Enhance workflow routing and agent configuration management
- Introduce WorkflowLLMRouter for pre-response LLM routing, allowing agents to determine the appropriate function to call based on user input. - Implement UserTurnRoutingProcessor to manage user turns before reaching the LLM, ensuring proper routing and handling of user messages. - Refactor WorkflowBrain to integrate new routing logic and enhance agent stage configuration, including entry modes and resource management. - Update service factory to support dynamic LLM resource configuration based on workflow settings. - Add tests for new routing functionality and ensure proper handling of user messages in various scenarios.
This commit is contained in:
Xin Wang
@@ -9,7 +9,10 @@ from pipecat.frames.frames import (
|
||||
LLMContextFrame,
|
||||
LLMFullResponseEndFrame,
|
||||
LLMFullResponseStartFrame,
|
||||
LLMRunFrame,
|
||||
LLMTextFrame,
|
||||
OutputTransportMessageUrgentFrame,
|
||||
TTSSpeakFrame,
|
||||
)
|
||||
from pipecat.processors.aggregators.llm_context import LLMContext
|
||||
from pipecat.processors.frame_processor import FrameDirection
|
||||
@@ -388,10 +391,108 @@ class PromptBrainTests(unittest.IsolatedAsyncioTestCase):
|
||||
|
||||
|
||||
class WorkflowBrainTests(unittest.IsolatedAsyncioTestCase):
|
||||
async def test_nodes_without_outgoing_edges_remain_active(self):
|
||||
queued = []
|
||||
|
||||
async def queue_frame(frame):
|
||||
queued.append(frame)
|
||||
|
||||
runtime = BrainRuntime(
|
||||
context=LLMContext(messages=[]),
|
||||
llm=FakeLLM(),
|
||||
queue_frame=queue_frame,
|
||||
set_system_prompt=lambda _prompt: None,
|
||||
set_tools=lambda _tools: None,
|
||||
call_end=FakeCallEnd(),
|
||||
)
|
||||
|
||||
class FakeManager:
|
||||
def __init__(self, current_node=None):
|
||||
self.current_node = current_node
|
||||
|
||||
async def initialize(self, config):
|
||||
self.current_node = config["name"]
|
||||
|
||||
start_brain = WorkflowBrain(
|
||||
{
|
||||
"specVersion": 3,
|
||||
"settings": {},
|
||||
"nodes": [{"id": "start", "type": "start", "data": {}}],
|
||||
"edges": [],
|
||||
}
|
||||
)
|
||||
start_brain._runtime = runtime
|
||||
start_brain._manager = FakeManager()
|
||||
await start_brain.on_connected()
|
||||
self.assertEqual(start_brain._manager.current_node, "start")
|
||||
|
||||
agent_brain = WorkflowBrain(
|
||||
{
|
||||
"specVersion": 3,
|
||||
"settings": {"globalPrompt": "全局规则"},
|
||||
"nodes": [
|
||||
{"id": "start", "type": "start", "data": {}},
|
||||
{
|
||||
"id": "agent",
|
||||
"type": "agent",
|
||||
"data": {"prompt": "持续回答"},
|
||||
},
|
||||
],
|
||||
"edges": [
|
||||
{
|
||||
"id": "begin",
|
||||
"source": "start",
|
||||
"target": "agent",
|
||||
"data": {"mode": "always", "priority": 0},
|
||||
}
|
||||
],
|
||||
}
|
||||
)
|
||||
agent_brain._runtime = runtime
|
||||
agent_brain._manager = FakeManager("agent")
|
||||
queued.clear()
|
||||
handled = await agent_brain.on_user_turn_end("请继续回答")
|
||||
self.assertTrue(handled)
|
||||
self.assertEqual(agent_brain._manager.current_node, "agent")
|
||||
self.assertTrue(any(isinstance(frame, LLMRunFrame) for frame in queued))
|
||||
|
||||
handoff_brain = WorkflowBrain(
|
||||
{
|
||||
"specVersion": 3,
|
||||
"settings": {},
|
||||
"nodes": [
|
||||
{"id": "start", "type": "start", "data": {}},
|
||||
{
|
||||
"id": "handoff",
|
||||
"type": "handoff",
|
||||
"data": {"targetType": "human"},
|
||||
},
|
||||
],
|
||||
"edges": [],
|
||||
}
|
||||
)
|
||||
handoff_brain._runtime = runtime
|
||||
handoff_config = await handoff_brain._resolve_path("handoff")
|
||||
self.assertEqual(handoff_config["name"], "handoff")
|
||||
self.assertTrue(
|
||||
any(
|
||||
isinstance(frame, OutputTransportMessageUrgentFrame)
|
||||
and frame.message.get("type") == "handoff-requested"
|
||||
for frame in queued
|
||||
)
|
||||
)
|
||||
|
||||
async def test_transition_and_end_are_owned_by_workflow_brain(self):
|
||||
graph = {
|
||||
"specVersion": 3,
|
||||
"settings": {"globalPrompt": "全局规则"},
|
||||
"settings": {
|
||||
"globalPrompt": "全局规则",
|
||||
"defaultLlmResourceId": "llm_global",
|
||||
"defaultAsrResourceId": "asr_global",
|
||||
"defaultTtsResourceId": "tts_global",
|
||||
"knowledgeBaseId": "kb_global",
|
||||
"knowledgeMode": "automatic",
|
||||
},
|
||||
"nodes": [
|
||||
{
|
||||
"id": "start",
|
||||
@@ -428,6 +529,7 @@ class WorkflowBrainTests(unittest.IsolatedAsyncioTestCase):
|
||||
"mode": "llm",
|
||||
"priority": 10,
|
||||
"condition": "需求已收集",
|
||||
"transitionSpeech": "正在为你结束流程",
|
||||
},
|
||||
}
|
||||
],
|
||||
@@ -447,6 +549,8 @@ class WorkflowBrainTests(unittest.IsolatedAsyncioTestCase):
|
||||
llm = FakeLLM()
|
||||
context = LLMContext(messages=[])
|
||||
queued = []
|
||||
service_switches = []
|
||||
knowledge_scopes = []
|
||||
call_end = FakeCallEnd()
|
||||
|
||||
class FakeWorker:
|
||||
@@ -478,6 +582,9 @@ class WorkflowBrainTests(unittest.IsolatedAsyncioTestCase):
|
||||
async def queue_frame(frame):
|
||||
queued.append(frame)
|
||||
|
||||
async def switch_services(llm_id, asr_id, tts_id):
|
||||
service_switches.append((llm_id, asr_id, tts_id))
|
||||
|
||||
runtime = BrainRuntime(
|
||||
context=context,
|
||||
llm=llm,
|
||||
@@ -487,29 +594,112 @@ class WorkflowBrainTests(unittest.IsolatedAsyncioTestCase):
|
||||
call_end=call_end,
|
||||
worker=worker,
|
||||
context_aggregator=pair,
|
||||
switch_services=switch_services,
|
||||
set_knowledge_scope=knowledge_scopes.append,
|
||||
)
|
||||
await brain.setup(cfg, runtime)
|
||||
await brain.on_connected()
|
||||
self.assertEqual(brain._manager.current_node, "agent")
|
||||
self.assertEqual(
|
||||
service_switches,
|
||||
[("llm_global", "asr_global", "tts_global")],
|
||||
)
|
||||
self.assertEqual(knowledge_scopes[-1]["knowledge_base_id"], "kb_global")
|
||||
|
||||
brain._engine.data("agent").update(
|
||||
{
|
||||
"inheritGlobalConfig": False,
|
||||
"llmResourceId": "llm_agent",
|
||||
"asrResourceId": "asr_agent",
|
||||
"ttsResourceId": "tts_agent",
|
||||
"knowledgeBaseId": "kb_agent",
|
||||
"knowledgeMode": "on_demand",
|
||||
}
|
||||
)
|
||||
await brain._apply_agent_stage("agent")
|
||||
self.assertEqual(
|
||||
service_switches[-1],
|
||||
("llm_agent", "asr_agent", "tts_agent"),
|
||||
)
|
||||
self.assertEqual(knowledge_scopes[-1]["knowledge_base_id"], "kb_agent")
|
||||
agent_config = brain._agent_config("agent")
|
||||
self.assertIn("王先生", agent_config["role_message"])
|
||||
self.assertIn("完成当前阶段任务", agent_config["role_message"])
|
||||
self.assertIn("工作流路由已在用户一轮输入结束时完成", agent_config["role_message"])
|
||||
self.assertEqual(agent_config["task_messages"], [])
|
||||
self.assertFalse(agent_config["respond_immediately"])
|
||||
self.assertFalse(any(isinstance(frame, LLMRunFrame) for frame in worker.frames))
|
||||
self.assertEqual(
|
||||
agent_config["context_strategy"].strategy.value,
|
||||
"reset",
|
||||
)
|
||||
|
||||
edge_function = next(
|
||||
function
|
||||
for function in brain._agent_config("agent")["functions"]
|
||||
if function.name == "goto_finish"
|
||||
brain._engine.data("agent")["entryMode"] = "generate"
|
||||
generate_config = brain._agent_config("agent")
|
||||
self.assertTrue(generate_config["respond_immediately"])
|
||||
worker.frames.clear()
|
||||
await brain._manager.set_node_from_config(generate_config)
|
||||
self.assertTrue(any(isinstance(frame, LLMRunFrame) for frame in worker.frames))
|
||||
|
||||
brain._engine.data("agent").update(
|
||||
{"entryMode": "fixed_speech", "entrySpeech": "您好，{{user_name}}"}
|
||||
)
|
||||
_, terminal = await edge_function.handler({}, brain._manager)
|
||||
self.assertEqual(terminal["name"], "end")
|
||||
fixed_config = brain._agent_config("agent")
|
||||
self.assertFalse(fixed_config["respond_immediately"])
|
||||
self.assertEqual(
|
||||
fixed_config["pre_actions"][0]["type"],
|
||||
"workflow_fixed_speech",
|
||||
)
|
||||
self.assertEqual(fixed_config["pre_actions"][0]["text"], "您好，王先生")
|
||||
self.assertEqual(
|
||||
fixed_config["task_messages"],
|
||||
[{"role": "assistant", "content": "您好，王先生"}],
|
||||
)
|
||||
worker.frames.clear()
|
||||
queued.clear()
|
||||
await brain._manager.set_node_from_config(fixed_config)
|
||||
self.assertTrue(any(isinstance(frame, TTSSpeakFrame) for frame in queued))
|
||||
self.assertFalse(any(isinstance(frame, LLMRunFrame) for frame in worker.frames))
|
||||
|
||||
self.assertFalse(
|
||||
any(
|
||||
function.name == "goto_finish"
|
||||
for function in brain._agent_config("agent")["functions"]
|
||||
)
|
||||
)
|
||||
await brain.on_assistant_text_end("old-turn", "需求已收集", False)
|
||||
self.assertEqual(brain._manager.current_node, "agent")
|
||||
|
||||
class FakeRouter:
|
||||
async def select_edge(self, **_kwargs):
|
||||
return "goto_finish"
|
||||
|
||||
brain._router = FakeRouter()
|
||||
handled = await brain.on_user_turn_end("我的需求已经说完了")
|
||||
self.assertTrue(handled)
|
||||
self.assertEqual(brain._manager.current_node, "end")
|
||||
self.assertIn("我的需求已经说完了", brain._store.values["system__conversation_history"])
|
||||
self.assertTrue(call_end.ending)
|
||||
self.assertTrue(call_end.armed)
|
||||
self.assertTrue(any(getattr(frame, "text", "") == "感谢来电" for frame in queued))
|
||||
assistant_transcripts = [
|
||||
frame.message.get("content")
|
||||
for frame in queued
|
||||
if isinstance(frame, OutputTransportMessageUrgentFrame)
|
||||
and frame.message.get("type") == "transcript"
|
||||
and frame.message.get("role") == "assistant"
|
||||
]
|
||||
self.assertEqual(
|
||||
assistant_transcripts,
|
||||
["您好，王先生", "正在为你结束流程", "感谢来电"],
|
||||
)
|
||||
self.assertIn(
|
||||
"正在为你结束流程",
|
||||
brain._store.values["system__conversation_history"],
|
||||
)
|
||||
self.assertIn(
|
||||
"感谢来电",
|
||||
brain._store.values["system__conversation_history"],
|
||||
)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
|
||||
Reference in New Issue
Block a user