sessionTitleGenerationModel": "应用会话标题生成模型",
     "asrModel": "语音转文字（ASR）模型",
     "asrModelTooltip": "用于工作台语音转文字场景",
+    "knowledgeAsrModelTooltip": "用于音视频解析中语音转文字场景",
     "ttsModel": "文字转语音（TTS）模型",
     "ttsModelTooltip": "用于工作台文字转语音场景",
     "saveFailed": "保存失败，请重试！",

src/frontend/platform/src/pages/ModelPage/manage/SystemModelConfig.tsx

@@ -147,7 +147,7 @@ export default function SystemModelConfig({
                         <WorkbenchModel llmOptions={llmOptions} embeddings={embeddings} asrModel={asrModel} ttsModel={ttsModel} onBack={onBack}></WorkbenchModel>
                     </TabsContent>
                     <TabsContent value="knowledge">
-                        <KnowledgeModle llmOptions={llmOptions} embeddings={embeddings} onBack={onBack}></KnowledgeModle>
+                        <KnowledgeModle llmOptions={llmOptions} embeddings={embeddings} asrModel={asrModel} onBack={onBack}></KnowledgeModle>
                     </TabsContent>
                     <TabsContent value="assis">
                         <AssisModel llmOptions={llmOptions} onBack={onBack}></AssisModel>

src/frontend/platform/src/pages/ModelPage/manage/tabs/KnowledgeModel.tsx

@@ -119,7 +119,7 @@ const PromptDialog = ({ value, onChange, onRestore, onSave, label, children }) =
     </Dialog>
 }
 
-export default function KnowledgeModel({ llmOptions, embeddings, onBack }) {
+export default function KnowledgeModel({ llmOptions, embeddings, asrModel = [], onBack }) {
     const { t } = useTranslation('model')
 
     const [form, setForm] = useState({
@@ -127,6 +127,7 @@ export default function KnowledgeModel({ llmOptions, embeddings, onBack }) {
         sourceModelId: null,
         extractModelId: null,
         qaSimilarModelId: null,
+        asrModelId: null,
         abstractEnabled: true,
         autoTagEnabled: true,
         abstractPrompt: '',
@@ -144,6 +145,7 @@ export default function KnowledgeModel({ llmOptions, embeddings, onBack }) {
             embedding_model_id,
             extract_title_model_id,
             qa_similar_model_id,
+            asr_model_id,
             source_model_id,
             abstract_enabled,
             auto_tag_enabled,
@@ -155,6 +157,7 @@ export default function KnowledgeModel({ llmOptions, embeddings, onBack }) {
             sourceModelId: source_model_id,
             extractModelId: extract_title_model_id,
             qaSimilarModelId: qa_similar_model_id,
+            asrModelId: asr_model_id ?? null,
             abstractEnabled: abstract_enabled ?? true,
             autoTagEnabled: auto_tag_enabled ?? true,
             abstractPrompt: abstract_prompt ?? defalutPrompt,
@@ -182,6 +185,7 @@ export default function KnowledgeModel({ llmOptions, embeddings, onBack }) {
             embeddingModelId,
             extractModelId,
             qaSimilarModelId,
+            asrModelId,
             sourceModelId,
             abstractEnabled,
             autoTagEnabled,
@@ -204,6 +208,7 @@ export default function KnowledgeModel({ llmOptions, embeddings, onBack }) {
             embedding_model_id: embeddingModelId,
             extract_title_model_id: extractModelId,
             qa_similar_model_id: qaSimilarModelId,
+            asr_model_id: asrModelId || null,
             source_model_id: sourceModelId,
             abstract_enabled: abstractEnabled,
             auto_tag_enabled: autoTagEnabled,
@@ -322,6 +327,14 @@ export default function KnowledgeModel({ llmOptions, embeddings, onBack }) {
                 options={llmOptions}
                 onChange={(val) => setFormAndClearInherited({ ...form, qaSimilarModelId: val })}
             />
+            <ModelSelect
+                close
+                label={t('model.asrModel')}
+                tooltipText={t('model.knowledgeAsrModelTooltip')}
+                value={form.asrModelId}
+                options={asrModel}
+                onChange={(val) => setFormAndClearInherited({ ...form, asrModelId: val })}
+            />
             <div className="mt-10 text-center space-x-6">
                 <Button className="px-6" variant="outline" onClick={onBack}>{t('model.cancel')}</Button>
                 <Button