File size: 16,200 Bytes
ac6a4ef |
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 |
"""Supervisor node implementation for the agent supervisor system."""
from typing import Dict, List, Literal, Optional, Union, Type, cast
from langchain_core.messages import HumanMessage, SystemMessage, AIMessage
from langchain_core.prompts import ChatPromptTemplate
from langgraph.graph import StateGraph, START, END
from langgraph.types import Command
from react_agent.configuration import Configuration
from react_agent.state import WORKERS, MEMBERS, ROUTING, VERDICTS, State, Router
from react_agent.utils import load_chat_model, format_system_prompt, get_message_text
from react_agent import prompts
# Compile-time type definitions
SupervisorDestinations = Literal["planner", "critic", "researcher", "coder", "final_answer", "__end__"]
def supervisor_node(state: State) -> Command[SupervisorDestinations]:
"""Supervising LLM that decides which specialized agent should act next.
Args:
state: The current state with messages
Returns:
Command with routing information
"""
# Get configuration to use supervisor_model
configuration = Configuration.from_context()
# Track steps to prevent infinite loops
steps_taken = state.get("steps_taken", 0)
steps_taken += 1
state_updates = {"steps_taken": steps_taken}
# Check if we've hit our step limit
if steps_taken >= configuration.recursion_limit - 5: # Buffer of 5 steps
# Extract the best answer we have from context if possible
context = state.get("context", {})
answer = extract_best_answer_from_context(context)
return Command(
goto="final_answer",
update={
"messages": [
HumanMessage(
content=f"Maximum steps ({steps_taken}) reached. Extracting best answer from available information.",
name="supervisor"
)
],
"draft_answer": f"FINAL ANSWER: {answer}",
"retry_exhausted": True, # Flag to indicate we've exhausted retries
"steps_taken": steps_taken
}
)
# Safety check - prevent infinite loops by forcing termination after too many retry steps
retry_count = state.get("retry_count", 0)
max_retries = 2 # Maximum number of allowed retries
if retry_count > max_retries:
# Extract the best answer we have from context if possible
context = state.get("context", {})
answer = extract_best_answer_from_context(context)
return Command(
goto="final_answer",
update={
"messages": [
HumanMessage(
content=f"Maximum retries ({max_retries}) reached. Extracting best answer from available information.",
name="supervisor"
)
],
"draft_answer": f"FINAL ANSWER: {answer}",
"retry_exhausted": True, # Flag to indicate we've exhausted retries
"steps_taken": steps_taken
}
)
# Check if we need a plan
if not state.get("plan"):
return Command(
goto="planner",
update={
**state_updates
}
)
# Validate that the plan has at least one step
plan = state.get("plan")
if not plan.get("steps") or len(plan.get("steps", [])) == 0:
# Plan has no steps, go back to planner with explicit instructions
return Command(
goto="planner",
update={
"messages": [
HumanMessage(
content="Previous plan had 0 steps. Please create a plan with at least 1 step to solve the user's question.",
name="supervisor"
)
],
"plan": None,
**state_updates
}
)
# Check if we have a critic verdict that requires replanning
critic_verdict = state.get("critic_verdict")
if critic_verdict:
if critic_verdict.get("verdict") == VERDICTS[0]: # CORRECT
# Final answer is approved, navigate to the final_answer node
# This will generate a polished response before ending
return Command(
goto="final_answer",
update={
"messages": [
HumanMessage(
content="Answer approved by critic. Generating final response.",
name="supervisor"
)
]
}
)
elif critic_verdict.get("verdict") == VERDICTS[1]: # RETRY
# IMPORTANT: Get the current retry count BEFORE incrementing
current_retry_count = state.get("retry_count", 0)
# Check if we're at the maximum allowed retries
if current_retry_count >= max_retries:
# Extract best answer and go to final_answer
context = state.get("context", {})
answer = extract_best_answer_from_context(context)
return Command(
goto="final_answer",
update={
"messages": [
HumanMessage(
content=f"Maximum retries ({max_retries}) reached. Proceeding with best available answer.",
name="supervisor"
)
],
"draft_answer": f"FINAL ANSWER: {answer}",
"retry_exhausted": True # Flag to indicate we've exhausted retries
}
)
# Reset the plan but KEEP the context from previous iterations
context = state.get("context", {})
worker_results = state.get("worker_results", {})
# Get the critic's reason for rejection, if any
reason = critic_verdict.get("reason", "")
if not reason or reason.strip() == "\"":
reason = "Answer did not meet format requirements"
# Check if this is a formatting issue
format_issues = [
"format", "concise", "explanation", "not formatted",
"instead of just", "contains explanations", "FINAL ANSWER"
]
is_format_issue = any(issue in reason.lower() for issue in format_issues)
# If we have enough information but the format is wrong, go directly to final answer
has_sufficient_info = has_sufficient_information(state)
if is_format_issue and has_sufficient_info and current_retry_count >= 0:
# We have information but formatting is wrong - skip planning and go to final answer
return Command(
goto="final_answer",
update={
"messages": [
HumanMessage(
content="We have sufficient information but formatting issues. Generating properly formatted answer.",
name="supervisor"
)
],
"retry_count": current_retry_count + 1 # Still increment retry count
}
)
# Increment the retry counter
next_retry_count = current_retry_count + 1
return Command(
goto="planner",
update={
"plan": None,
"current_step_index": None,
"draft_answer": None,
"critic_verdict": None,
# Keep the context and worker_results
"context": context,
"worker_results": worker_results,
# Track retries - IMPORTANT: store the incremented count
"retry_count": next_retry_count,
# Add a message about the retry (using the INCREMENTED count)
"messages": [
HumanMessage(
content=f"Retrying with new plan (retry #{next_retry_count}). Reason: {reason}",
name="supervisor"
)
]
}
)
# Get the current step from the plan
plan = state["plan"]
current_step_index = state.get("current_step_index", 0)
# Check if we've completed all steps
if current_step_index >= len(plan["steps"]):
# Use context to compile the draft answer
context = state.get("context", {})
# Combine the most recent worker outputs as the draft answer
worker_results = []
for worker in WORKERS:
if worker in context:
worker_results.append(f"**{worker.title()}**: {context[worker]}")
# Compile the draft answer from all worker outputs
draft_content = "\n\n".join(worker_results)
# Send to the critic for evaluation
return Command(
goto="critic",
update={
"draft_answer": draft_content,
# Add a message about moving to evaluation
"messages": [
HumanMessage(
content="All steps completed. Evaluating the answer.",
name="supervisor"
)
]
}
)
# Get the current step
current_step = plan["steps"][current_step_index]
worker = current_step["worker"]
instruction = current_step["instruction"]
# Extract only the most relevant context for the current worker and task
context_info = ""
if state.get("context"):
# Filter context by relevance to the current task
relevant_context = {}
# For the coder, extract numerical data and parameters from researcher
if worker == "coder" and "researcher" in state["context"]:
relevant_context["researcher"] = state["context"]["researcher"]
# For the researcher, previous coder calculations might be relevant
if worker == "researcher" and "coder" in state["context"]:
# Only include numerical results from coder, not code snippets
coder_content = state["context"]["coder"]
if len(coder_content) < 100: # Only short results are likely just numbers
relevant_context["coder"] = coder_content
# Format the relevant context items
context_items = []
for key, value in relevant_context.items():
# Summarize if value is too long
if len(value) > 200:
# Find first sentence or up to 200 chars
summary = value[:200]
if '.' in summary:
summary = summary.split('.')[0] + '.'
context_items.append(f"Previous {key} found: {summary}...")
else:
context_items.append(f"Previous {key} found: {value}")
if context_items:
context_info = "\n\nRelevant context: " + "\n".join(context_items)
# Enhance the instruction with context
enhanced_instruction = f"{instruction}{context_info}"
# Add guidance based on worker type
if worker == "coder":
enhanced_instruction += "\nProvide both your calculation method AND the final result value."
elif worker == "researcher":
enhanced_instruction += "\nFocus on gathering factual information related to the task."
# Add the instruction to the messages
messages_update = [
HumanMessage(
content=f"Step {current_step_index + 1}: {enhanced_instruction}",
name="supervisor"
)
]
# Cast worker to appropriate type to satisfy type checking
worker_destination = cast(SupervisorDestinations, worker)
# Move to the appropriate worker
return Command(
goto=worker_destination,
update={
"messages": messages_update,
"next": worker, # For backward compatibility
**state_updates
}
)
def extract_best_answer_from_context(context):
"""Extract the best available answer from context.
This is a generic function to extract answers from any type of question context.
It progressively tries different strategies to find a suitable answer.
Args:
context: The state context containing worker outputs
Returns:
Best answer found or "unknown" if nothing suitable is found
"""
answer = "unknown"
# First check if the coder already provided a properly formatted answer
if "coder" in context:
coder_content = context["coder"]
# Look for "FINAL ANSWER: X" pattern in the coder output
import re
answer_match = re.search(r"FINAL ANSWER:\s*(.*?)(?:\n|$)", coder_content, re.IGNORECASE)
if answer_match:
return answer_match.group(1).strip()
# If no answer in coder output, check researcher content
if "researcher" in context:
researcher_content = context["researcher"]
# Look for lists in the researcher content (common pattern)
import re
# Look for bulleted list items
list_items = re.findall(r"[-•*]\s+([^:\n]+)", researcher_content)
if list_items:
# Format as comma-separated list
answer = ",".join(item.strip() for item in list_items)
return answer
# Look for emphasized/bold items which might be key information
bold_items = re.findall(r"\*\*([^*]+)\*\*", researcher_content)
if bold_items:
# Join the important items as a comma-separated list
processed_items = []
for item in bold_items:
# Remove common filler words and clean up the item
clean_item = re.sub(r'(^|\s)(a|an|the|is|are|was|were|be|been)(\s|$)', ' ', item)
clean_item = clean_item.strip()
if clean_item and len(clean_item) < 30: # Only include reasonably short items
processed_items.append(clean_item)
if processed_items:
answer = ",".join(processed_items)
return answer
# If we still don't have an answer, try to extract common entities
combined_content = ""
for worker_type, content in context.items():
combined_content += " " + content
# Look for numbers in the content
import re
numbers = re.findall(r'\b(\d+(?:\.\d+)?)\b', combined_content)
if numbers:
answer = numbers[0] # Use the first number found
return answer
def has_sufficient_information(state):
"""Determine if we have enough information to generate a final answer.
Args:
state: The current conversation state
Returns:
Boolean indicating if we have sufficient information
"""
context = state.get("context", {})
# If we have both researcher and coder outputs, we likely have enough info
if "researcher" in context and "coder" in context:
return True
# If we have a substantial researcher output, that might be enough
if "researcher" in context and len(context["researcher"]) > 150:
return True
# If we have any worker output that contains lists or formatted data
for worker, content in context.items():
if content and (
"- " in content or # Bullet point
"•" in content or # Bullet point
"*" in content or # Emphasis or bullet
":" in content # Definition or explanation
):
return True
return False |