feat(roundtable): 第二/三步模型自动容错(出问题自动换下一个模型)
模型出问题时按 config.yaml models[] 顺序自动切到下一个模型继续,并告知用户 「『X』模型异常 → 已切换『Y』继续」。前台实时 + 后台挂起两条路径都覆盖。 - 中间件兜底/熔断 AIMessage 打 LLM_ERROR_MARKER 标记,is_llm_error_message 可识别 - roundtable_model_fallback: fallback_model_chain 候选链 + reason_text + 总开关 env ROUNDTABLE_MODEL_FALLBACK - 触发: 异常(start_run失败/idle卡死) / 兜底错误文案 / 交付为空(仅seat/leader) - 后台 InProcessRoundtableGateway._run_turn_with_fallback 包住每轮 run - 前台 _special_run/_leader_run 收尾检测失败→model_switch 帧+换模型重流 - 结构化抽取 flow_extract round1 astream 同款换模型循环 - Step2 单席位异常隔离: 单席位失败不拖垮整批/整链/整个作业 - 时间线插「系统提示」气泡 + toast + 诊断 model_switch 事件 - Tests: test_roundtable_model_fallback.py Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
1 parent
ae2706dad3
commit
d383b0e9dd
18 files changed
+1033
-174
No files matched your search
@@ -31,12 +31,12 @@ from fastapi.responses import StreamingResponse
|
||||
from langchain_core.messages import HumanMessage, SystemMessage
|
||||
from pydantic import BaseModel, Field
|
||||
|
||||
from deerflow.agents.roundtable_orchestrator.extraction_spec import build_structure_directive
|
||||
|
||||
from app.gateway.roundtable_model_fallback import fallback_model_chain, reason_text
|
||||
from app.gateway.routers._roundtable_seed import (
|
||||
STRUCTURE_AGENT_ID,
|
||||
ensure_roundtable_functional_agents,
|
||||
)
|
||||
from deerflow.agents.roundtable_orchestrator.extraction_spec import build_structure_directive
|
||||
from deerflow.config import get_app_config
|
||||
from deerflow.config.agents_config import load_agent_config, load_agent_soul
|
||||
from deerflow.models.factory import create_chat_model
|
||||
@@ -596,28 +596,60 @@ async def extract_flow_stream(request: Request, payload: FlowExtractRequest = Bo
|
||||
return
|
||||
|
||||
# ① 抽取 / 编辑(正文 + 思考(reasoning) 拼成 `<think>…</think>正文` 增量逐帧吐,前端实时显示)
|
||||
# 模型自动容错:结构化抽取非常重要——某模型 astream 抛错 / 产出空时,按 config.yaml
|
||||
# models[] 顺序切到下一个模型重抽,并发 model_switch 帧让前端把抽取面板已累计的半截内容重置。
|
||||
round1_messages = _edit_messages(payload) if is_edit else _extract_messages(payload)
|
||||
model = create_chat_model(name=resolved_model, thinking_enabled=True)
|
||||
chain = fallback_model_chain(resolved_model) or [resolved_model]
|
||||
extracted = ""
|
||||
think_open = False
|
||||
async for chunk in model.astream(round1_messages):
|
||||
c, r = _chunk_parts(chunk)
|
||||
out = ""
|
||||
if r:
|
||||
if not think_open:
|
||||
out += "<think>"
|
||||
think_open = True
|
||||
out += r
|
||||
if c:
|
||||
used_model = chain[0]
|
||||
prev_model: str | None = None
|
||||
prev_reason: str | None = None
|
||||
for idx, candidate in enumerate(chain):
|
||||
if idx > 0:
|
||||
yield _frame(
|
||||
{
|
||||
"event": "model_switch",
|
||||
"agent_name": "结构化抽取",
|
||||
"failed_model": prev_model or "",
|
||||
"next_model": candidate or "",
|
||||
"reason": reason_text(prev_reason),
|
||||
}
|
||||
)
|
||||
used_model = candidate
|
||||
extracted = ""
|
||||
think_open = False
|
||||
try:
|
||||
model = create_chat_model(name=candidate, thinking_enabled=True)
|
||||
async for chunk in model.astream(round1_messages):
|
||||
c, r = _chunk_parts(chunk)
|
||||
out = ""
|
||||
if r:
|
||||
if not think_open:
|
||||
out += "<think>"
|
||||
think_open = True
|
||||
out += r
|
||||
if c:
|
||||
if think_open:
|
||||
out += "</think>"
|
||||
think_open = False
|
||||
out += c
|
||||
extracted += c
|
||||
if out:
|
||||
yield _frame({"event": "extract_delta", "content": out})
|
||||
if think_open:
|
||||
out += "</think>"
|
||||
think_open = False
|
||||
out += c
|
||||
extracted += c
|
||||
if out:
|
||||
yield _frame({"event": "extract_delta", "content": out})
|
||||
if think_open:
|
||||
yield _frame({"event": "extract_delta", "content": "</think>"})
|
||||
yield _frame({"event": "extract_delta", "content": "</think>"})
|
||||
except Exception as exc: # noqa: BLE001 — 该模型抽取失败:换下一个模型重抽
|
||||
prev_model, prev_reason = candidate, "exception"
|
||||
logger.warning("flow-extract round1 model %s failed: %s", candidate, exc)
|
||||
if idx + 1 < len(chain):
|
||||
continue
|
||||
raise
|
||||
if extracted.strip():
|
||||
break
|
||||
# 产出为空:换下一个模型(已是最后一个则继续走后续解析兜底)。
|
||||
prev_model, prev_reason = candidate, "empty"
|
||||
if idx + 1 >= len(chain):
|
||||
break
|
||||
yield _frame({"event": "extract_done"})
|
||||
|
||||
# ② 校验 / 修复(同样流式 + 思考实时显示)
|
||||
@@ -625,7 +657,7 @@ async def extract_flow_stream(request: Request, payload: FlowExtractRequest = Bo
|
||||
validated = ""
|
||||
graph: dict[str, Any] | None = None
|
||||
try:
|
||||
vmodel = create_chat_model(name=resolved_model, thinking_enabled=True)
|
||||
vmodel = create_chat_model(name=used_model, thinking_enabled=True)
|
||||
think_open = False
|
||||
async for chunk in vmodel.astream(_validate_messages(payload, extracted)):
|
||||
c, r = _chunk_parts(chunk)
|
||||
|
||||
Reference in new issue
Block a user