wu981526092's picture
add
7bc750c
"""
Unified Knowledge Extraction Method (1-Task Approach)
Copied from core/agent_monitoring_unified.py and adapted for evaluation framework.
Uses the unified 1-task CrewAI approach with a single agent that performs all
knowledge extraction tasks in one step.
"""
# Import the LiteLLM fix FIRST, before any other imports that might use LiteLLM
import os
import sys
# Add the parent directory to the path to ensure imports work correctly
sys.path.append(os.path.dirname(os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))))
import json
import logging
import time
from datetime import datetime
from typing import Any, Dict
from crewai import Agent, Crew, Process, Task
from evaluation.knowledge_extraction.baselines.base_method import BaseKnowledgeExtractionMethod
from evaluation.knowledge_extraction.utils.models import KnowledgeGraph
# Import shared prompt templates
from evaluation.knowledge_extraction.utils.prompts import (
ENTITY_EXTRACTION_INSTRUCTION_PROMPT,
ENTITY_EXTRACTION_SYSTEM_PROMPT,
GRAPH_BUILDER_SYSTEM_PROMPT,
RELATION_EXTRACTION_INSTRUCTION_PROMPT,
RELATION_EXTRACTION_SYSTEM_PROMPT,
)
from utils.fix_litellm_stop_param import * # This applies the patches # noqa: F403
# Configure logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
# Set higher log levels for noisy libraries
logging.getLogger("openai").setLevel(logging.WARNING)
logging.getLogger("httpx").setLevel(logging.WARNING)
logging.getLogger("litellm").setLevel(logging.WARNING)
logging.getLogger("chromadb").setLevel(logging.WARNING)
# Set default verbosity level
verbose_level = 0
# Set environment variables
os.environ["OPENAI_MODEL_NAME"] = "gpt-5-mini"
class UnifiedKnowledgeExtractionMethod(BaseKnowledgeExtractionMethod):
"""Unified 1-task knowledge extraction method using CrewAI."""
def __init__(self, **kwargs):
super().__init__("unified_method", **kwargs)
self._setup_agent_and_task()
def _setup_agent_and_task(self):
"""Set up the CrewAI agent and task."""
# Create unified agent
self.unified_knowledge_graph_agent = Agent(
role="Unified Knowledge Graph Analyst",
goal="Create comprehensive knowledge graphs from agent system data in a single analysis pass",
backstory=f"""{ENTITY_EXTRACTION_SYSTEM_PROMPT}
{RELATION_EXTRACTION_SYSTEM_PROMPT}
{GRAPH_BUILDER_SYSTEM_PROMPT}.""",
verbose=bool(verbose_level),
llm=os.environ["OPENAI_MODEL_NAME"]
)
# Create unified task
self.unified_knowledge_graph_task = Task(
description=f"""
Extract entities:
{ENTITY_EXTRACTION_INSTRUCTION_PROMPT}
Also extract relationships:
{RELATION_EXTRACTION_INSTRUCTION_PROMPT}
Finally, build the knowledge graph:
""",
agent=self.unified_knowledge_graph_agent,
expected_output="A complete knowledge graph with entities, relations, and metadata",
output_pydantic=KnowledgeGraph,
)
# Create crew
self.unified_agent_monitoring_crew = Crew(
agents=[self.unified_knowledge_graph_agent],
tasks=[self.unified_knowledge_graph_task],
verbose=bool(verbose_level),
memory=False,
planning=False,
process=Process.sequential,
)
def process_text(self, text: str) -> Dict[str, Any]:
"""
Process input text using the unified 1-task CrewAI approach.
Args:
text: Input text to process
Returns:
Dictionary with kg_data, metadata, success, and optional error
"""
start_time = time.time()
try:
logger.info(f"process_text called with text length: {len(text)}")
logger.info(f"text first 200 chars: {repr(text[:200])}")
logger.info("Starting crew execution with input_data...")
# Run the crew with proper input mechanism
result = self.unified_agent_monitoring_crew.kickoff(inputs={"input_data": text})
logger.info(f"Crew execution completed, result type: {type(result)}")
processing_time = time.time() - start_time
# Extract the knowledge graph from the result
if hasattr(result, 'pydantic') and result.pydantic:
kg_data = result.pydantic.dict()
elif hasattr(result, 'raw'):
# Try to parse as JSON
try:
kg_data = json.loads(result.raw)
except: # noqa: E722
kg_data = {"entities": [], "relations": [], "error": "Failed to parse result"}
else:
kg_data = {"entities": [], "relations": [], "error": "Unknown result format"}
# Validate kg_data structure
if not isinstance(kg_data, dict):
raise ValueError("kg_data is not a dict after parsing")
if not ("entities" in kg_data and "relations" in kg_data):
raise ValueError("kg_data missing 'entities' or 'relations'")
# Add metadata
if "metadata" not in kg_data:
kg_data["metadata"] = {}
kg_data["metadata"]["processing_info"] = {
"method": "unified_single_task",
"processing_time_seconds": processing_time,
"processed_at": datetime.now().isoformat(),
"agent_count": 1,
"task_count": 1,
"api_calls": 1
}
# Calculate statistics
entity_count = len(kg_data.get("entities", []))
relation_count = len(kg_data.get("relations", []))
return {
"success": True,
"kg_data": kg_data,
"metadata": {
"approach": "unified_1_task",
"tasks_executed": 1,
"agents_used": 1,
"method": self.method_name,
"processing_time_seconds": processing_time,
"entity_count": entity_count,
"relation_count": relation_count,
"entities_per_second": entity_count / processing_time if processing_time > 0 else 0,
"relations_per_second": relation_count / processing_time if processing_time > 0 else 0,
"api_calls": 1
}
}
except Exception as e:
processing_time = time.time() - start_time
logger.error(f"Error in unified knowledge extraction method: {e}")
logger.error(f"Error type: {type(e).__name__}")
import traceback
logger.error(f"Traceback: {traceback.format_exc()}")
return {
"success": False,
"error": str(e),
"kg_data": {"entities": [], "relations": []},
"metadata": {
"approach": "unified_1_task",
"tasks_executed": 0,
"agents_used": 0,
"method": self.method_name,
"processing_time_seconds": processing_time,
"api_calls": 1
}
}
def extract_knowledge_graph(self, trace_data: str) -> Dict[str, Any]:
"""
Extract knowledge graph from trace data.
Args:
trace_data: Agent trace data as JSON string
Returns:
Dictionary with entities and relations
"""
try:
# Debug logging
logger.info(f"extract_knowledge_graph called with trace_data type: {type(trace_data)}")
if isinstance(trace_data, str):
logger.info(f"trace_data length: {len(trace_data)}")
logger.info(f"trace_data first 200 chars: {repr(trace_data[:200])}")
# Pass the JSON string directly to process_text without re-encoding
result = self.process_text(trace_data)
# Return just the knowledge graph data
if result.get("success", False):
return result.get("kg_data", {"entities": [], "relations": []})
else:
# Return empty knowledge graph on failure
return {"entities": [], "relations": []}
except Exception as e:
logger.error(f"Error in extract_knowledge_graph: {e}")
logger.error(f"trace_data type: {type(trace_data)}")
if isinstance(trace_data, str):
logger.error(f"trace_data content (first 200 chars): {repr(trace_data[:200])}")
return {"entities": [], "relations": []}