Spaces:
Running
Running
File size: 19,162 Bytes
c2ea5ed 7bc750c c2ea5ed 7bc750c | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 | #!/usr/bin/env python3
"""
RAG-based Prompt Reconstructor for Agent Monitoring
This module uses Retrieval-Augmented Generation (RAG) to reconstruct prompts from knowledge graphs.
It leverages CrewAI's RAG capabilities to intelligently search through trace content and reconstruct
the actual prompts that would have been sent to LLMs during system execution.
"""
import json
import logging
import tempfile
import os
from typing import Dict, List, Any, Optional, Tuple
from datetime import datetime
import copy
from crewai import Agent, Task, Crew, LLM
from crewai_tools import RagTool
from concurrent.futures import ThreadPoolExecutor, as_completed
logger = logging.getLogger(__name__)
class RagPromptReconstructor:
"""
RAG-based prompt reconstructor that uses CrewAI to intelligently reconstruct
prompts by searching through vectorized trace content.
"""
def __init__(self, knowledge_graph: Dict[str, Any], original_trace: str, llm_config: Optional[Dict] = None):
"""Initialize the RAG-based prompt reconstructor."""
if not knowledge_graph or 'entities' not in knowledge_graph or 'relations' not in knowledge_graph:
raise ValueError("Invalid knowledge graph data - must contain 'entities' and 'relations'")
if not original_trace or not original_trace.strip():
raise ValueError("Original trace content is required for RAG reconstruction")
self.kg = knowledge_graph
self.original_trace = original_trace
# Create lookup dictionaries
self.entities = {entity["id"]: entity for entity in self.kg["entities"]}
self.relations = {}
self.relations_by_source = {}
self.relations_by_target = {}
# Organize relations for lookup
for relation in self.kg["relations"]:
self.relations[relation["id"]] = relation
if relation["source"] not in self.relations_by_source:
self.relations_by_source[relation["source"]] = []
self.relations_by_source[relation["source"]].append(relation)
if relation["target"] not in self.relations_by_target:
self.relations_by_target[relation["target"]] = []
self.relations_by_target[relation["target"]].append(relation)
# Initialize components
self.llm = self._init_llm(llm_config)
self.rag_tool = self._init_rag_tool()
self.query_agent = self._create_query_agent()
self.reconstruction_agent = self._create_reconstruction_agent()
logger.info(f"Initialized RagPromptReconstructor with {len(self.entities)} entities and {len(self.relations)} relations")
def _init_llm(self, llm_config: Optional[Dict]) -> LLM:
"""Initialize LLM for CrewAI agents."""
if llm_config:
return LLM(**llm_config)
return LLM(
model="gpt-5-mini",
temperature=0.1,
)
def _init_rag_tool(self) -> RagTool:
"""Initialize RAG tool with trace content."""
try:
# Create RAG tool first
rag_tool = RagTool()
# Add content directly as text instead of file
# This avoids file type detection issues
rag_tool.add(source=self.original_trace, data_type="text")
logger.info("Successfully initialized RAG tool with trace content")
return rag_tool
except Exception as e:
logger.error(f"Failed to initialize RAG tool: {e}")
# If text doesn't work, try alternative approaches
try:
# Alternative: try as raw text content
rag_tool = RagTool()
# Create a temporary file and try text_file type
with tempfile.NamedTemporaryFile(mode='w', suffix='.txt', delete=False) as tmp_file:
tmp_file.write(self.original_trace)
tmp_file_path = tmp_file.name
rag_tool.add(source=tmp_file_path, data_type="text_file")
os.unlink(tmp_file_path)
logger.info("Successfully initialized RAG tool with alternative method")
return rag_tool
except Exception as e2:
logger.error(f"Alternative RAG tool initialization also failed: {e2}")
raise RuntimeError(f"RAG tool initialization failed: {e}. Alternative also failed: {e2}")
def _create_query_agent(self) -> Agent:
"""Create agent specialized in generating semantic search queries."""
return Agent(
role="Query Generation Specialist",
goal="Generate precise search queries to find relevant trace content for prompt reconstruction",
backstory="""You are an expert at understanding conversation flows and generating
semantic search queries. Your job is to analyze relationships between entities
and create targeted queries that will retrieve the exact trace content needed
to reconstruct original prompts.""",
tools=[self.rag_tool],
llm=self.llm,
verbose=False,
)
def _create_reconstruction_agent(self) -> Agent:
"""Create agent specialized in reconstructing prompts from retrieved content."""
return Agent(
role="Prompt Reconstruction Expert",
goal="Reconstruct natural, accurate prompts from retrieved trace content",
backstory="""You are an expert at understanding how AI systems communicate
and reconstructing the exact prompts that would be sent between components.
You can identify user inputs, agent responses, system prompts, tool calls,
and conversation context.""",
tools=[self.rag_tool],
llm=self.llm,
verbose=False,
)
def _generate_search_queries(self, relation: Dict[str, Any], source_entity: Dict[str, Any],
target_entity: Dict[str, Any]) -> List[str]:
"""Generate semantic search queries for a specific relationship."""
context = {
"relation_type": relation["type"],
"source_name": source_entity["name"],
"source_type": source_entity["type"],
"target_name": target_entity["name"],
"target_type": target_entity["type"],
"interaction_prompt": relation.get("interaction_prompt", ""),
}
# Define task for query generation
query_task = Task(
description=f"""
Generate 3-5 semantic search queries to find trace content.
Relationship: {context['relation_type']}
Source: {context['source_name']} ({context['source_type']})
Target: {context['target_name']} ({context['target_type']})
Return as JSON list of strings.
""",
agent=self.query_agent,
expected_output="JSON list of search query strings"
)
try:
crew = Crew(agents=[self.query_agent], tasks=[query_task], verbose=False)
result = crew.kickoff()
if isinstance(result, str):
queries = json.loads(result)
else:
queries = result
if isinstance(queries, list):
return [str(q) for q in queries]
else:
return self._fallback_queries(context)
except Exception as e:
logger.warning(f"Query generation failed: {e}")
return self._fallback_queries(context)
def _fallback_queries(self, context: Dict[str, Any]) -> List[str]:
"""Generate fallback queries."""
return [
f"{context['source_name']} {context['target_name']}",
f"{context['relation_type'].lower()} {context['target_name']}",
f"interaction between {context['source_name']} and {context['target_name']}"
]
def _retrieve_and_reconstruct(self, relation: Dict[str, Any], source_entity: Dict[str, Any],
target_entity: Dict[str, Any], queries: List[str]) -> str:
"""Retrieve relevant content and reconstruct the prompt."""
context = {
"relation_type": relation["type"],
"source": source_entity,
"target": target_entity,
"interaction_prompt": relation.get("interaction_prompt", ""),
"queries": queries
}
# Define reconstruction task
reconstruction_task = Task(
description=f"""
Use the RAG tool to search for trace content and reconstruct the original prompt.
Relationship: {context['relation_type']}
Source: {context['source']['name']} ({context['source']['type']})
Target: {context['target']['name']} ({context['target']['type']})
Search Queries: {', '.join(queries)}
Use the RAG tool to search and reconstruct the exact prompt.
Format as natural conversation. Remove line numbers or artifacts.
Return ONLY the reconstructed prompt content.
""",
agent=self.reconstruction_agent,
expected_output="The reconstructed prompt as it would appear in the actual system"
)
try:
crew = Crew(agents=[self.reconstruction_agent], tasks=[reconstruction_task], verbose=False)
result = crew.kickoff()
if isinstance(result, str):
return result.strip()
else:
return str(result).strip()
except Exception as e:
logger.error(f"Prompt reconstruction failed: {e}")
return self._fallback_reconstruction(context)
def _fallback_reconstruction(self, context: Dict[str, Any]) -> str:
"""Generate fallback reconstruction when agent-based reconstruction fails."""
source = context["source"]
target = context["target"]
relation_type = context["relation_type"]
interaction = context.get("interaction_prompt", "")
if relation_type == "PERFORMS" and source["type"] == "Input" and target["type"] == "Agent":
user_content = source.get("raw_prompt", interaction)
return f"User: {user_content}"
# Generic fallback
source_content = source.get("raw_prompt", "")
result = f"{source['name']}: {source_content}"
if interaction:
result += f"\nInteraction: {interaction}"
return result.strip()
def reconstruct_relation_prompt(self, relation_id: str) -> Dict[str, Any]:
"""Reconstruct the actual prompt for a specific relation using RAG."""
if relation_id not in self.relations:
return {"error": f"Relation {relation_id} not found in knowledge graph"}
relation = self.relations[relation_id]
source_id = relation["source"]
target_id = relation["target"]
dependencies = {
"entities": {source_id, target_id},
"relations": {relation_id}
}
if source_id not in self.entities or target_id not in self.entities:
return {"error": f"Source or target entity for relation {relation_id} not found"}
source_entity = self.entities[source_id]
target_entity = self.entities[target_id]
# Generate queries and reconstruct
queries = self._generate_search_queries(relation, source_entity, target_entity)
reconstructed_prompt = self._retrieve_and_reconstruct(relation, source_entity, target_entity, queries)
return {
"relation_id": relation_id,
"relation_type": relation["type"],
"source": {"id": source_id, "name": source_entity["name"], "type": source_entity["type"]},
"target": {"id": target_id, "name": target_entity["name"], "type": target_entity["type"]},
"reconstructed_prompt": reconstructed_prompt,
"reconstruction_method": "rag_based",
"search_queries_used": queries,
"dependencies": {"entities": list(dependencies["entities"]), "relations": list(dependencies["relations"])}
}
def reconstruct_relations(self, parallel: bool = True, max_workers: int = 4) -> List[Dict[str, Any]]:
"""
Reconstruct all relations using RAG-based approach with optional parallel processing.
Args:
parallel: Whether to process relations in parallel (default: True)
max_workers: Maximum number of parallel workers (default: 4)
Returns:
List of dictionaries containing reconstructed prompts for each relation
"""
# Filter valid relations first
valid_relations = []
for relation_id, relation in self.relations.items():
if relation.get("type") in ["REQUIRES_TOOL", "NEXT"]:
continue
source_id = relation["source"]
target_id = relation["target"]
if source_id in self.entities and target_id in self.entities:
valid_relations.append((relation_id, relation))
if not valid_relations:
return []
if not parallel or len(valid_relations) <= 1:
# Sequential processing
reconstructed_relations = []
for relation_id, relation in valid_relations:
reconstructed = self.reconstruct_relation_prompt(relation_id)
if "error" not in reconstructed:
relation_entry = copy.deepcopy(relation)
relation_entry["prompt"] = reconstructed.get("reconstructed_prompt", "")
relation_entry["reconstruction_method"] = "rag_based"
relation_entry["dependencies"] = reconstructed.get("dependencies", {"entities": [], "relations": []})
reconstructed_relations.append(relation_entry)
return reconstructed_relations
# Parallel processing
reconstructed_relations = []
def process_relation(relation_tuple):
relation_id, relation = relation_tuple
reconstructed = self.reconstruct_relation_prompt(relation_id)
if "error" not in reconstructed:
relation_entry = copy.deepcopy(relation)
relation_entry["prompt"] = reconstructed.get("reconstructed_prompt", "")
relation_entry["reconstruction_method"] = "rag_based_parallel"
relation_entry["dependencies"] = reconstructed.get("dependencies", {"entities": [], "relations": []})
return relation_entry
else:
logger.warning(f"Failed to reconstruct relation {relation_id}: {reconstructed.get('error', 'Unknown error')}")
return None
with ThreadPoolExecutor(max_workers=min(max_workers, len(valid_relations))) as executor:
# Submit all reconstruction tasks
future_to_relation = {
executor.submit(process_relation, relation_tuple): relation_tuple[0]
for relation_tuple in valid_relations
}
# Collect results as they complete
for future in as_completed(future_to_relation):
relation_id = future_to_relation[future]
try:
result = future.result()
if result is not None:
reconstructed_relations.append(result)
logger.info(f"Completed parallel reconstruction for relation {relation_id}")
except Exception as e:
logger.error(f"Failed to reconstruct relation {relation_id} in parallel: {e}")
return reconstructed_relations
# Pure functions for external API compatibility
def reconstruct_prompts_from_knowledge_graph_rag(knowledge_graph: Dict[str, Any],
original_trace: str,
llm_config: Optional[Dict] = None,
parallel: bool = True,
max_workers: int = 4) -> List[Dict[str, Any]]:
"""
Pure function to reconstruct prompts from knowledge graph using RAG approach.
Args:
knowledge_graph: Knowledge graph data with entities and relations
original_trace: Original trace content for RAG vectorization
llm_config: Optional LLM configuration for CrewAI agents
parallel: Whether to process relations in parallel (default: True)
max_workers: Maximum number of parallel workers (default: 4)
Returns:
List of dictionaries containing reconstructed prompts for each relation
"""
reconstructor = RagPromptReconstructor(knowledge_graph, original_trace, llm_config)
return reconstructor.reconstruct_relations(parallel=parallel, max_workers=max_workers)
def enrich_knowledge_graph_with_prompts_rag(knowledge_graph: Dict[str, Any],
original_trace: str,
llm_config: Optional[Dict] = None,
parallel: bool = True,
max_workers: int = 4) -> Dict[str, Any]:
"""
Pure function to enrich a knowledge graph with RAG-reconstructed prompts.
Args:
knowledge_graph: Knowledge graph data with entities and relations
original_trace: Original trace content for RAG vectorization
llm_config: Optional LLM configuration for CrewAI agents
parallel: Whether to process relations in parallel (default: True)
max_workers: Maximum number of parallel workers (default: 4)
Returns:
Enhanced knowledge graph with prompt_reconstructions field containing
the RAG-reconstructed prompts for each relation
"""
reconstructor = RagPromptReconstructor(knowledge_graph, original_trace, llm_config)
reconstructed_relations = reconstructor.reconstruct_relations(parallel=parallel, max_workers=max_workers)
# Create enhanced knowledge graph
enhanced_kg = copy.deepcopy(knowledge_graph)
enhanced_kg["prompt_reconstructions"] = reconstructed_relations
enhanced_kg["reconstruction_metadata"] = {
"method": "rag_based",
"reconstructed_at": datetime.now().isoformat(),
"total_relations_processed": len(reconstructed_relations),
"original_trace_length": len(original_trace)
}
return enhanced_kg
|