GuangYu/QuestionProject
4
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix: align agent no-hit handling

Browse Source
This commit is contained in:
Nelson
2026-07-30 14:36:09 +08:00
parent d1f573108e
commit 77a899e24c
5 changed files with 71 additions and 25 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
ai_knowledge_base_v2/apps/backend/app/services/agent_debug_service.py
View File

@@ -23,12 +23,14 @@ class AgentDebugService:
SimpleNamespace(id=index + 1, role=item.role, content=item.content)
for index, item in enumerate(payload.history)
]
preview_knowledge_ids = payload.knowledgeIds or None
version_overrides = payload.knowledgeVersions or None
rag_result = await KnowledgeAgentService.build_result(
db,
question=payload.question,
history=history,
version_overrides=payload.knowledgeVersions,
preview_knowledge_ids=payload.knowledgeIds,
version_overrides=version_overrides,
preview_knowledge_ids=preview_knowledge_ids,
prompt_override=payload.promptContent,
response_depth=payload.responseDepth,
)

2
ai_knowledge_base_v2/apps/backend/app/services/model_service.py
View File

@@ -38,7 +38,7 @@ class ModelClientService:
if model is None:
raise ExternalServiceError("未启用可用模型，请先在模型管理中启用一个模型。", provider="model")
model_name = model.model_name
answer = _call_configured_model(model, rag_result, allow_no_hit=rag_result.allow_general_knowledge)
answer = _call_configured_model(model, rag_result, allow_no_hit=True)
return ModelCompletion(
answer=answer,
model_id=model.id if model is not None else None,

26
ai_knowledge_base_v2/apps/backend/app/services/model_stream_service.py
View File

@@ -30,7 +30,7 @@ from app.services.model_service import (
_system_and_turn_messages,
_system_config_bool,
)
from app.services.rag_service import NO_HIT_ANSWER, RagResult
from app.services.rag_service import RagResult
@dataclass(frozen=True)
@@ -64,14 +64,6 @@ class ModelStreamService:
chunks=_display_chunks(model, _mock_answer(rag_result)),
)
if not rag_result.is_hit and not rag_result.allow_general_knowledge:
return StreamingModelResponse(
model_id=model.id if model is not None else None,
model_name=model.model_name if model is not None else "no-hit",
input_token=_rough_token_count(rag_result.prompt),
chunks=_display_chunks(model, NO_HIT_ANSWER),
)
if model is None:
raise ExternalServiceError("未启用可用模型，请先在模型管理中启用一个模型。", provider="model")
if not (model.api_url or model.base_url) or not model.api_key:
@@ -98,14 +90,6 @@ class ModelStreamService:
chunks=_async_display_chunks(model, _mock_answer(rag_result)),
)
if not rag_result.is_hit and not rag_result.allow_general_knowledge:
return AsyncStreamingModelResponse(
model_id=model.id if model is not None else None,
model_name=model.model_name if model is not None else "no-hit",
input_token=_rough_token_count(rag_result.prompt),
chunks=_async_display_chunks(model, NO_HIT_ANSWER),
)
if model is None:
raise ExternalServiceError("未启用可用模型，请先在模型管理中启用一个模型。", provider="model")
if not (model.api_url or model.base_url) or not model.api_key:
@@ -147,19 +131,19 @@ def _get_enabled_model(db: Session) -> ModelConfig | None:
def _stream_configured_model(model: ModelConfig, rag_result: RagResult) -> Iterator[str]:
api_type = model.api_type or "openai_compatible"
if model.stream_enabled != 1:
return iter((_call_configured_model(model, rag_result, allow_no_hit=rag_result.allow_general_knowledge),))
return iter((_call_configured_model(model, rag_result, allow_no_hit=True),))
if api_type == "anthropic_messages":
return _stream_anthropic_messages(model, rag_result)
if api_type == "openai_compatible":
return _stream_openai_compatible_model(model, rag_result)
return iter((_call_configured_model(model, rag_result, allow_no_hit=rag_result.allow_general_knowledge),))
return iter((_call_configured_model(model, rag_result, allow_no_hit=True),))
async def _stream_configured_model_async(model: ModelConfig, rag_result: RagResult) -> AsyncIterator[str]:
api_type = model.api_type or "openai_compatible"
if model.stream_enabled != 1:
answer = await asyncio.to_thread(
_call_configured_model, model, rag_result, allow_no_hit=rag_result.allow_general_knowledge
_call_configured_model, model, rag_result, allow_no_hit=True
)
yield answer
return
@@ -175,7 +159,7 @@ async def _stream_configured_model_async(model: ModelConfig, rag_result: RagResu
return
answer = await asyncio.to_thread(
_call_configured_model, model, rag_result, allow_no_hit=rag_result.allow_general_knowledge
_call_configured_model, model, rag_result, allow_no_hit=True
)
yield answer

60
ai_knowledge_base_v2/apps/backend/tests/test_agent_runtime_config.py
View File

@@ -16,6 +16,7 @@ from app.schemas.admin import AgentDebugRequest, AgentRuntimeConfigSaveRequest
from app.services.agent_debug_service import AgentDebugService
from app.services.model_stream_service import (
AsyncStreamingModelResponse,
ModelStreamService,
_openai_stream_payload,
_stream_configured_model_async,
)
@@ -194,12 +195,71 @@ def test_debug_preview_passes_conversation_history_and_replaces_saved_prompt():
kwargs = build_result.await_args.kwargs
assert [item.content for item in kwargs["history"]] == ["最开始的问题", "第一次回答"]
assert kwargs["preview_knowledge_ids"] is None
assert kwargs["version_overrides"] is None
assert kwargs["prompt_override"] == "调试主提示词"
assert kwargs["response_depth"] == 35
assert result.messages == rag_result.messages
assert result.prompt == "调试提示词渲染结果"
def test_agent_debug_empty_knowledge_selection_uses_formal_open_catalog():
payload = AgentDebugRequest(
promptContent="调试主提示词",
modelId=1,
knowledgeIds=[],
knowledgeVersions={},
question="测试问题",
)
rag_result = RagResult(
question="测试问题",
knowledge_scopes=[],
chunks=[],
prompt="测试",
allow_general_knowledge=True,
)
with _database() as db:
build_result = AsyncMock(return_value=rag_result)
with patch("app.services.agent_debug_service.KnowledgeAgentService.build_result", build_result):
asyncio.run(AgentDebugService.build_result(db, payload))
kwargs = build_result.await_args.kwargs
assert kwargs["preview_knowledge_ids"] is None
assert kwargs["version_overrides"] is None
def test_user_stream_no_hit_still_calls_model_for_cautious_answer():
async def configured_chunks(_model, _rag_result):
yield "我暂时没有可靠课程依据，但可以先帮你整理需要确认的点。"
async def collect():
with _database() as db:
model = _model()
db.add_all([
model,
SystemConfig(config_key="mock_model_enabled", config_value="false"),
])
db.commit()
rag_result = RagResult(
question="课程里的特殊练习怎么做？",
knowledge_scopes=[],
chunks=[],
prompt="本轮没有可靠的正式知识章节。",
allow_general_knowledge=False,
)
with patch(
"app.services.model_stream_service._stream_configured_model_async",
configured_chunks,
):
response = ModelStreamService.stream_async(db, rag_result)
return [chunk async for chunk in response.chunks]
chunks = asyncio.run(collect())
assert chunks == ["我暂时没有可靠课程依据，但可以先帮你整理需要确认的点。"]
def test_prompt_includes_response_depth_instruction():
with _database() as db:
db.add(SystemConfig(config_key="agent_response_depth", config_value="20"))