- Revisiting RAG_SPECIALIST
- Adapt Catalogs & Retrievers to use specific types, removing tagging_fields - Adding CrewAI Implementation Guide
This commit is contained in:
Josako
@@ -6,6 +6,7 @@ from crewai.flow.flow import start, listen, and_
|
||||
from flask import current_app
|
||||
from pydantic import BaseModel, Field
|
||||
|
||||
from common.services.utils.translation_services import TranslationServices
|
||||
from common.utils.business_event_context import current_event
|
||||
from eveai_chat_workers.retrievers.retriever_typing import RetrieverArguments
|
||||
from eveai_chat_workers.specialists.crewai_base_specialist import CrewAIBaseSpecialistExecutor
|
||||
@@ -13,6 +14,10 @@ from eveai_chat_workers.specialists.specialist_typing import SpecialistResult, S
|
||||
from eveai_chat_workers.outputs.globals.rag.rag_v1_0 import RAGOutput
|
||||
from eveai_chat_workers.specialists.crewai_base_classes import EveAICrewAICrew, EveAICrewAIFlow, EveAIFlowState
|
||||
|
||||
INSUFFICIENT_INFORMATION_MESSAGE = (
|
||||
"We do not have the necessary information to provide you with the requested answers. "
|
||||
"Please accept our apologies. Don't hesitate to ask other questions, and I'll do my best to answer them.")
|
||||
|
||||
|
||||
class SpecialistExecutor(CrewAIBaseSpecialistExecutor):
|
||||
"""
|
||||
@@ -40,6 +45,9 @@ class SpecialistExecutor(CrewAIBaseSpecialistExecutor):
|
||||
def _config_pydantic_outputs(self):
|
||||
self._add_pydantic_output("rag_task", RAGOutput, "rag_output")
|
||||
|
||||
def _config_state_result_relations(self):
|
||||
self._add_state_result_relation("rag_output")
|
||||
|
||||
def _instantiate_specialist(self):
|
||||
verbose = self.tuning
|
||||
|
||||
@@ -61,40 +69,84 @@ class SpecialistExecutor(CrewAIBaseSpecialistExecutor):
|
||||
def execute(self, arguments: SpecialistArguments, formatted_context, citations) -> SpecialistResult:
|
||||
self.log_tuning("RAG Specialist execution started", {})
|
||||
|
||||
flow_inputs = {
|
||||
"language": arguments.language,
|
||||
"query": arguments.query,
|
||||
"context": formatted_context,
|
||||
"history": self.formatted_history,
|
||||
"name": self.specialist.configuration.get('name', ''),
|
||||
"company": self.specialist.configuration.get('company', ''),
|
||||
}
|
||||
# crew_results = self.rag_crew.kickoff(inputs=flow_inputs)
|
||||
# current_app.logger.debug(f"Test Crew Output received: {crew_results}")
|
||||
flow_results = self.flow.kickoff(inputs=flow_inputs)
|
||||
current_app.logger.debug(f"Arguments: {arguments.model_dump()}")
|
||||
current_app.logger.debug(f"Formatted Context: {formatted_context}")
|
||||
current_app.logger.debug(f"Formatted History: {self._formatted_history}")
|
||||
current_app.logger.debug(f"Cached Chat Session: {self._cached_session}")
|
||||
|
||||
flow_state = self.flow.state
|
||||
if not self._cached_session.interactions:
|
||||
specialist_phase = "initial"
|
||||
else:
|
||||
specialist_phase = self._cached_session.interactions[-1].specialist_results.get('phase', 'initial')
|
||||
|
||||
results = RAGSpecialistResult.create_for_type(self.type, self.type_version)
|
||||
update_data = {}
|
||||
if flow_state.rag_output: # Fallback
|
||||
update_data["rag_output"] = flow_state.rag_output
|
||||
results = None
|
||||
current_app.logger.debug(f"Specialist Phase: {specialist_phase}")
|
||||
|
||||
results = results.model_copy(update=update_data)
|
||||
match specialist_phase:
|
||||
case "initial":
|
||||
results = self.execute_initial_state(arguments, formatted_context, citations)
|
||||
case "rag":
|
||||
results = self.execute_rag_state(arguments, formatted_context, citations)
|
||||
|
||||
self.log_tuning(f"RAG Specialist execution ended", {"Results": results.model_dump()})
|
||||
|
||||
return results
|
||||
|
||||
def execute_initial_state(self, arguments: SpecialistArguments, formatted_context, citations) -> SpecialistResult:
|
||||
self.log_tuning("RAG Specialist initial_state execution started", {})
|
||||
|
||||
welcome_message = self.specialist.configuration.get('welcome_message', 'Welcome! You can start asking questions')
|
||||
welcome_message = TranslationServices.translate(self.tenant_id, welcome_message, arguments.language)
|
||||
|
||||
self.flow.state.answer = welcome_message
|
||||
self.flow.state.phase = "rag"
|
||||
|
||||
results = RAGSpecialistResult.create_for_type(self.type, self.type_version)
|
||||
|
||||
return results
|
||||
|
||||
def execute_rag_state(self, arguments: SpecialistArguments, formatted_context, citations) -> SpecialistResult:
|
||||
self.log_tuning("RAG Specialist rag_state execution started", {})
|
||||
|
||||
insufficient_info_message = TranslationServices.translate(self.tenant_id,
|
||||
INSUFFICIENT_INFORMATION_MESSAGE,
|
||||
arguments.language)
|
||||
|
||||
if formatted_context:
|
||||
flow_inputs = {
|
||||
"language": arguments.language,
|
||||
"question": arguments.question,
|
||||
"context": formatted_context,
|
||||
"history": self.formatted_history,
|
||||
"name": self.specialist.configuration.get('name', ''),
|
||||
"welcome_message": self.specialist.configuration.get('welcome_message', '')
|
||||
}
|
||||
|
||||
flow_results = self.flow.kickoff(inputs=flow_inputs)
|
||||
|
||||
if flow_results.rag_output.insufficient_info:
|
||||
flow_results.rag_output.answer = insufficient_info_message
|
||||
|
||||
rag_output = flow_results.rag_output
|
||||
|
||||
else:
|
||||
rag_output = RAGOutput(answer=insufficient_info_message, insufficient_info=True)
|
||||
|
||||
self.flow.state.rag_output = rag_output
|
||||
self.flow.state.answer = rag_output.answer
|
||||
self.flow.state.phase = "rag"
|
||||
|
||||
results = RAGSpecialistResult.create_for_type(self.type, self.type_version)
|
||||
|
||||
|
||||
class RAGSpecialistInput(BaseModel):
|
||||
language: Optional[str] = Field(None, alias="language")
|
||||
query: Optional[str] = Field(None, alias="query")
|
||||
question: Optional[str] = Field(None, alias="question")
|
||||
context: Optional[str] = Field(None, alias="context")
|
||||
citations: Optional[List[int]] = Field(None, alias="citations")
|
||||
history: Optional[str] = Field(None, alias="history")
|
||||
name: Optional[str] = Field(None, alias="name")
|
||||
company: Optional[str] = Field(None, alias="company")
|
||||
welcome_message: Optional[str] = Field(None, alias="welcome_message")
|
||||
|
||||
|
||||
class RAGSpecialistResult(SpecialistResult):
|
||||
|
||||
197
eveai_chat_workers/specialists/globals/RAG_SPECIALIST/1_1.py
Normal file
197
eveai_chat_workers/specialists/globals/RAG_SPECIALIST/1_1.py
Normal file
@@ -0,0 +1,197 @@
|
||||
import json
|
||||
from os import wait
|
||||
from typing import Optional, List
|
||||
|
||||
from crewai.flow.flow import start, listen, and_
|
||||
from flask import current_app
|
||||
from pydantic import BaseModel, Field
|
||||
|
||||
from common.services.utils.translation_services import TranslationServices
|
||||
from common.utils.business_event_context import current_event
|
||||
from eveai_chat_workers.retrievers.retriever_typing import RetrieverArguments
|
||||
from eveai_chat_workers.specialists.crewai_base_specialist import CrewAIBaseSpecialistExecutor
|
||||
from eveai_chat_workers.specialists.specialist_typing import SpecialistResult, SpecialistArguments
|
||||
from eveai_chat_workers.outputs.globals.rag.rag_v1_0 import RAGOutput
|
||||
from eveai_chat_workers.specialists.crewai_base_classes import EveAICrewAICrew, EveAICrewAIFlow, EveAIFlowState
|
||||
|
||||
INSUFFICIENT_INFORMATION_MESSAGE = (
|
||||
"We do not have the necessary information to provide you with the requested answers. "
|
||||
"Please accept our apologies. Don't hesitate to ask other questions, and I'll do my best to answer them.")
|
||||
|
||||
|
||||
class SpecialistExecutor(CrewAIBaseSpecialistExecutor):
|
||||
"""
|
||||
type: RAG_SPECIALIST
|
||||
type_version: 1.0
|
||||
RAG Specialist Executor class
|
||||
"""
|
||||
|
||||
def __init__(self, tenant_id, specialist_id, session_id, task_id, **kwargs):
|
||||
self.rag_crew = None
|
||||
|
||||
super().__init__(tenant_id, specialist_id, session_id, task_id)
|
||||
|
||||
@property
|
||||
def type(self) -> str:
|
||||
return "RAG_SPECIALIST"
|
||||
|
||||
@property
|
||||
def type_version(self) -> str:
|
||||
return "1.1"
|
||||
|
||||
def _config_task_agents(self):
|
||||
self._add_task_agent("rag_task", "rag_agent")
|
||||
|
||||
def _config_pydantic_outputs(self):
|
||||
self._add_pydantic_output("rag_task", RAGOutput, "rag_output")
|
||||
|
||||
def _config_state_result_relations(self):
|
||||
self._add_state_result_relation("rag_output")
|
||||
|
||||
def _instantiate_specialist(self):
|
||||
verbose = self.tuning
|
||||
|
||||
rag_agents = [self.rag_agent]
|
||||
rag_tasks = [self.rag_task]
|
||||
self.rag_crew = EveAICrewAICrew(
|
||||
self,
|
||||
"Rag Crew",
|
||||
agents=rag_agents,
|
||||
tasks=rag_tasks,
|
||||
verbose=verbose,
|
||||
)
|
||||
|
||||
self.flow = RAGFlow(
|
||||
self,
|
||||
self.rag_crew,
|
||||
)
|
||||
|
||||
def execute(self, arguments: SpecialistArguments, formatted_context, citations) -> SpecialistResult:
|
||||
self.log_tuning("RAG Specialist execution started", {})
|
||||
|
||||
current_app.logger.debug(f"Arguments: {arguments.model_dump()}")
|
||||
current_app.logger.debug(f"Formatted Context: {formatted_context}")
|
||||
current_app.logger.debug(f"Formatted History: {self._formatted_history}")
|
||||
current_app.logger.debug(f"Cached Chat Session: {self._cached_session}")
|
||||
|
||||
if not self._cached_session.interactions:
|
||||
specialist_phase = "initial"
|
||||
else:
|
||||
specialist_phase = self._cached_session.interactions[-1].specialist_results.get('phase', 'initial')
|
||||
|
||||
results = None
|
||||
current_app.logger.debug(f"Specialist Phase: {specialist_phase}")
|
||||
|
||||
match specialist_phase:
|
||||
case "initial":
|
||||
results = self.execute_initial_state(arguments, formatted_context, citations)
|
||||
case "rag":
|
||||
results = self.execute_rag_state(arguments, formatted_context, citations)
|
||||
|
||||
self.log_tuning(f"RAG Specialist execution ended", {"Results": results.model_dump()})
|
||||
|
||||
return results
|
||||
|
||||
def execute_initial_state(self, arguments: SpecialistArguments, formatted_context, citations) -> SpecialistResult:
|
||||
self.log_tuning("RAG Specialist initial_state execution started", {})
|
||||
|
||||
welcome_message = self.specialist.configuration.get('welcome_message', 'Welcome! You can start asking questions')
|
||||
welcome_message = TranslationServices.translate(self.tenant_id, welcome_message, arguments.language)
|
||||
|
||||
self.flow.state.answer = welcome_message
|
||||
self.flow.state.phase = "rag"
|
||||
|
||||
results = RAGSpecialistResult.create_for_type(self.type, self.type_version)
|
||||
|
||||
return results
|
||||
|
||||
def execute_rag_state(self, arguments: SpecialistArguments, formatted_context, citations) -> SpecialistResult:
|
||||
self.log_tuning("RAG Specialist rag_state execution started", {})
|
||||
|
||||
insufficient_info_message = TranslationServices.translate(self.tenant_id,
|
||||
INSUFFICIENT_INFORMATION_MESSAGE,
|
||||
arguments.language)
|
||||
|
||||
if formatted_context:
|
||||
flow_inputs = {
|
||||
"language": arguments.language,
|
||||
"question": arguments.question,
|
||||
"context": formatted_context,
|
||||
"history": self.formatted_history,
|
||||
"name": self.specialist.configuration.get('name', ''),
|
||||
"welcome_message": self.specialist.configuration.get('welcome_message', '')
|
||||
}
|
||||
|
||||
flow_results = self.flow.kickoff(inputs=flow_inputs)
|
||||
|
||||
if flow_results.rag_output.insufficient_info:
|
||||
flow_results.rag_output.answer = insufficient_info_message
|
||||
|
||||
rag_output = flow_results.rag_output
|
||||
|
||||
else:
|
||||
rag_output = RAGOutput(answer=insufficient_info_message, insufficient_info=True)
|
||||
|
||||
self.flow.state.rag_output = rag_output
|
||||
self.flow.state.answer = rag_output.answer
|
||||
self.flow.state.phase = "rag"
|
||||
|
||||
results = RAGSpecialistResult.create_for_type(self.type, self.type_version)
|
||||
|
||||
|
||||
class RAGSpecialistInput(BaseModel):
|
||||
language: Optional[str] = Field(None, alias="language")
|
||||
question: Optional[str] = Field(None, alias="question")
|
||||
context: Optional[str] = Field(None, alias="context")
|
||||
history: Optional[str] = Field(None, alias="history")
|
||||
name: Optional[str] = Field(None, alias="name")
|
||||
welcome_message: Optional[str] = Field(None, alias="welcome_message")
|
||||
|
||||
|
||||
class RAGSpecialistResult(SpecialistResult):
|
||||
rag_output: Optional[RAGOutput] = Field(None, alias="Rag Output")
|
||||
|
||||
|
||||
class RAGFlowState(EveAIFlowState):
|
||||
"""Flow state for RAG specialist that automatically updates from task outputs"""
|
||||
input: Optional[RAGSpecialistInput] = None
|
||||
rag_output: Optional[RAGOutput] = None
|
||||
|
||||
|
||||
class RAGFlow(EveAICrewAIFlow[RAGFlowState]):
|
||||
def __init__(self,
|
||||
specialist_executor: CrewAIBaseSpecialistExecutor,
|
||||
rag_crew: EveAICrewAICrew,
|
||||
**kwargs):
|
||||
super().__init__(specialist_executor, "RAG Specialist Flow", **kwargs)
|
||||
self.specialist_executor = specialist_executor
|
||||
self.rag_crew = rag_crew
|
||||
self.exception_raised = False
|
||||
|
||||
@start()
|
||||
def process_inputs(self):
|
||||
return ""
|
||||
|
||||
@listen(process_inputs)
|
||||
async def execute_rag(self):
|
||||
inputs = self.state.input.model_dump()
|
||||
try:
|
||||
crew_output = await self.rag_crew.kickoff_async(inputs=inputs)
|
||||
self.specialist_executor.log_tuning("RAG Crew Output", crew_output.model_dump())
|
||||
output_pydantic = crew_output.pydantic
|
||||
if not output_pydantic:
|
||||
raw_json = json.loads(crew_output.raw)
|
||||
output_pydantic = RAGOutput.model_validate(raw_json)
|
||||
self.state.rag_output = output_pydantic
|
||||
return crew_output
|
||||
except Exception as e:
|
||||
current_app.logger.error(f"CREW rag_crew Kickoff Error: {str(e)}")
|
||||
self.exception_raised = True
|
||||
raise e
|
||||
|
||||
async def kickoff_async(self, inputs=None):
|
||||
current_app.logger.debug(f"Async kickoff {self.name}")
|
||||
current_app.logger.debug(f"Inputs: {inputs}")
|
||||
self.state.input = RAGSpecialistInput.model_validate(inputs)
|
||||
result = await super().kickoff_async(inputs)
|
||||
return self.state
|
||||
Reference in New Issue
Block a user