Chris
Final 4
225a75e
raw
history blame
11.9 kB
#!/usr/bin/env python3
"""
GAIA Agent LangGraph Workflow
Main orchestration workflow for the GAIA benchmark agent system
"""
import logging
from typing import Dict, Any, List, Literal
from langgraph.graph import StateGraph, END
from langgraph.checkpoint.memory import MemorySaver
from agents.state import GAIAAgentState, AgentRole, QuestionType
from agents.router import RouterAgent
from agents.web_researcher import WebResearchAgent
from agents.file_processor_agent import FileProcessorAgent
from agents.reasoning_agent import ReasoningAgent
from agents.synthesizer import SynthesizerAgent
from models.qwen_client import QwenClient
logger = logging.getLogger(__name__)
class GAIAWorkflow:
"""
Main GAIA agent workflow using LangGraph
Orchestrates router β†’ specialized agents β†’ synthesizer pipeline
"""
def __init__(self, llm_client: QwenClient):
self.llm_client = llm_client
# Initialize all agents
self.router = RouterAgent(llm_client)
self.web_researcher = WebResearchAgent(llm_client)
self.file_processor = FileProcessorAgent(llm_client)
self.reasoning_agent = ReasoningAgent(llm_client)
self.synthesizer = SynthesizerAgent(llm_client)
# Create workflow graph
self.workflow = self._create_workflow()
# Compile workflow with memory
self.app = self.workflow.compile(checkpointer=MemorySaver())
def _create_workflow(self) -> StateGraph:
"""Create the LangGraph workflow"""
# Define the workflow graph
workflow = StateGraph(GAIAAgentState)
# Add nodes
workflow.add_node("router", self._router_node)
workflow.add_node("web_researcher", self._web_researcher_node)
workflow.add_node("file_processor", self._file_processor_node)
workflow.add_node("reasoning_agent", self._reasoning_agent_node)
workflow.add_node("synthesizer", self._synthesizer_node)
# Define entry point
workflow.set_entry_point("router")
# Add conditional edges from router to agents
workflow.add_conditional_edges(
"router",
self._route_to_agents,
{
"web_researcher": "web_researcher",
"file_processor": "file_processor",
"reasoning_agent": "reasoning_agent",
"multi_agent": "web_researcher", # Start with web researcher for multi-agent
"synthesizer": "synthesizer" # Direct to synthesizer if no agents needed
}
)
# Add edges from agents to synthesizer
workflow.add_edge("web_researcher", "synthesizer")
workflow.add_edge("file_processor", "synthesizer")
workflow.add_edge("reasoning_agent", "synthesizer")
# Add conditional edges for multi-agent scenarios
workflow.add_conditional_edges(
"synthesizer",
self._check_if_complete,
{
"complete": END,
"need_more_agents": "file_processor" # Route to next agent if needed
}
)
return workflow
def _router_node(self, state: GAIAAgentState) -> GAIAAgentState:
"""Router node - classifies question and selects agents"""
logger.info("🧭 Executing router node")
return self.router.route_question(state)
def _web_researcher_node(self, state: GAIAAgentState) -> GAIAAgentState:
"""Web researcher node"""
logger.info("🌐 Executing web researcher node")
return self.web_researcher.process(state)
def _file_processor_node(self, state: GAIAAgentState) -> GAIAAgentState:
"""File processor node"""
logger.info("πŸ“ Executing file processor node")
return self.file_processor.process(state)
def _reasoning_agent_node(self, state: GAIAAgentState) -> GAIAAgentState:
"""Reasoning agent node"""
logger.info("🧠 Executing reasoning agent node")
return self.reasoning_agent.process(state)
def _synthesizer_node(self, state: GAIAAgentState) -> GAIAAgentState:
"""Synthesizer node - combines agent results"""
logger.info("πŸ”— Executing synthesizer node")
return self.synthesizer.process(state)
def _route_to_agents(self, state: GAIAAgentState) -> str:
"""Determine which agent(s) to route to based on router decision"""
selected_agents = state.selected_agents
# Remove synthesizer from routing decision (it's always last)
agent_roles = [agent for agent in selected_agents if agent != AgentRole.SYNTHESIZER]
if not agent_roles:
# No specific agents selected, go directly to synthesizer
return "synthesizer"
elif len(agent_roles) == 1:
# Single agent selected
agent = agent_roles[0]
if agent == AgentRole.WEB_RESEARCHER:
return "web_researcher"
elif agent == AgentRole.FILE_PROCESSOR:
return "file_processor"
elif agent == AgentRole.REASONING_AGENT:
return "reasoning_agent"
else:
return "synthesizer"
else:
# Multiple agents - start with web researcher
# The workflow will handle additional agents in subsequent steps
return "multi_agent"
def _check_if_complete(self, state: GAIAAgentState) -> str:
"""Check if processing is complete or if more agents are needed"""
# If synthesis is complete, we're done
if state.is_complete:
return "complete"
# Check if we need to run additional agents
selected_agents = state.selected_agents
executed_agents = set(state.agent_results.keys())
# Find agents that haven't been executed yet
remaining_agents = [
agent for agent in selected_agents
if agent not in executed_agents and agent != AgentRole.SYNTHESIZER
]
if remaining_agents:
# Route to next agent
next_agent = remaining_agents[0]
if next_agent == AgentRole.FILE_PROCESSOR:
return "need_more_agents" # This will route to file_processor
elif next_agent == AgentRole.REASONING_AGENT:
return "need_more_agents" # Would need additional routing logic
else:
return "complete"
else:
return "complete"
def process_question(self, question: str, file_path: str = None, file_name: str = None,
task_id: str = None, difficulty_level: int = 1) -> GAIAAgentState:
"""
Process a GAIA question through the complete workflow
Args:
question: The question to process
file_path: Optional path to associated file
file_name: Optional name of associated file
task_id: Optional task identifier
difficulty_level: Question difficulty (1-3)
Returns:
GAIAAgentState with final results
"""
logger.info(f"πŸš€ Processing question: {question[:100]}...")
# Initialize state
initial_state = GAIAAgentState()
initial_state.task_id = task_id or f"workflow_{hash(question) % 10000}"
initial_state.question = question
initial_state.file_path = file_path
initial_state.file_name = file_name
initial_state.difficulty_level = difficulty_level
try:
# Execute workflow
final_state = self.app.invoke(
initial_state,
config={"configurable": {"thread_id": initial_state.task_id}}
)
logger.info(f"βœ… Workflow complete: {final_state.final_answer[:100]}...")
return final_state
except Exception as e:
error_msg = f"Workflow execution failed: {str(e)}"
logger.error(error_msg)
# Create error state
initial_state.add_error(error_msg)
initial_state.final_answer = "Workflow execution failed"
initial_state.final_confidence = 0.0
initial_state.final_reasoning = error_msg
initial_state.is_complete = True
initial_state.requires_human_review = True
return initial_state
def get_workflow_visualization(self) -> str:
"""Get a text representation of the workflow"""
return """
GAIA Agent Workflow:
β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚ Router β”‚ ← Entry Point
β””β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”˜
β”‚
β”œβ”€ Web Researcher ──┐
β”œβ”€ File Processor ───
β”œβ”€ Reasoning Agent ──
β”‚ β”‚
β–Ό β–Ό
β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β” β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚ Synthesizer β”‚ ←─── Agent Results β”‚
β””β”€β”€β”€β”€β”€β”€β”¬β”€β”€β”€β”€β”€β”€β”˜ β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜
β”‚
β–Ό
β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚ END β”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜
Flow:
1. Router classifies question and selects appropriate agent(s)
2. Selected agents process question in parallel/sequence
3. Synthesizer combines results into final answer
4. Workflow completes with final state
"""
# Simplified workflow for cases where we don't need full LangGraph
class SimpleGAIAWorkflow:
"""
Simplified workflow that doesn't require LangGraph for basic cases
Useful for testing and lightweight deployments
"""
def __init__(self, llm_client: QwenClient):
self.llm_client = llm_client
self.router = RouterAgent(llm_client)
self.web_researcher = WebResearchAgent(llm_client)
self.file_processor = FileProcessorAgent(llm_client)
self.reasoning_agent = ReasoningAgent(llm_client)
self.synthesizer = SynthesizerAgent(llm_client)
def process_question(self, question: str, file_path: str = None, file_name: str = None,
task_id: str = None, difficulty_level: int = 1) -> GAIAAgentState:
"""Process question with simplified sequential workflow"""
# Initialize state
state = GAIAAgentState()
state.task_id = task_id or f"simple_{hash(question) % 10000}"
state.question = question
state.file_path = file_path
state.file_name = file_name
state.difficulty_level = difficulty_level
try:
# Step 1: Route
state = self.router.route_question(state)
# Step 2: Execute agents
for agent_role in state.selected_agents:
if agent_role == AgentRole.WEB_RESEARCHER:
state = self.web_researcher.process(state)
elif agent_role == AgentRole.FILE_PROCESSOR:
state = self.file_processor.process(state)
elif agent_role == AgentRole.REASONING_AGENT:
state = self.reasoning_agent.process(state)
# Skip synthesizer for now
# Step 3: Synthesize
state = self.synthesizer.process(state)
return state
except Exception as e:
error_msg = f"Simple workflow failed: {str(e)}"
state.add_error(error_msg)
state.final_answer = "Processing failed"
state.final_confidence = 0.0
state.final_reasoning = error_msg
state.is_complete = True
return state