Spaces:
Running
Running
File size: 10,550 Bytes
c2ea5ed 7bc750c c2ea5ed |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 |
"""
Direct LLM Knowledge Extraction Method
A streamlined approach that uses direct LLM API calls with structured output
instead of the CrewAI framework for better performance and cost efficiency.
"""
import json
import logging
import os
import sys
import time
from datetime import datetime
from typing import Any, Dict
from openai import OpenAI
from pydantic import ValidationError
from evaluation.knowledge_extraction.baselines.unified_method import KnowledgeGraph
# 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__))))))
from evaluation.knowledge_extraction.baselines.base_method import BaseKnowledgeExtractionMethod
# Import shared prompt templates
from evaluation.knowledge_extraction.utils.prompts import (
ENTITY_EXTRACTION_INSTRUCTION_PROMPT,
ENTITY_EXTRACTION_SYSTEM_PROMPT,
GRAPH_BUILDER_INSTRUCTION_PROMPT,
GRAPH_BUILDER_SYSTEM_PROMPT,
RELATION_EXTRACTION_INSTRUCTION_PROMPT,
RELATION_EXTRACTION_SYSTEM_PROMPT,
)
# 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)
class DirectLLMKnowledgeExtractor(BaseKnowledgeExtractionMethod):
"""Direct LLM knowledge extraction method using OpenAI API with structured output."""
def __init__(self, model: str = "gpt-5-mini", **kwargs):
super().__init__("direct_llm_method", **kwargs)
self.client = OpenAI()
self.model = model
self.max_retries = 3
self.retry_delay = 1.0
def _get_optimized_system_prompt(self) -> str:
"""Get the optimized system prompt for knowledge graph extraction."""
# Combine all system prompts for a unified extraction
return f"""{ENTITY_EXTRACTION_SYSTEM_PROMPT}
{RELATION_EXTRACTION_SYSTEM_PROMPT}
{GRAPH_BUILDER_SYSTEM_PROMPT}"""
def _get_extraction_instruction(self, text: str) -> str:
"""Get the extraction instruction with the input text."""
# Combine entity and relation extraction instructions
entity_instruction = ENTITY_EXTRACTION_INSTRUCTION_PROMPT.format(input_data=text)
relation_instruction = RELATION_EXTRACTION_INSTRUCTION_PROMPT.format(input_data=text)
graph_instruction = GRAPH_BUILDER_INSTRUCTION_PROMPT
return f"""Extract a complete knowledge graph from the following agent system data.
First, extract entities following these instructions:
{entity_instruction}
Then, extract relations following these instructions:
{relation_instruction}
Finally, build the knowledge graph following these instructions:
{graph_instruction}
"""
def _extract_with_retry(self, text: str) -> Dict[str, Any]:
"""Extract knowledge graph with retry logic using new Structured Outputs API."""
last_error = None
for attempt in range(self.max_retries):
try:
logger.info(f"Extraction attempt {attempt + 1}/{self.max_retries}")
# Use the beta API with structured outputs
response = self.client.beta.chat.completions.parse(
model=self.model,
messages=[
{"role": "system", "content": self._get_optimized_system_prompt()},
{"role": "user", "content": self._get_extraction_instruction(text)},
],
response_format=KnowledgeGraph,
temperature=0,
)
# Get the parsed response
parsed_response = response.choices[0].message.parsed
# Handle refusal
if response.choices[0].message.refusal:
raise ValueError(f"Model refused: {response.choices[0].message.refusal}")
if not parsed_response:
raise ValueError("Empty parsed response from LLM")
# Convert to dict
kg_dict = parsed_response.model_dump()
# Add metadata
kg_dict["metadata"] = {
"method": "direct_llm",
"model": self.model,
"attempt": attempt + 1,
"timestamp": datetime.now().isoformat(),
"token_usage": {
"prompt_tokens": response.usage.prompt_tokens if response.usage else 0,
"completion_tokens": response.usage.completion_tokens if response.usage else 0,
"total_tokens": response.usage.total_tokens if response.usage else 0,
},
}
logger.info(
f"Successfully extracted KG with {len(kg_dict['entities'])} entities and {len(kg_dict['relations'])} relations"
)
return kg_dict
except json.JSONDecodeError as e:
last_error = f"JSON parsing error: {e}"
logger.warning(f"Attempt {attempt + 1} failed: {last_error}")
except ValidationError as e:
last_error = f"Validation error: {e}"
logger.warning(f"Attempt {attempt + 1} failed: {last_error}")
except Exception as e:
last_error = f"API error: {e}"
logger.warning(f"Attempt {attempt + 1} failed: {last_error}")
if attempt < self.max_retries - 1:
time.sleep(self.retry_delay * (2**attempt)) # Exponential backoff
# If all attempts failed, return empty structure
logger.error(f"All extraction attempts failed. Last error: {last_error}")
return {
"entities": [],
"relations": [],
"system_name": "Failed Extraction",
"system_summary": "Knowledge graph extraction failed after multiple attempts.",
"metadata": {"error": last_error, "method": "direct_llm"},
}
def process_text(self, text: str) -> Dict[str, Any]:
"""
Process input text using direct LLM API calls.
Args:
text: Input text to process
Returns:
Dictionary with kg_data, metadata, success, and optional error
"""
start_time = time.time()
try:
logger.info(f"Processing text with Direct LLM method (length: {len(text)})")
# Extract knowledge graph
kg_data = self._extract_with_retry(text)
processing_time = time.time() - start_time
# Check if extraction was successful
success = len(kg_data.get("entities", [])) > 0 or len(kg_data.get("relations", [])) > 0
# Calculate statistics
entity_count = len(kg_data.get("entities", []))
relation_count = len(kg_data.get("relations", []))
# Add processing metadata
if "metadata" not in kg_data:
kg_data["metadata"] = {}
kg_data["metadata"].update(
{
"processing_info": {
"method": "direct_llm",
"processing_time_seconds": processing_time,
"processed_at": datetime.now().isoformat(),
"model": self.model,
"api_calls": 1,
"entity_count": entity_count,
"relation_count": relation_count,
}
}
)
return {
"success": success,
"kg_data": kg_data,
"metadata": {
"approach": "direct_llm",
"model": self.model,
"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,
"token_usage": kg_data.get("metadata", {}).get("token_usage", {}),
},
}
except Exception as e:
processing_time = time.time() - start_time
logger.error(f"Error in direct LLM knowledge extraction: {e}")
import traceback
logger.error(f"Traceback: {traceback.format_exc()}")
return {
"success": False,
"error": str(e),
"kg_data": {"entities": [], "relations": []},
"metadata": {
"approach": "direct_llm",
"model": self.model,
"method": self.method_name,
"processing_time_seconds": processing_time,
"api_calls": 1,
"error": str(e),
},
}
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:
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])}")
# Process the trace data
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": []}
|